1 ========================
2 USB Gadget API for Linux
3 ========================
5 :Author: David Brownell
11 This document presents a Linux-USB "Gadget" kernel mode API, for use
12 within peripherals and other USB devices that embed Linux. It provides
13 an overview of the API structure, and shows how that fits into a system
14 development project. This is the first such API released on Linux to
15 address a number of important problems, including:
17 - Supports USB 2.0, for high speed devices which can stream data at
18 several dozen megabytes per second.
20 - Handles devices with dozens of endpoints just as well as ones with
21 just two fixed-function ones. Gadget drivers can be written so
22 they're easy to port to new hardware.
24 - Flexible enough to expose more complex USB device capabilities such
25 as multiple configurations, multiple interfaces, composite devices,
26 and alternate interface settings.
28 - USB "On-The-Go" (OTG) support, in conjunction with updates to the
31 - Sharing data structures and API models with the Linux-USB host side
32 API. This helps the OTG support, and looks forward to more-symmetric
33 frameworks (where the same I/O model is used by both host and device
36 - Minimalist, so it's easier to support new device controller hardware.
37 I/O processing doesn't imply large demands for memory or CPU
40 Most Linux developers will not be able to use this API, since they have
41 USB ``host`` hardware in a PC, workstation, or server. Linux users with
42 embedded systems are more likely to have USB peripheral hardware. To
43 distinguish drivers running inside such hardware from the more familiar
44 Linux "USB device drivers", which are host side proxies for the real USB
45 devices, a different term is used: the drivers inside the peripherals
46 are "USB gadget drivers". In USB protocol interactions, the device
47 driver is the master (or "client driver") and the gadget driver is the
48 slave (or "function driver").
50 The gadget API resembles the host side Linux-USB API in that both use
51 queues of request objects to package I/O buffers, and those requests may
52 be submitted or canceled. They share common definitions for the standard
53 USB *Chapter 9* messages, structures, and constants. Also, both APIs
54 bind and unbind drivers to devices. The APIs differ in detail, since the
55 host side's current URB framework exposes a number of implementation
56 details and assumptions that are inappropriate for a gadget API. While
57 the model for control transfers and configuration management is
58 necessarily different (one side is a hardware-neutral master, the other
59 is a hardware-aware slave), the endpoint I/0 API used here should also
60 be usable for an overhead-reduced host side API.
62 Structure of Gadget Drivers
63 ===========================
65 A system running inside a USB peripheral normally has at least three
66 layers inside the kernel to handle USB protocol processing, and may have
67 additional layers in user space code. The ``gadget`` API is used by the
68 middle layer to interact with the lowest level (which directly handles
71 In Linux, from the bottom up, these layers are:
73 *USB Controller Driver*
74 This is the lowest software level. It is the only layer that talks
75 to hardware, through registers, fifos, dma, irqs, and the like. The
76 ``<linux/usb/gadget.h>`` API abstracts the peripheral controller
77 endpoint hardware. That hardware is exposed through endpoint
78 objects, which accept streams of IN/OUT buffers, and through
79 callbacks that interact with gadget drivers. Since normal USB
80 devices only have one upstream port, they only have one of these
81 drivers. The controller driver can support any number of different
82 gadget drivers, but only one of them can be used at a time.
84 Examples of such controller hardware include the PCI-based NetChip
85 2280 USB 2.0 high speed controller, the SA-11x0 or PXA-25x UDC
86 (found within many PDAs), and a variety of other products.
89 The lower boundary of this driver implements hardware-neutral USB
90 functions, using calls to the controller driver. Because such
91 hardware varies widely in capabilities and restrictions, and is used
92 in embedded environments where space is at a premium, the gadget
93 driver is often configured at compile time to work with endpoints
94 supported by one particular controller. Gadget drivers may be
95 portable to several different controllers, using conditional
96 compilation. (Recent kernels substantially simplify the work
97 involved in supporting new hardware, by *autoconfiguring* endpoints
98 automatically for many bulk-oriented drivers.) Gadget driver
99 responsibilities include:
101 - handling setup requests (ep0 protocol responses) possibly
102 including class-specific functionality
104 - returning configuration and string descriptors
106 - (re)setting configurations and interface altsettings, including
107 enabling and configuring endpoints
109 - handling life cycle events, such as managing bindings to
110 hardware, USB suspend/resume, remote wakeup, and disconnection
113 - managing IN and OUT transfers on all currently enabled endpoints
115 Such drivers may be modules of proprietary code, although that
116 approach is discouraged in the Linux community.
