diff options
| author | Mauro Carvalho Chehab <mchehab@s-opensource.com> | 2017-04-05 10:22:57 -0300 |
|---|---|---|
| committer | Jonathan Corbet <corbet@lwn.net> | 2017-04-11 14:37:04 -0600 |
| commit | 4ad4b21b1b81ce215c1d45850bd5a67e2179c60a (patch) | |
| tree | 96ff7ca62e81affaf9f1b20247c3d73f50e55a56 /Documentation/driver-api/usb | |
| parent | d76a085bc87f68c5098e0150973e0b319a258a8c (diff) | |
docs-rst: convert usb docbooks to ReST
As we're moving out of DocBook, let's convert the remaining
USB docbooks to ReST.
The transformation itself on this patch is a no-brainer
conversion using pandoc via this script:
Documentation/sphinx/tmplcvt
Signed-off-by: Mauro Carvalho Chehab <mchehab@s-opensource.com>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Jonathan Corbet <corbet@lwn.net>
Diffstat (limited to 'Documentation/driver-api/usb')
| -rw-r--r-- | Documentation/driver-api/usb/gadget.rst | 533 | ||||
| -rw-r--r-- | Documentation/driver-api/usb/index.rst | 17 | ||||
| -rw-r--r-- | Documentation/driver-api/usb/usb.rst | 748 | ||||
| -rw-r--r-- | Documentation/driver-api/usb/writing_musb_glue_layer.rst | 737 | ||||
| -rw-r--r-- | Documentation/driver-api/usb/writing_usb_driver.rst | 344 |
5 files changed, 2379 insertions, 0 deletions
diff --git a/Documentation/driver-api/usb/gadget.rst b/Documentation/driver-api/usb/gadget.rst new file mode 100644 index 000000000000..52b299b1ca6d --- /dev/null +++ b/Documentation/driver-api/usb/gadget.rst @@ -0,0 +1,533 @@ +======================== +USB Gadget API for Linux +======================== + +:Author: David Brownell +:Date: 20 August 2004 + +Introduction +============ + +This document presents a Linux-USB "Gadget" kernel mode API, for use +within peripherals and other USB devices that embed Linux. It provides +an overview of the API structure, and shows how that fits into a system +development project. This is the first such API released on Linux to +address a number of important problems, including: + +- Supports USB 2.0, for high speed devices which can stream data at + several dozen megabytes per second. + +- Handles devices with dozens of endpoints just as well as ones with + just two fixed-function ones. Gadget drivers can be written so + they're easy to port to new hardware. + +- Flexible enough to expose more complex USB device capabilities such + as multiple configurations, multiple interfaces, composite devices, + and alternate interface settings. + +- USB "On-The-Go" (OTG) support, in conjunction with updates to the + Linux-USB host side. + +- Sharing data structures and API models with the Linux-USB host side + API. This helps the OTG support, and looks forward to more-symmetric + frameworks (where the same I/O model is used by both host and device + side drivers). + +- Minimalist, so it's easier to support new device controller hardware. + I/O processing doesn't imply large demands for memory or CPU + resources. + +Most Linux developers will not be able to use this API, since they have +USB "host" hardware in a PC, workstation, or server. Linux users with +embedded systems are more likely to have USB peripheral hardware. To +distinguish drivers running inside such hardware from the more familiar +Linux "USB device drivers", which are host side proxies for the real USB +devices, a different term is used: the drivers inside the peripherals +are "USB gadget drivers". In USB protocol interactions, the device +driver is the master (or "client driver") and the gadget driver is the +slave (or "function driver"). + +The gadget API resembles the host side Linux-USB API in that both use +queues of request objects to package I/O buffers, and those requests may +be submitted or canceled. They share common definitions for the standard +USB *Chapter 9* messages, structures, and constants. Also, both APIs +bind and unbind drivers to devices. The APIs differ in detail, since the +host side's current URB framework exposes a number of implementation +details and assumptions that are inappropriate for a gadget API. While +the model for control transfers and configuration management is +necessarily different (one side is a hardware-neutral master, the other +is a hardware-aware slave), the endpoint I/0 API used here should also +be usable for an overhead-reduced host side API. + +Structure of Gadget Drivers +=========================== + +A system running inside a USB peripheral normally has at least three +layers inside the kernel to handle USB protocol processing, and may have +additional layers in user space code. The "gadget" API is used by the +middle layer to interact with the lowest level (which directly handles +hardware). + +In Linux, from the bottom up, these layers are: + +*USB Controller Driver* + This is the lowest software level. It is the only layer that talks + to hardware, through registers, fifos, dma, irqs, and the like. The + ``<linux/usb/gadget.h>`` API abstracts the peripheral controller + endpoint hardware. That hardware is exposed through endpoint + objects, which accept streams of IN/OUT buffers, and through + callbacks that interact with gadget drivers. Since normal USB + devices only have one upstream port, they only have one of these + drivers. The controller driver can support any number of different + gadget drivers, but only one of them can be used at a time. + + Examples of such controller hardware include the PCI-based NetChip + 2280 USB 2.0 high speed controller, the SA-11x0 or PXA-25x UDC + (found within many PDAs), and a variety of other products. + +*Gadget Driver* + The lower boundary of this driver implements hardware-neutral USB + functions, using calls to the controller driver. Because such + hardware varies widely in capabilities and restrictions, and is used + in embedded environments where space is at a premium, the gadget + driver is often configured at compile time to work with endpoints + supported by one particular controller. Gadget drivers may be + portable to several different controllers, using conditional + compilation. (Recent kernels substantially simplify the work + involved in supporting new hardware, by *autoconfiguring* endpoints + automatically for many bulk-oriented drivers.) Gadget driver + responsibilities include: + + - handling setup requests (ep0 protocol responses) possibly + including class-specific functionality + + - returning configuration and string descriptors + + - (re)setting configurations and interface altsettings, including + enabling and configuring endpoints + + - handling life cycle events, such as managing bindings to + hardware, USB suspend/resume, remote wakeup, and disconnection + from the USB host. + + - managing IN and OUT transfers on all currently enabled endpoints + + Such drivers may be modules of proprietary code, although that + approach is discouraged in the Linux community. + +*Upper Level* + Most gadget drivers have an upper boundary that connects to some + Linux driver or framework in Linux. Through that boundary flows the + data which the gadget driver produces and/or consumes through + protocol transfers over USB. Examples include: + + - user mode code, using generic (gadgetfs) or application specific + files in ``/dev`` + + - networking subsystem (for network gadgets, like the CDC Ethernet + Model gadget driver) + + - data capture drivers, perhaps video4Linux or a scanner driver; or + test and measurement hardware. + + - input subsystem (for HID gadgets) + + - sound subsystem (for audio gadgets) + + - file system (for PTP gadgets) + + - block i/o subsystem (for usb-storage gadgets) + + - ... and more + +*Additional Layers* + Other layers may exist. These could include kernel layers, such as + network protocol stacks, as well as user mode applications building + on standard POSIX system call APIs such as *open()*, *close()*, + *read()* and *write()*. On newer systems, POSIX Async I/O calls may + be an option. Such user mode