119 Most gadget drivers have an upper boundary that connects to some
120 Linux driver or framework in Linux. Through that boundary flows the
121 data which the gadget driver produces and/or consumes through
122 protocol transfers over USB. Examples include:
124 - user mode code, using generic (gadgetfs) or application specific
127 - networking subsystem (for network gadgets, like the CDC Ethernet
130 - data capture drivers, perhaps video4Linux or a scanner driver; or
131 test and measurement hardware.
133 - input subsystem (for HID gadgets)
135 - sound subsystem (for audio gadgets)
137 - file system (for PTP gadgets)
139 - block i/o subsystem (for usb-storage gadgets)
144 Other layers may exist. These could include kernel layers, such as
145 network protocol stacks, as well as user mode applications building
146 on standard POSIX system call APIs such as ``open()``, ``close()``,
147 ``read()`` and ``write()``. On newer systems, POSIX Async I/O calls may
148 be an option. Such user mode code will not necessarily be subject to
149 the GNU General Public License (GPL).
151 OTG-capable systems will also need to include a standard Linux-USB host
152 side stack, with ``usbcore``, one or more *Host Controller Drivers*
153 (HCDs), *USB Device Drivers* to support the OTG "Targeted Peripheral
154 List", and so forth. There will also be an *OTG Controller Driver*,
155 which is visible to gadget and device driver developers only indirectly.
156 That helps the host and device side USB controllers implement the two
157 new OTG protocols (HNP and SRP). Roles switch (host to peripheral, or
158 vice versa) using HNP during USB suspend processing, and SRP can be
159 viewed as a more battery-friendly kind of device wakeup protocol.
161 Over time, reusable utilities are evolving to help make some gadget
162 driver tasks simpler. For example, building configuration descriptors
163 from vectors of descriptors for the configurations interfaces and
164 endpoints is now automated, and many drivers now use autoconfiguration
165 to choose hardware endpoints and initialize their descriptors. A
166 potential example of particular interest is code implementing standard
167 USB-IF protocols for HID, networking, storage, or audio classes. Some
168 developers are interested in KDB or KGDB hooks, to let target hardware
169 be remotely debugged. Most such USB protocol code doesn't need to be
170 hardware-specific, any more than network protocols like X11, HTTP, or
171 NFS are. Such gadget-side interface drivers should eventually be
172 combined, to implement composite devices.
174 Kernel Mode Gadget API
175 ======================
177 Gadget drivers declare themselves through a struct
178 :c:type:`usb_gadget_driver`, which is responsible for most parts of enumeration
179 for a struct :c:type:`usb_gadget`. The response to a set_configuration usually
180 involves enabling one or more of the struct :c:type:`usb_ep` objects exposed by
181 the gadget, and submitting one or more struct :c:type:`usb_request` buffers to
182 transfer data. Understand those four data types, and their operations,
183 and you will understand how this API works.
187 Other than the "Chapter 9" data types, most of the significant data
188 types and functions are described here.
190 However, some relevant information is likely omitted from what you
191 are reading. One example of such information is endpoint
192 autoconfiguration. You'll have to read the header file, and use
193 example source code (such as that for "Gadget Zero"), to fully
196 The part of the API implementing some basic driver capabilities is
197 specific to the version of the Linux kernel that's in use. The 2.6
198 and upper kernel versions include a *driver model* framework that has
199 no analogue on earlier kernels; so those parts of the gadget API are
200 not fully portable. (They are implemented on 2.4 kernels, but in a
201 different way.) The driver model state is another part of this API that is
202 ignored by the kerneldoc tools.
204 The core API does not expose every possible hardware feature, only the
205 most widely available ones. There are significant hardware features,
206 such as device-to-device DMA (without temporary storage in a memory
207 buffer) that would be added using hardware-specific APIs.
209 This API allows drivers to use conditional compilation to handle
210 endpoint capabilities of different hardware, but doesn't require that.
211 Hardware tends to have arbitrary restrictions, relating to transfer
212 types, addressing, packet sizes, buffering, and availability. As a rule,
213 such differences only matter for "endpoint zero" logic that handles
214 device configuration and management. The API supports limited run-time
215 detection of capabilities, through naming conventions for endpoints.
216 Many drivers will be able to at least partially autoconfigure
217 themselves. In particular, driver init sections will often have endpoint
218 autoconfiguration logic that scans the hardware's list of endpoints to
219 find ones matching the driver requirements (relying on those
220 conventions), to eliminate some of the most common reasons for
221 conditional compilation.