code will not necessarily be subject to + the GNU General Public License (GPL). + +OTG-capable systems will also need to include a standard Linux-USB host +side stack, with *usbcore*, one or more *Host Controller Drivers* +(HCDs), *USB Device Drivers* to support the OTG "Targeted Peripheral +List", and so forth. There will also be an *OTG Controller Driver*, +which is visible to gadget and device driver developers only indirectly. +That helps the host and device side USB controllers implement the two +new OTG protocols (HNP and SRP). Roles switch (host to peripheral, or +vice versa) using HNP during USB suspend processing, and SRP can be +viewed as a more battery-friendly kind of device wakeup protocol. + +Over time, reusable utilities are evolving to help make some gadget +driver tasks simpler. For example, building configuration descriptors +from vectors of descriptors for the configurations interfaces and +endpoints is now automated, and many drivers now use autoconfiguration +to choose hardware endpoints and initialize their descriptors. A +potential example of particular interest is code implementing standard +USB-IF protocols for HID, networking, storage, or audio classes. Some +developers are interested in KDB or KGDB hooks, to let target hardware +be remotely debugged. Most such USB protocol code doesn't need to be +hardware-specific, any more than network protocols like X11, HTTP, or +NFS are. Such gadget-side interface drivers should eventually be +combined, to implement composite devices. + +Kernel Mode Gadget API +====================== + +Gadget drivers declare themselves through a *struct +usb_gadget_driver*, which is responsible for most parts of enumeration +for a *struct usb_gadget*. The response to a set_configuration usually +involves enabling one or more of the *struct usb_ep* objects exposed by +the gadget, and submitting one or more *struct usb_request* buffers to +transfer data. Understand those four data types, and their operations, +and you will understand how this API works. + + **Note** + + This documentation was prepared using the standard Linux kernel + ``docproc`` tool, which turns text and in-code comments into SGML + DocBook and then into usable formats such as HTML or PDF. Other than + the "Chapter 9" data types, most of the significant data types and + functions are described here. + + However, docproc does not understand all the C constructs that are + used, so some relevant information is likely omitted from what you + are reading. One example of such information is endpoint + autoconfiguration. You'll have to read the header file, and use + example source code (such as that for "Gadget Zero"), to fully + understand the API. + + The part of the API implementing some basic driver capabilities is + specific to the version of the Linux kernel that's in use. The 2.6 + kernel includes a *driver model* framework that has no analogue on + earlier kernels; so those parts of the gadget API are not fully + portable. (They are implemented on 2.4 kernels, but in a different + way.) The driver model state is another part of this API that is + ignored by the kerneldoc tools. + +The core API does not expose every possible hardware feature, only the +most widely available ones. There are significant hardware features, +such as device-to-device DMA (without temporary storage in a memory +buffer) that would be added using hardware-specific APIs. + +This API allows drivers to use conditional compilation to handle +endpoint capabilities of different hardware, but doesn't require that. +Hardware tends to have arbitrary restrictions, relating to transfer +types, addressing, packet sizes, buffering, and availability. As a rule, +such differences only matter for "endpoint zero" logic that handles +device configuration and management. The API supports limited run-time +detection of capabilities, through naming conventions for endpoints. +Many drivers will be able to at least partially autoconfigure +themselves. In particular, driver init sections will often have endpoint +autoconfiguration logic that scans the hardware's list of endpoints to +find ones matching the driver requirements (relying on those +conventions), to eliminate some of the most common reasons for +conditional compilation. + +Like the Linux-USB host side API, this API exposes the "chunky" nature +of USB messages: I/O requests are in terms of one or more "packets", and +packet boundaries are visible to drivers. Compared to RS-232 serial +protocols, USB resembles synchronous protocols like HDLC (N bytes per +frame, multipoint addressing, host as the primary station and devices as +secondary stations) more than asynchronous ones (tty style: 8 data bits +per frame, no parity, one stop bit). So for example the controller +drivers won't buffer two single byte writes into a single two-byte USB +IN packet, although gadget drivers may do so when they implement +protocols where packet boundaries (and "short packets") are not +significant. + +Driver Life Cycle +----------------- + +Gadget drivers make endpoint I/O requests to hardware without needing to +know many details of the hardware, but driver setup/configuration code +needs to handle some differences. Use the API like this: + +1. Register a driver for the particular device side usb controller + hardware, such as the net2280 on PCI (USB 2.0), sa11x0 or pxa25x as + found in Linux PDAs, and so on. At this point the device is logically + in the USB ch9 initial state ("attached"), drawing no power and not + usable (since it does not yet support enumeration). Any host should + not see the device, since it's not activated the data line pullup + used by the host to detect a device, even if VBUS power is available. + +2. Register a gadget driver that implements some higher level device + function. That will then bind() to a usb_gadget, which activates the + data line pullup sometime after detecting VBUS. + +3. The hardware driver can now start enumerating. The steps it handles + are to accept USB power and set_address requests. Other steps are + handled by the gadget driver. If the gadget driver module is unloaded + before the host starts to enumerate, steps before step 7 are skipped. + +4. The gadget driver's setup() call returns usb descriptors, based both + on what the bus interface hardware provides and on the functionality + being implemented. That can involve alternate settings or + configurations, unless the hardware prevents such operation. For OTG + devices, each configuration descriptor includes an OTG descriptor. + +5. The gadget driver handles the last step of enumeration, when the USB + host issues a set_configuration call. It enables all endpoints used + in that configuration, with all interfaces in their default settings. + That involves using a list of the hardware's endpoints, enabling each + endpoint according to its descriptor. It may also involve using + :c:func:`usb_gadget_vbus_draw()` to let more power be drawn + from VBUS, as allowed by that configuration. For OTG devices, setting + a configuration may also involve reporting HNP capabilities through a + user interface. + +6. Do real work and perform data transfers, possibly involving changes + to interface settings or switching to new configurations, until the + device is disconnect()ed from the host. Queue any number of transfer + requests to each endpoint. It may be suspended and resumed several + times before being disconnected. On disconnect, the drivers go back + to step 3 (above). + +7. When the gadget driver module is being unloaded, the driver unbind() + callback is issued. That lets the controller driver be unloaded. + +Drivers will normally be arranged so that just loading the gadget driver +module (or statically linking it into a Linux kernel) allows the +peripheral device to be enumerated, but some drivers will defer +enumeration until some higher level component (like a user mode daemon) +enables it. Note that at this lowest level there are no policies about +how ep0 configuration logic is implemented, except that it should obey +USB specifications. Such issues are in the domain of gadget drivers, +including knowing about implementation constraints imposed by some USB +controllers or understanding that composite devices might happen to be +built by integrating reusable components. + +Note that the lifecycle above can be slightly different for OTG devices. +Other than providing an additional OTG descriptor in each configuration, +only the HNP-related differences are particularly visible to driver +code. They involve reporting requirements during the SET_CONFIGURATION +request, and the option to invoke HNP during some suspend callbacks. +Also, SRP changes the semantics of :c:func:`usb_gadget_wakeup()` +slightly. + +USB 2.0 Chapter 9 Types and Constants +------------------------------------- + +Gadget drivers rely on common USB structures and constants defined in +the ``<linux/usb/ch9.h>`` header file, which is standard in Linux 2.6 +kernels. These are the same types and constants used by host side +drivers (and usbcore). + +.. kernel-doc:: include/linux/usb/ch9.h + :internal: + +Core Objects and Methods +------------------------ + +These are declared in ``<linux/usb/gadget.h>``, and are used by gadget +drivers to interact with USB peripheral controller drivers. + +.. kernel-doc:: include/linux/usb/gadget.h + :internal: + +Optional Utilities +------------------ + +The core API is sufficient for writing a USB Gadget Driver, but some +optional utilities are provided to simplify common tasks. These +utilities include endpoint autoconfiguration. + +.. kernel-doc:: drivers/usb/gadget/usbstring.c + :export: + +.. kernel-doc:: drivers/usb/gadget/config.c + :export: + +Composite Device Framework +-------------------------- + +The core API is sufficient for writing drivers for composite USB devices +(with more than one function in a given configuration), and also +multi-configuration devices (also more than one function, but not +necessarily sharing a given configuration). There is however an optional +framework which makes it easier to reuse and combine functions. + +Devices using this framework provide a *struct usb_composite_driver*, +which in turn provides one or more *struct usb_configuration* +instances. Each such configuration includes at least one *struct +usb_function*, which packages a user visible role such as "network +link" or "mass storage device". Management functions may also exist, +such as "Device Firmware Upgrade". + +.. kernel-doc:: include/linux/usb/composite.h + :internal: + +.. kernel-doc:: drivers/usb/gadget/composite.c + :export: + +Composite Device Functions +-------------------------- + +At this writing, a few of the current gadget drivers have been converted +to this framework. Near-term plans include converting all of them, +except for "gadgetfs". + +.. kernel-doc:: drivers/usb/gadget/function/f_acm.c + :export: + +.. kernel-doc:: drivers/usb/gadget/function/f_ecm.c + :export: + +.. kernel-doc:: drivers/usb/gadget/function/f_subset.c + :export: + +.. kernel-doc:: drivers/usb/gadget/function/f_obex.c + :export: + +.. kernel-doc:: drivers/usb/gadget/function/f_serial.c + :export: + +Peripheral Controller Drivers +============================= + +The first hardware supporting this API was the NetChip 2280 controller, +which supports USB 2.0 high speed and is based on PCI. This is the +``net2280`` driver module. The driver supports Linux kernel versions 2.4 +and 2.6; contact NetChip Technologies for development boards and product +information. + +Other hardware working in the "gadget" framework includes: Intel's PXA +25x and IXP42x series processors (``pxa2xx_udc``), Toshiba TC86c001 +"Goku-S" (``goku_udc``), Renesas SH7705/7727 (``sh_udc``), MediaQ 11xx +(``mq11xx_udc``), Hynix HMS30C7202 (``h7202_udc``), National 9303/4 +(``n9604_udc``), Texas Instruments OMAP (``omap_udc``), Sharp LH7A40x +(``lh7a40x_udc``), and more. Most of those are full speed controllers. + +At this writing, there are people at work on drivers in this framework +for several other USB device controllers, with plans to make many of +them be widely available. + +A partial USB simulator, the ``dummy_hcd`` driver, is available. It can +act like a net2280, a pxa25x, or an sa11x0 in terms of available +endpoints and device speeds; and it simulates control, bulk, and to some +extent interrupt transfers. That lets you develop some parts of a gadget +driver on a normal PC, without any special hardware, and perhaps with +the assistance of tools such as GDB running with User Mode Linux. At +least one person has expressed interest in adapting that approach, +hooking it up to a simulator for a microcontroller. Such simulators can +help debug subsystems where the runtime hardware is unfriendly to +software development, or is not yet available. + +Support for other controllers is expected to be developed and +contributed over time, as this driver framework evolves. + +Gadget Drivers +============== + +In addition to *Gadget Zero* (used primarily for testing and development +with drivers for usb controller hardware), other gadget drivers exist. + +There's an *ethernet* gadget driver, which implements one of the most +useful *Communications Device Class* (CDC) models. One of the standards +for cable modem interoperability even specifies the use of this ethernet +model as one of two mandatory options. Gadgets using this code look to a +USB host as if they're an Ethernet adapter. It provides access to a +network where the gadget's CPU is one host, which could easily be +bridging, routing, or firewalling access to other networks. Since some +hardware can't fully implement the CDC Ethernet requirements, this +driver also implements a "good parts only" subset of CDC Ethernet. (That +subset doesn't advertise itself as CDC Ethernet, to avoid creating +problems.) + +Support for Microsoft's *RNDIS* protocol has been contributed by +Pengutronix and Auerswald GmbH. This is like CDC Ethernet, but it runs +on more slightly USB hardware (but less than the CDC subset). However, +its main claim to fame is being able to connect directly to recent +versions of Windows, using drivers that Microsoft bundles and supports, +making it much simpler to network with Windows. + +There is also support for user mode gadget drivers, using *gadgetfs*. +This provides a *User Mode API* that presents each endpoint as a single +file descriptor. I/O is done using normal *read()* and *read()* calls. +Familiar tools like GDB and pthreads can be used to develop and debug +user mode drivers, so that once a robust controller driver is available +many applications for it won't require new kernel mode software. Linux +2.6 *Async I/O (AIO)* support is available, so that user mode software +can stream data with only slightly more overhead than a kernel driver. + +There's a USB Mass Storage class driver, which provides a different +solution for interoperability with systems such as MS-Windows and MacOS. +That *Mass Storage* driver uses a file or block device as backing store +for a drive, like the ``loop`` driver. The USB host uses the BBB, CB, or +CBI versions of the mass storage class specification, using transparent +SCSI commands to access the data from the backing store. + +There's a "serial line" driver, useful for TTY style operation over USB. +The latest version of that driver supports CDC ACM style operation, like +a USB modem, and so on most hardware it can interoperate easily with +MS-Windows. One interesting use of that driver is in boot firmware (like +a BIOS), which can sometimes use that model with very small systems +without real serial lines. + +Support for other kinds of gadget is expected to be developed and +contributed over time, as this driver framework evolves. + +USB