223 Like the Linux-USB host side API, this API exposes the "chunky" nature
224 of USB messages: I/O requests are in terms of one or more "packets", and
225 packet boundaries are visible to drivers. Compared to RS-232 serial
226 protocols, USB resembles synchronous protocols like HDLC (N bytes per
227 frame, multipoint addressing, host as the primary station and devices as
228 secondary stations) more than asynchronous ones (tty style: 8 data bits
229 per frame, no parity, one stop bit). So for example the controller
230 drivers won't buffer two single byte writes into a single two-byte USB
231 IN packet, although gadget drivers may do so when they implement
232 protocols where packet boundaries (and "short packets") are not
238 Gadget drivers make endpoint I/O requests to hardware without needing to
239 know many details of the hardware, but driver setup/configuration code
240 needs to handle some differences. Use the API like this:
242 1. Register a driver for the particular device side usb controller
243 hardware, such as the net2280 on PCI (USB 2.0), sa11x0 or pxa25x as
244 found in Linux PDAs, and so on. At this point the device is logically
245 in the USB ch9 initial state (``attached``), drawing no power and not
246 usable (since it does not yet support enumeration). Any host should
247 not see the device, since it's not activated the data line pullup
248 used by the host to detect a device, even if VBUS power is available.
250 2. Register a gadget driver that implements some higher level device
251 function. That will then bind() to a :c:type:`usb_gadget`, which activates
252 the data line pullup sometime after detecting VBUS.
254 3. The hardware driver can now start enumerating. The steps it handles
255 are to accept USB ``power`` and ``set_address`` requests. Other steps are
256 handled by the gadget driver. If the gadget driver module is unloaded
257 before the host starts to enumerate, steps before step 7 are skipped.
259 4. The gadget driver's ``setup()`` call returns usb descriptors, based both
260 on what the bus interface hardware provides and on the functionality
261 being implemented. That can involve alternate settings or
262 configurations, unless the hardware prevents such operation. For OTG
263 devices, each configuration descriptor includes an OTG descriptor.
265 5. The gadget driver handles the last step of enumeration, when the USB
266 host issues a ``set_configuration`` call. It enables all endpoints used
267 in that configuration, with all interfaces in their default settings.
268 That involves using a list of the hardware's endpoints, enabling each
269 endpoint according to its descriptor. It may also involve using
270 ``usb_gadget_vbus_draw`` to let more power be drawn from VBUS, as
271 allowed by that configuration. For OTG devices, setting a
272 configuration may also involve reporting HNP capabilities through a
275 6. Do real work and perform data transfers, possibly involving changes
276 to interface settings or switching to new configurations, until the
277 device is disconnect()ed from the host. Queue any number of transfer
278 requests to each endpoint. It may be suspended and resumed several
279 times before being disconnected. On disconnect, the drivers go back
282 7. When the gadget driver module is being unloaded, the driver unbind()
283 callback is issued. That lets the controller driver be unloaded.
285 Drivers will normally be arranged so that just loading the gadget driver
286 module (or statically linking it into a Linux kernel) allows the
287 peripheral device to be enumerated, but some drivers will defer
288 enumeration until some higher level component (like a user mode daemon)
289 enables it. Note that at this lowest level there are no policies about
290 how ep0 configuration logic is implemented, except that it should obey
291 USB specifications. Such issues are in the domain of gadget drivers,
292 including knowing about implementation constraints imposed by some USB
293 controllers or understanding that composite devices might happen to be
294 built by integrating reusable components.
296 Note that the lifecycle above can be slightly different for OTG devices.
297 Other than providing an additional OTG descriptor in each configuration,
298 only the HNP-related differences are particularly visible to driver
299 code. They involve reporting requirements during the ``SET_CONFIGURATION``
300 request, and the option to invoke HNP during some suspend callbacks.
301 Also, SRP changes the semantics of ``usb_gadget_wakeup`` slightly.
303 USB 2.0 Chapter 9 Types and Constants
304 -------------------------------------
306 Gadget drivers rely on common USB structures and constants defined in
307 the :ref:`linux/usb/ch9.h <usb_chapter9>` header file, which is standard in
308 Linux 2.6+ kernels. These are the same types and constants used by host side
309 drivers (and usbcore).
311 Core Objects and Methods
312 ------------------------
314 These are declared in ``<linux/usb/gadget.h>``, and are used by gadget
315 drivers to interact with USB peripheral controller drivers.