On-The-GO (OTG) +=================== + +USB OTG support on Linux 2.6 was initially developed by Texas +Instruments for `OMAP <http://www.omap.com>`__ 16xx and 17xx series +processors. Other OTG systems should work in similar ways, but the +hardware level details could be very different. + +Systems need specialized hardware support to implement OTG, notably +including a special *Mini-AB* jack and associated transceiver to support +*Dual-Role* operation: they can act either as a host, using the standard +Linux-USB host side driver stack, or as a peripheral, using this +"gadget" framework. To do that, the system software relies on small +additions to those programming interfaces, and on a new internal +component (here called an "OTG Controller") affecting which driver stack +connects to the OTG port. In each role, the system can re-use the +existing pool of hardware-neutral drivers, layered on top of the +controller driver interfaces (*usb_bus* or *usb_gadget*). Such drivers +need at most minor changes, and most of the calls added to support OTG +can also benefit non-OTG products. + +- Gadget drivers test the *is_otg* flag, and use it to determine + whether or not to include an OTG descriptor in each of their + configurations. + +- Gadget drivers may need changes to support the two new OTG protocols, + exposed in new gadget attributes such as *b_hnp_enable* flag. HNP + support should be reported through a user interface (two LEDs could + suffice), and is triggered in some cases when the host suspends the + peripheral. SRP support can be user-initiated just like remote + wakeup, probably by pressing the same button. + +- On the host side, USB device drivers need to be taught to trigger HNP + at appropriate moments, using :c:func:`usb_suspend_device()`. + That also conserves battery power, which is useful even for non-OTG + configurations. + +- Also on the host side, a driver must support the OTG "Targeted + Peripheral List". That's just a whitelist, used to reject peripherals + not supported with a given Linux OTG host. *This whitelist is + product-specific; each product must modify ``otg_whitelist.h`` to + match its interoperability specification.* + + Non-OTG Linux hosts, like PCs and workstations, normally have some + solution for adding drivers, so that peripherals that aren't + recognized can eventually be supported. That approach is unreasonable + for consumer products that may never have their firmware upgraded, + and where it's usually unrealistic to expect traditional + PC/workstation/server kinds of support model to work. For example, + it's often impractical to change device firmware once the product has + been distributed, so driver bugs can't normally be fixed if they're + found after shipment. + +Additional changes are needed below those hardware-neutral *usb_bus* +and *usb_gadget* driver interfaces; those aren't discussed here in any +detail. Those affect the hardware-specific code for each USB Host or +Peripheral controller, and how the HCD initializes (since OTG can be +active only on a single port). They also involve what may be called an +*OTG Controller Driver*, managing the OTG transceiver and the OTG state +machine logic as well as much of the root hub behavior for the OTG port. +The OTG controller driver needs to activate and deactivate USB +controllers depending on the relevant device role. Some related changes +were needed inside usbcore, so that it can identify OTG-capable devices +and respond appropriately to HNP or SRP protocols. diff --git a/Documentation/driver-api/usb/index.rst b/Documentation/driver-api/usb/index.rst new file mode 100644 index 000000000000..cf2fa2e8d236 --- /dev/null +++ b/Documentation/driver-api/usb/index.rst @@ -0,0 +1,17 @@ +============= +Linux USB API +============= + +.. toctree:: + + usb + gadget + writing_usb_driver + writing_musb_glue_layer + +.. only:: subproject and html + + Indices + ======= + + * :ref:`genindex` diff --git a/Documentation/driver-api/usb/usb.rst b/Documentation/driver-api/usb/usb.rst new file mode 100644 index 000000000000..b856abb3200e --- /dev/null +++ b/Documentation/driver-api/usb/usb.rst @@ -0,0 +1,748 @@ +=========================== +The Linux-USB Host Side API +=========================== + +Introduction to USB on Linux +============================ + +A Universal Serial Bus (USB) is used to connect a host, such as a PC or +workstation, to a number of peripheral devices. USB uses a tree +structure, with the host as the root (the system's master), hubs as +interior nodes, and peripherals as leaves (and slaves). Modern PCs +support several such trees of USB devices, usually +a few USB 3.0 (5 GBit/s) or USB 3.1 (10 GBit/s) and some legacy +USB 2.0 (480 MBit/s) busses just in case. + +That master/slave asymmetry was designed-in for a number of reasons, one +being ease of use. It is not physically possible to mistake upstream and +downstream or it does not matter with a type C plug (or they are built into the +peripheral). Also, the host software doesn't need to deal with +distributed auto-configuration since the pre-designated master node +manages all that. + +Kernel developers added USB support to Linux early in the 2.2 kernel +series and have been developing it further since then. Besides support +for each new generation of USB, various host controllers gained support, +new drivers for peripherals have been added and advanced features for latency +measurement and improved power management introduced. + +Linux can run inside USB devices as well as on the hosts that control +the devices. But USB device drivers running inside those peripherals +don't do the same things as the ones running inside hosts, so they've +been given a different name: *gadget drivers*. This document does not +cover gadget drivers. + +USB Host-Side API Model +======================= + +Host-side drivers for USB devices talk to the "usbcore" APIs. There are +two. One is intended for *general-purpose* drivers (exposed through +driver frameworks), and the other is for drivers that are *part of the +core*. Such core drivers include the *hub* driver (which manages trees +of USB devices) and several different kinds of *host controller +drivers*, which control individual busses. + +The device model seen by USB drivers is relatively complex. + +- USB supports four kinds of data transfers (control, bulk, interrupt, + and isochronous). Two of them (control and bulk) use bandwidth as + it's available, while the other two (interrupt and isochronous) are + scheduled to provide guaranteed bandwidth. + +- The device description model includes one or more "configurations" + per device, only one of which is active at a time. Devices are supposed + to be capable of operating at lower than their top + speeds and may provide a BOS descriptor showing the lowest speed they + remain fully operational at. + +- From USB 3.0 on configurations have one or more "functions", which + provide a common functionality and are grouped together for purposes + of power management. + +- Configurations or functions have one or more "interfaces", each of which may have + "alternate settings". Interfaces may be standardized by USB "Class" + specifications, or may be specific to a vendor or device. + + USB device drivers actually bind to interfaces, not devices. Think of + them as "interface drivers", though you may not see many devices + where the distinction is important. *Most USB devices are simple, + with only one function, one configuration, one interface, and one alternate + setting.* + +- Interfaces have one or more "endpoints", each of which supports one + type and direction of data transfer such as "bulk out" or "interrupt + in". The entire configuration may have up to sixteen endpoints in + each direction, allocated as needed among all the interfaces. + +- Data transfer on USB is packetized; each endpoint has a maximum + packet size. Drivers must often be aware of conventions such as + flagging the end of bulk transfers using "short" (including zero + length) packets. + +- The Linux USB API supports synchronous calls for control and bulk + messages. It also supports asynchronous calls for all kinds of data + transfer, using request structures called "URBs" (USB Request + Blocks). + +Accordingly, the USB Core API exposed to device drivers covers quite a +lot of territory. You'll probably need to consult the USB 3.0 +specification, available online from www.usb.org at no cost, as well as +class or device specifications. + +The only host-side drivers that actually touch hardware (reading/writing +registers, handling IRQs, and so on) are the HCDs. In theory, all HCDs +provide the same functionality through the same API. In practice, that's +becoming more true, but there are still differences +that crop up especially with fault handling on the less common controllers. +Different controllers don't +necessarily report the same aspects of failures, and recovery from +faults (including software-induced ones like unlinking an URB) isn't yet +fully consistent. Device driver authors should make a point of doing +disconnect testing (while the device is active) with each different host +controller driver, to make sure drivers don't have bugs of their own as +well as to make sure they aren't relying on some HCD-specific behavior. + +USB-Standard Types +================== + +In ``<linux/usb/ch9.h>`` you will find the USB data types defined in +chapter 9 of the USB specification. These data types are used throughout +USB, and in APIs including this host side API, gadget APIs, and usbfs. + +.. kernel-doc:: include/linux/usb/ch9.h + :internal: + +Host-Side Data Types and Macros +=============================== + +The host side API exposes several layers to drivers, some of which are +more necessary than others. These support lifecycle models for host side +drivers and devices, and support passing buffers through usbcore to some +HCD that performs the I/O for the device driver. + +.. kernel-doc:: include/linux/usb.h + :internal: + +USB Core APIs +============= + +There are two basic I/O models in the USB API. The most elemental one is +asynchronous: drivers submit requests in the form of an URB, and the +URB's completion callback handles the next step. All USB transfer types +support that model, although there are special cases for control URBs +(which always have setup and status stages, but may not have a data +stage) and isochronous URBs (which allow large packets and include +per-packet fault reports). Built on top of that is synchronous API +support, where a driver calls a routine that allocates one or more URBs, +submits them, and waits until they complete. There are synchronous +wrappers for single-buffer control and bulk transfers (which are awkward +to use in some driver disconnect scenarios), and for scatterlist based +streaming i/o (bulk or interrupt). + +USB drivers need to provide buffers that can be used for DMA, although +they don't necessarily need to provide the DMA mapping themselves. There +are APIs to use used when allocating DMA buffers, which can prevent use +of bounce buffers on some systems. In some cases, drivers may be able to +rely on 64bit DMA to eliminate another kind of bounce buffer. + +.. kernel-doc:: drivers/usb/core/urb.c + :export: + +.. kernel-doc:: drivers/usb/core/message.c + :export: + +.. kernel-doc:: drivers/usb/core/file.c + :export: + +.. kernel-doc:: drivers/usb/core/driver.c + :export: + +.. kernel-doc:: drivers/usb/core/usb.c + :export: + +.. kernel-doc:: drivers/usb/core/hub.c + :export: + +Host Controller APIs +==================== + +These APIs are only for use by host controller drivers, most of which +implement standard register interfaces such as XHCI, EHCI, OHCI, or UHCI. UHCI +was one of the first interfaces, designed by Intel and also used by VIA; +it doesn't do much in hardware. OHCI was designed later, to have the +hardware do more work (bigger transfers, tracking protocol state, and so +on). EHCI was designed with USB 2.0; its design has features that +resemble OHCI (hardware does much more work) as well as UHCI (some parts +of ISO support, TD list processing). XHCI was designed with USB 3.0. It +continues to shift support for functionality into hardware. + +There are host controllers other than the "big three", although most PCI +based controllers (and a few non-PCI based ones) use one of those +interfaces. Not all host controllers use DMA; some use PIO, and there is +also a simulator and a virtual host controller to pipe USB over the network. + +The same basic APIs are available to drivers for all those controllers. +For historical reasons they are in two layers: :c:type:`struct +usb_bus <usb_bus>` is a rather thin layer that became available +in the 2.2 kernels, while :c:type:`struct usb_hcd <usb_hcd>` +is a more featureful layer +that lets HCDs share common code, to shrink driver size and +significantly reduce hcd-specific behaviors. + +.. kernel-doc:: drivers/usb/core/hcd.c + :export: + +.. kernel-doc:: drivers/usb/core/hcd-pci.c + :export: + +.. kernel-doc:: drivers/usb/core/buffer.c + :internal: + +The USB Filesystem (usbfs) +========================== + +This chapter presents the Linux *usbfs*. You may prefer to avoid writing +new kernel code for your USB driver; that's the problem that usbfs set +out to solve. User mode device drivers are usually packaged as +applications or libraries, and may use usbfs through some programming +library that wraps it. Such libraries include +`libusb <http://libusb.sourceforge.net>`__ for C/C++, and +`jUSB <http://jUSB.sourceforge.net>`__ for Java. + + **Note** + + This particular documentation is incomplete, especially with respect + to the asynchronous mode. As of kernel 2.5.66 the code and this + (new) documentation need to be cross-reviewed. + +Configure usbfs into Linux kernels by enabling the *USB filesystem* +option (CONFIG_USB_DEVICEFS), and you get basic support for user mode +USB device drivers. Until relatively recently it was often (confusingly) +called *usbdevfs* although it wasn't solving what *devfs* was. Every USB +device will appear in usbfs, regardless of whether or not it has a +kernel driver. + +What files are in "usbfs"? +-------------------------- + +Conventionally mounted at ``/proc/bus/usb``, usbfs features include: + +- ``/proc/bus/usb/devices`` ... a text file showing each of the USB + devices on known to the kernel, and their configuration descriptors. + You can also poll() this to learn about new devices. + +- ``/proc/bus/usb/BBB/DDD`` ... magic files exposing the each device's + configuration descriptors, and supporting a series of ioctls for + making device requests, including I/O to devices. (Purely for access + by programs.) + +Each bus is given a number (BBB) based on when it was enumerated; within +each bus, each device is given a similar number (DDD). Those BBB/DDD +paths are not "stable" identifiers; expect them to change even if you +always leave the devices plugged in to the same hub port. *Don't even +think of saving these in application configuration files.