317 .. kernel-doc:: include/linux/usb/gadget.h
323 The core API is sufficient for writing a USB Gadget Driver, but some
324 optional utilities are provided to simplify common tasks. These
325 utilities include endpoint autoconfiguration.
327 .. kernel-doc:: drivers/usb/gadget/usbstring.c
330 .. kernel-doc:: drivers/usb/gadget/config.c
333 Composite Device Framework
334 --------------------------
336 The core API is sufficient for writing drivers for composite USB devices
337 (with more than one function in a given configuration), and also
338 multi-configuration devices (also more than one function, but not
339 necessarily sharing a given configuration). There is however an optional
340 framework which makes it easier to reuse and combine functions.
342 Devices using this framework provide a struct :c:type:`usb_composite_driver`,
343 which in turn provides one or more struct :c:type:`usb_configuration`
344 instances. Each such configuration includes at least one struct
345 :c:type:`usb_function`, which packages a user visible role such as "network
346 link" or "mass storage device". Management functions may also exist,
347 such as "Device Firmware Upgrade".
349 .. kernel-doc:: include/linux/usb/composite.h
352 .. kernel-doc:: drivers/usb/gadget/composite.c
355 Composite Device Functions
356 --------------------------
358 At this writing, a few of the current gadget drivers have been converted
359 to this framework. Near-term plans include converting all of them,
360 except for ``gadgetfs``.
362 Peripheral Controller Drivers
363 =============================
365 The first hardware supporting this API was the NetChip 2280 controller,
366 which supports USB 2.0 high speed and is based on PCI. This is the
367 ``net2280`` driver module. The driver supports Linux kernel versions 2.4
368 and 2.6; contact NetChip Technologies for development boards and product
371 Other hardware working in the ``gadget`` framework includes: Intel's PXA
372 25x and IXP42x series processors (``pxa2xx_udc``), Toshiba TC86c001
373 "Goku-S" (``goku_udc``), Renesas SH7705/7727 (``sh_udc``), MediaQ 11xx
374 (``mq11xx_udc``), Hynix HMS30C7202 (``h7202_udc``), National 9303/4
375 (``n9604_udc``), Texas Instruments OMAP (``omap_udc``), Sharp LH7A40x
376 (``lh7a40x_udc``), and more. Most of those are full speed controllers.
378 At this writing, there are people at work on drivers in this framework
379 for several other USB device controllers, with plans to make many of
380 them be widely available.
382 A partial USB simulator, the ``dummy_hcd`` driver, is available. It can
383 act like a net2280, a pxa25x, or an sa11x0 in terms of available
384 endpoints and device speeds; and it simulates control, bulk, and to some
385 extent interrupt transfers. That lets you develop some parts of a gadget
386 driver on a normal PC, without any special hardware, and perhaps with
387 the assistance of tools such as GDB running with User Mode Linux. At
388 least one person has expressed interest in adapting that approach,
389 hooking it up to a simulator for a microcontroller. Such simulators can
390 help debug subsystems where the runtime hardware is unfriendly to
391 software development, or is not yet available.
393 Support for other controllers is expected to be developed and
394 contributed over time, as this driver framework evolves.
399 In addition to *Gadget Zero* (used primarily for testing and development
400 with drivers for usb controller hardware), other gadget drivers exist.
402 There's an ``ethernet`` gadget driver, which implements one of the most
403 useful *Communications Device Class* (CDC) models. One of the standards
404 for cable modem interoperability even specifies the use of this ethernet
405 model as one of two mandatory options. Gadgets using this code look to a
406 USB host as if they're an Ethernet adapter. It provides access to a
407 network where the gadget's CPU is one host, which could easily be
408 bridging, routing, or firewalling access to other networks. Since some
409 hardware can't fully implement the CDC Ethernet requirements, this
410 driver also implements a "good parts only" subset of CDC Ethernet. (That
411 subset doesn't advertise itself as CDC Ethernet, to avoid creating
414 Support for Microsoft's ``RNDIS`` protocol has been contributed by
415 Pengutronix and Auerswald GmbH. This is like CDC Ethernet, but it runs
416 on more slightly USB hardware (but less than the CDC subset). However,
417 its main claim to fame is being able to connect directly to recent
418 versions of Windows, using drivers that Microsoft bundles and supports,
419 making it much simpler to network with Windows.