* Stable +identifiers are available, for user mode applications that want to use +them. HID and networking devices expose these stable IDs, so that for +example you can be sure that you told the right UPS to power down its +second server. "usbfs" doesn't (yet) expose those IDs. + +Mounting and Access Control +--------------------------- + +There are a number of mount options for usbfs, which will be of most +interest to you if you need to override the default access control +policy. That policy is that only root may read or write device files +(``/proc/bus/BBB/DDD``) although anyone may read the ``devices`` or +``drivers`` files. I/O requests to the device also need the +CAP_SYS_RAWIO capability, + +The significance of that is that by default, all user mode device +drivers need super-user privileges. You can change modes or ownership in +a driver setup when the device hotplugs, or maye just start the driver +right then, as a privileged server (or some activity within one). That's +the most secure approach for multi-user systems, but for single user +systems ("trusted" by that user) it's more convenient just to grant +everyone all access (using the *devmode=0666* option) so the driver can +start whenever it's needed. + +The mount options for usbfs, usable in /etc/fstab or in command line +invocations of *mount*, are: + +*busgid*\ =NNNNN + Controls the GID used for the /proc/bus/usb/BBB directories. + (Default: 0) + +*busmode*\ =MMM + Controls the file mode used for the /proc/bus/usb/BBB directories. + (Default: 0555) + +*busuid*\ =NNNNN + Controls the UID used for the /proc/bus/usb/BBB directories. + (Default: 0) + +*devgid*\ =NNNNN + Controls the GID used for the /proc/bus/usb/BBB/DDD files. (Default: + 0) + +*devmode*\ =MMM + Controls the file mode used for the /proc/bus/usb/BBB/DDD files. + (Default: 0644) + +*devuid*\ =NNNNN + Controls the UID used for the /proc/bus/usb/BBB/DDD files. (Default: + 0) + +*listgid*\ =NNNNN + Controls the GID used for the /proc/bus/usb/devices and drivers + files. (Default: 0) + +*listmode*\ =MMM + Controls the file mode used for the /proc/bus/usb/devices and + drivers files. (Default: 0444) + +*listuid*\ =NNNNN + Controls the UID used for the /proc/bus/usb/devices and drivers + files. (Default: 0) + +Note that many Linux distributions hard-wire the mount options for usbfs +in their init scripts, such as ``/etc/rc.d/rc.sysinit``, rather than +making it easy to set this per-system policy in ``/etc/fstab``. + +/proc/bus/usb/devices +--------------------- + +This file is handy for status viewing tools in user mode, which can scan +the text format and ignore most of it. More detailed device status +(including class and vendor status) is available from device-specific +files. For information about the current format of this file, see the +``Documentation/usb/proc_usb_info.txt`` file in your Linux kernel +sources. + +This file, in combination with the poll() system call, can also be used +to detect when devices are added or removed: + +:: + + int fd; + struct pollfd pfd; + + fd = open("/proc/bus/usb/devices", O_RDONLY); + pfd = { fd, POLLIN, 0 }; + for (;;) { + /* The first time through, this call will return immediately. */ + poll(&pfd, 1, -1); + + /* To see what's changed, compare the file's previous and current + contents or scan the filesystem. (Scanning is more precise.) */ + } + +Note that this behavior is intended to be used for informational and +debug purposes. It would be more appropriate to use programs such as +udev or HAL to initialize a device or start a user-mode helper program, +for instance. + +/proc/bus/usb/BBB/DDD +--------------------- + +Use these files in one of these basic ways: + +*They can be read,* producing first the device descriptor (18 bytes) and +then the descriptors for the current configuration. See the USB 2.0 spec +for details about those binary data formats. You'll need to convert most +multibyte values from little endian format to your native host byte +order, although a few of the fields in the device descriptor (both of +the BCD-encoded fields, and the vendor and product IDs) will be +byteswapped for you. Note that configuration descriptors include +descriptors for interfaces, altsettings, endpoints, and maybe additional +class descriptors. + +*Perform USB operations* using *ioctl()* requests to make endpoint I/O +requests (synchronously or asynchronously) or manage the device. These +requests need the CAP_SYS_RAWIO capability, as well as filesystem +access permissions. Only one ioctl request can be made on one of these +device files at a time. This means that if you are synchronously reading +an endpoint from one thread, you won't be able to write to a different +endpoint from another thread until the read completes. This works for +*half duplex* protocols, but otherwise you'd use asynchronous i/o +requests. + +Life Cycle of User Mode Drivers +------------------------------- + +Such a driver first needs to find a device file for a device it knows +how to handle. Maybe it was told about it because a ``/sbin/hotplug`` +event handling agent chose that driver to handle the new device. Or +maybe it's an application that scans all the /proc/bus/usb device files, +and ignores most devices. In either case, it should :c:func:`read()` +all the descriptors from the device file, and check them against what it +knows how to handle. It might just reject everything except a particular +vendor and product ID, or need a more complex policy. + +Never assume there will only be one such device on the system at a time! +If your code can't handle more than one device at a time, at least +detect when there's more than one, and have your users choose which +device to use. + +Once your user mode driver knows what device to use, it interacts with +it in either of two styles. The simple style is to make only control +requests; some devices don't need more complex interactions than those. +(An example might be software using vendor-specific control requests for +some initialization or configuration tasks, with a kernel driver for the +rest.) + +More likely, you need a more complex style driver: one using non-control +endpoints, reading or writing data and claiming exclusive use of an +interface. *Bulk* transfers are easiest to use, but only their sibling +*interrupt* transfers work with low speed devices. Both interrupt and +*isochronous* transfers offer service guarantees because their bandwidth +is reserved. Such "periodic" transfers are awkward to use through usbfs, +unless you're using the asynchronous calls. However, interrupt transfers +can also be used in a synchronous "one shot" style. + +Your user-mode driver should never need to worry about cleaning up +request state when the device is disconnected, although it should close +its open file descriptors as soon as it starts seeing the ENODEV errors. + +The ioctl() Requests +-------------------- + +To use these ioctls, you need to include the following headers in your +userspace program: + +:: + + #include <linux/usb.h> + #include <linux/usbdevice_fs.h> + #include <asm/byteorder.h> + +The standard USB device model requests, from "Chapter 9" of the USB 2.0 +specification, are automatically included from the ``<linux/usb/ch9.h>`` +header. + +Unless noted otherwise, the ioctl requests described here will update +the modification time on the usbfs file to which they are applied +(unless they fail). A return of zero indicates success; otherwise, a +standard USB error code is returned. (These are documented in +``Documentation/usb/error-codes.txt`` in your kernel sources.) + +Each of these files multiplexes access to several I/O streams, one per +endpoint. Each device has one control endpoint (endpoint zero) which +supports a limited RPC style RPC access. Devices are configured by +hub_wq (in the kernel) setting a device-wide *configuration* that +affects things like power consumption and basic functionality. The +endpoints are part of USB *interfaces*, which may have *altsettings* +affecting things like which endpoints are available. Many devices only +have a single configuration and interface, so drivers for them will +ignore configurations and altsettings. + +Management/Status Requests +~~~~~~~~~~~~~~~~~~~~~~~~~~ + +A number of usbfs requests don't deal very directly with device I/O. +They mostly relate to device management and status. These are all +synchronous requests. + +USBDEVFS_CLAIMINTERFACE + This is used to force usbfs to claim a specific interface, which has + not previously been claimed by usbfs or any other kernel driver. The + ioctl parameter is an integer holding the number of the interface + (bInterfaceNumber from descriptor). + + Note that if your driver doesn't claim an interface before trying to + use one of its endpoints, and no other driver has bound to it, then + the interface is automatically claimed by usbfs. + + This claim will be released by a RELEASEINTERFACE ioctl, or by + closing the file descriptor. File modification time is not updated + by this request. + +USBDEVFS_CONNECTINFO + Says whether the device is lowspeed. The ioctl parameter points to a + structure like this: + + :: + + struct usbdevfs_connectinfo { + unsigned int devnum; + unsigned char slow; + }; + + File modification time is not updated by this request. + + *You can't tell whether a "not slow" device is connected at high + speed (480 MBit/sec) or just full speed (12 MBit/sec).