421 There is also support for user mode gadget drivers, using ``gadgetfs``.
422 This provides a *User Mode API* that presents each endpoint as a single
423 file descriptor. I/O is done using normal ``read()`` and ``read()`` calls.
424 Familiar tools like GDB and pthreads can be used to develop and debug
425 user mode drivers, so that once a robust controller driver is available
426 many applications for it won't require new kernel mode software. Linux
427 2.6 *Async I/O (AIO)* support is available, so that user mode software
428 can stream data with only slightly more overhead than a kernel driver.
430 There's a USB Mass Storage class driver, which provides a different
431 solution for interoperability with systems such as MS-Windows and MacOS.
432 That *Mass Storage* driver uses a file or block device as backing store
433 for a drive, like the ``loop`` driver. The USB host uses the BBB, CB, or
434 CBI versions of the mass storage class specification, using transparent
435 SCSI commands to access the data from the backing store.
437 There's a "serial line" driver, useful for TTY style operation over USB.
438 The latest version of that driver supports CDC ACM style operation, like
439 a USB modem, and so on most hardware it can interoperate easily with
440 MS-Windows. One interesting use of that driver is in boot firmware (like
441 a BIOS), which can sometimes use that model with very small systems
442 without real serial lines.
444 Support for other kinds of gadget is expected to be developed and
445 contributed over time, as this driver framework evolves.
450 USB OTG support on Linux 2.6 was initially developed by Texas
451 Instruments for `OMAP <http://www.omap.com>`__ 16xx and 17xx series
452 processors. Other OTG systems should work in similar ways, but the
453 hardware level details could be very different.
455 Systems need specialized hardware support to implement OTG, notably
456 including a special *Mini-AB* jack and associated transceiver to support
457 *Dual-Role* operation: they can act either as a host, using the standard
458 Linux-USB host side driver stack, or as a peripheral, using this
459 ``gadget`` framework. To do that, the system software relies on small
460 additions to those programming interfaces, and on a new internal
461 component (here called an "OTG Controller") affecting which driver stack
462 connects to the OTG port. In each role, the system can re-use the
463 existing pool of hardware-neutral drivers, layered on top of the
464 controller driver interfaces (:c:type:`usb_bus` or :c:type:`usb_gadget`).
465 Such drivers need at most minor changes, and most of the calls added to
466 support OTG can also benefit non-OTG products.
468 - Gadget drivers test the ``is_otg`` flag, and use it to determine
469 whether or not to include an OTG descriptor in each of their
472 - Gadget drivers may need changes to support the two new OTG protocols,
473 exposed in new gadget attributes such as ``b_hnp_enable`` flag. HNP
474 support should be reported through a user interface (two LEDs could
475 suffice), and is triggered in some cases when the host suspends the
476 peripheral. SRP support can be user-initiated just like remote
477 wakeup, probably by pressing the same button.
479 - On the host side, USB device drivers need to be taught to trigger HNP
480 at appropriate moments, using ``usb_suspend_device()``. That also
481 conserves battery power, which is useful even for non-OTG
484 - Also on the host side, a driver must support the OTG "Targeted
485 Peripheral List". That's just a whitelist, used to reject peripherals
486 not supported with a given Linux OTG host. *This whitelist is
487 product-specific; each product must modify* ``otg_whitelist.h`` *to
488 match its interoperability specification.*
490 Non-OTG Linux hosts, like PCs and workstations, normally have some
491 solution for adding drivers, so that peripherals that aren't
492 recognized can eventually be supported. That approach is unreasonable
493 for consumer products that may never have their firmware upgraded,
494 and where it's usually unrealistic to expect traditional
495 PC/workstation/server kinds of support model to work. For example,
496 it's often impractical to change device firmware once the product has
497 been distributed, so driver bugs can't normally be fixed if they're
498 found after shipment.
500 Additional changes are needed below those hardware-neutral :c:type:`usb_bus`
501 and :c:type:`usb_gadget` driver interfaces; those aren't discussed here in any
502 detail. Those affect the hardware-specific code for each USB Host or
503 Peripheral controller, and how the HCD initializes (since OTG can be
504 active only on a single port). They also involve what may be called an
505 *OTG Controller Driver*, managing the OTG transceiver and the OTG state
506 machine logic as well as much of the root hub behavior for the OTG port.
507 The OTG controller driver needs to activate and deactivate USB
508 controllers depending on the relevant device role. Some related changes
509 were needed inside usbcore, so that it can identify OTG-capable devices
510 and respond appropriately to HNP or SRP protocols.