* You should + know the devnum value already, it's the DDD value of the device file + name. + +USBDEVFS_GETDRIVER + Returns the name of the kernel driver bound to a given interface (a + string). Parameter is a pointer to this structure, which is + modified: + + :: + + struct usbdevfs_getdriver { + unsigned int interface; + char driver[USBDEVFS_MAXDRIVERNAME + 1]; + }; + + File modification time is not updated by this request. + +USBDEVFS_IOCTL + Passes a request from userspace through to a kernel driver that has + an ioctl entry in the *struct usb_driver* it registered. + + :: + + struct usbdevfs_ioctl { + int ifno; + int ioctl_code; + void *data; + }; + + /* user mode call looks like this. + * 'request' becomes the driver->ioctl() 'code' parameter. + * the size of 'param' is encoded in 'request', and that data + * is copied to or from the driver->ioctl() 'buf' parameter. + */ + static int + usbdev_ioctl (int fd, int ifno, unsigned request, void *param) + { + struct usbdevfs_ioctl wrapper; + + wrapper.ifno = ifno; + wrapper.ioctl_code = request; + wrapper.data = param; + + return ioctl (fd, USBDEVFS_IOCTL, &wrapper); + } + + File modification time is not updated by this request. + + This request lets kernel drivers talk to user mode code through + filesystem operations even when they don't create a character or + block special device. It's also been used to do things like ask + devices what device special file should be used. Two pre-defined + ioctls are used to disconnect and reconnect kernel drivers, so that + user mode code can completely manage binding and configuration of + devices. + +USBDEVFS_RELEASEINTERFACE + This is used to release the claim usbfs made on interface, either + implicitly or because of a USBDEVFS_CLAIMINTERFACE call, before the + file descriptor is closed. The ioctl parameter is an integer holding + the number of the interface (bInterfaceNumber from descriptor); File + modification time is not updated by this request. + + **Warning** + + *No security check is made to ensure that the task which made + the claim is the one which is releasing it. This means that user + mode driver may interfere other ones.* + +USBDEVFS_RESETEP + Resets the data toggle value for an endpoint (bulk or interrupt) to + DATA0. The ioctl parameter is an integer endpoint number (1 to 15, + as identified in the endpoint descriptor), with USB_DIR_IN added + if the device's endpoint sends data to the host. + + **Warning** + + *Avoid using this request. It should probably be removed.* Using + it typically means the device and driver will lose toggle + synchronization. If you really lost synchronization, you likely + need to completely handshake with the device, using a request + like CLEAR_HALT or SET_INTERFACE. + +USBDEVFS_DROP_PRIVILEGES + This is used to relinquish the ability to do certain operations + which are considered to be privileged on a usbfs file descriptor. + This includes claiming arbitrary interfaces, resetting a device on + which there are currently claimed interfaces from other users, and + issuing USBDEVFS_IOCTL calls. The ioctl parameter is a 32 bit mask + of interfaces the user is allowed to claim on this file descriptor. + You may issue this ioctl more than one time to narrow said mask. + +Synchronous I/O Support +~~~~~~~~~~~~~~~~~~~~~~~ + +Synchronous requests involve the kernel blocking until the user mode +request completes, either by finishing successfully or by reporting an +error. In most cases this is the simplest way to use usbfs, although as +noted above it does prevent performing I/O to more than one endpoint at +a time. + +USBDEVFS_BULK + Issues a bulk read or write request to the device. The ioctl + parameter is a pointer to this structure: + + :: + + struct usbdevfs_bulktransfer { + unsigned int ep; + unsigned int len; + unsigned int timeout; /* in milliseconds */ + void *data; + }; + + The "ep" value identifies a bulk endpoint number (1 to 15, as + identified in an endpoint descriptor), masked with USB_DIR_IN when + referring to an endpoint which sends data to the host from the + device. The length of the data buffer is identified by "len"; Recent + kernels support requests up to about 128KBytes. *FIXME say how read + length is returned, and how short reads are handled.*. + +USBDEVFS_CLEAR_HALT + Clears endpoint halt (stall) and resets the endpoint toggle. This is + only meaningful for bulk or interrupt endpoints. The ioctl parameter + is an integer endpoint number (1 to 15, as identified in an endpoint + descriptor), masked with USB_DIR_IN when referring to an endpoint + which sends data to the host from the device. + + Use this on bulk or interrupt endpoints which have stalled, + returning *-EPIPE* status to a data transfer request. Do not issue + the control request directly, since that could invalidate the host's + record of the data toggle. + +USBDEVFS_CONTROL + Issues a control request to the device. The ioctl parameter points + to a structure like this: + + :: + + struct usbdevfs_ctrltransfer { + __u8 bRequestType; + __u8 bRequest; + __u16 wValue; + __u16 wIndex; + __u16 wLength; + __u32 timeout; /* in milliseconds */ + void *data; + }; + + The first eight bytes of this structure are the contents of the + SETUP packet to be sent to the device; see the USB 2.0 specification + for details. The bRequestType value is composed by combining a + USB_TYPE_\* value, a USB_DIR_\* value, and a USB_RECIP_\* + value (from *<linux/usb.h>*). If wLength is nonzero, it describes + the length of the data buffer, which is either written to the device + (USB_DIR_OUT) or read from the device (USB_DIR_IN). + + At this writing, you can't transfer more than 4 KBytes of data to or + from a device; usbfs has a limit, and some host controller drivers + have a limit. (That's not usually a problem.) *Also* there's no way + to say it's not OK to get a short read back from the device. + +USBDEVFS_RESET + Does a USB level device reset. The ioctl parameter is ignored. After + the reset, this rebinds all device interfaces. File modification + time is not updated by this request. + + **Warning** + + *Avoid using this call* until some usbcore bugs get fixed, since + it does not fully synchronize device, interface, and driver (not + just usbfs) state. + +USBDEVFS_SETINTERFACE + Sets the alternate setting for an interface. The ioctl parameter is + a pointer to a structure like this: + + :: + + struct usbdevfs_setinterface { + unsigned int interface; + unsigned int altsetting; + }; + + File modification time is not updated by this request. + + Those struct members are from some interface descriptor applying to + the current configuration. The interface number is the + bInterfaceNumber value, and the altsetting number is the + bAlternateSetting value. (This resets each endpoint in the + interface.) + +USBDEVFS_SETCONFIGURATION + Issues the :c:func:`usb_set_configuration()` call for the + device. The parameter is an integer holding the number of a + configuration (bConfigurationValue from descriptor). File + modification time is not updated by this request. + + **Warning** + + *Avoid using this call* until some usbcore bugs get fixed, since + it does not fully synchronize device, interface, and driver (not + just usbfs) state. + +Asynchronous I/O Support +~~~~~~~~~~~~~~~~~~~~~~~~ + +As mentioned above, there are situations where it may be important to +initiate concurrent operations from user mode code. This is particularly +important for periodic transfers (interrupt and isochronous), but it can +be used for other kinds of USB requests too. In such cases, the +asynchronous requests described here are essential. Rather than +submitting one request and having the kernel block until it completes, +the blocking is separate. + +These requests are packaged into a structure that resembles the URB used +by kernel device drivers. (No POSIX Async I/O support here, sorry.) It +identifies the endpoint type (USBDEVFS_URB_TYPE_\*), endpoint +(number, masked with USB_DIR_IN as appropriate), buffer and length, +and a user "context" value serving to uniquely identify each request. +(It's usually a pointer to per-request data.) Flags can modify requests +(not as many as supported for kernel drivers). + +Each request can specify a realtime signal number (between SIGRTMIN and +SIGRTMAX, inclusive) to request a signal be sent when the request +completes. + +When usbfs returns these urbs, the status value is updated, and the +buffer may have been modified. Except for isochronous transfers, the +actual_length is updated to say how many bytes were transferred; if the +USBDEVFS_URB_DISABLE_SPD flag is set ("short packets are not OK"), if +fewer bytes were read than were requested then you get an error report. + +:: + + struct usbdevfs_iso_packet_desc { + unsigned int length; + unsigned int actual_length; + unsigned int status; + }; + + struct usbdevfs_urb { + unsigned char type; + unsigned char endpoint; + int status; + unsigned int flags; + void *buffer; + int buffer_length; + int actual_length; + int start_frame; + int number_of_packets; + int error_count; + unsigned int signr; + void *usercontext; + struct usbdevfs_iso_packet_desc iso_frame_desc[]; + }; + +For these asynchronous requests, the file modification time reflects +when the request was initiated. This contrasts with their use with the +synchronous requests, where it reflects when requests complete. + +USBDEVFS_DISCARDURB + *TBS* File modification time is not updated by this request. + +USBDEVFS_DISCSIGNAL + *TBS* File modification time is not updated by this request. + +USBDEVFS_REAPURB + *TBS* File modification time is not updated by this request. + +USBDEVFS_REAPURBNDELAY + *TBS* File modification time is not updated by this request. + +USBDEVFS_SUBMITURB + *TBS* diff --git a/Documentation/driver-api/usb/writing_musb_glue_layer.rst b/Documentation/driver-api/usb/writing_musb_glue_layer.rst new file mode 100644 index 000000000000..52700c7031f9 --- /dev/null +++ b/Documentation/driver-api/usb/writing_musb_glue_layer.rst @@ -0,0 +1,737 @@ +========================== +Writing an MUSB Glue Layer +========================== + +:Author: Apelete Seketeli + +Introduction +============ + +The Linux MUSB subsystem is part of the larger Linux USB subsystem. It +provides support for embedded USB Device Controllers (UDC) that do not +use Universal Host Controller Interface (UHCI) or Open Host Controller +Interface (OHCI). + +Instead, these embedded UDC rely on the USB On-the-Go (OTG) +specification which they implement at least partially. The silicon +reference design used in most cases is the Multipoint USB Highspeed +Dual-Role Controller (MUSB HDRC) found in the Mentor Graphics Inventra™ +design. + +As a self-taught exercise I have written an MUSB glue layer for the +Ingenic JZ4740 SoC, modelled after the many MUSB glue layers in the +kernel source tree. This layer can be found at +drivers/usb/musb/jz4740.c. In this documentation I will walk through the +basics of the jz4740.c glue layer, explaining the different pieces and +what needs to be done in order to write your own device glue layer. + +Linux MUSB Basics +================= + +To get started on the topic, please read USB On-the-Go Basics (see +Resources) which provides an introduction of USB OTG operation at the +hardware level. A couple of wiki pages by Texas Instruments and Analog +Devices also provide an overview of the Linux kernel MUSB configuration, +albeit focused on some specific devices provided by these companies. +Finally, getting acquainted with the USB specification at USB home page +may come in handy, with practical instance provided through the Writing +USB Device Drivers documentation (again, see Resources). + +Linux USB stack is a layered architecture in which the MUSB controller +hardware sits at the lowest. The MUSB controller driver abstract the +MUSB controller hardware to the Linux USB stack. + +:: + + ------------------------ + | | <------- drivers/usb/gadget + | Linux USB Core Stack | <------- drivers/usb/host + | | <------- drivers/usb/core + ------------------------ + ⬍ + -------------------------- + | | <------ drivers/usb/musb/musb_gadget.c + | MUSB Controller driver | <------ drivers/usb/musb/musb_host.c + | | <------ drivers/usb/musb/musb_core.c + -------------------------- + ⬍ + --------------------------------- + | MUSB Platform Specific Driver | + | | <-- drivers/usb/musb/jz4740.c + | aka "Glue Layer" | + --------------------------------- + ⬍ + --------------------------------- + | MUSB Controller Hardware | + --------------------------------- + + +As outlined above, the glue layer is actually the platform specific code +sitting in between the controller driver and the controller hardware. + +Just like a Linux USB driver needs to register itself with the Linux USB +subsystem, the MUSB glue layer needs first to register itself with the +MUSB controller driver. This will allow the controller driver to know +about which device the glue layer supports and which functions to call +when a supported device is detected or released; remember we are talking +about an embedded controller chip here, so no insertion or removal at +run-time. + +All of this information is passed to the MUSB controller driver through +a platform_driver structure defined in the glue layer as: + +:: + + static struct platform_driver jz4740_driver = { + .probe = jz4740_probe, + .remove = jz4740_remove, + .driver = { + .name = "musb-jz4740", + }, + }; + + +The probe and remove function pointers are called when a matching device +is detected and, respectively, released. The name string describes the +device supported by this glue layer. In the current case it matches a +platform_device structure declared in arch/mips/jz4740/platform.c. Note +that we are not using device tree bindings here. + +In order to register itself to the controller driver, the glue layer +goes through a few steps, basically allocating the controller hardware +resources and initialising a couple of circuits. To do so, it needs to +keep track of the information used throughout these steps. This is done +by defining a private jz4740_glue structure: + +:: + + struct jz4740_glue { + struct device *dev; + struct platform_device *musb; + struct clk *clk; + }; + + +The dev and musb members are both device structure variables. The first +one holds generic information about the device, since it's the basic +device structure, and the latter holds information more closely related +to the subsystem the device is registered to. The clk variable keeps +information related to the device clock operation. + +Let's go through the steps of the probe function that leads the glue +layer to register itself to the controller driver. + +N.B.: For the sake of readability each function will be split in logical +parts, each part being shown as if it was independent from the others. + +:: + + static int jz4740_probe(struct platform_device *pdev) + { + struct platform_device *musb; + struct jz4740_glue *glue; + struct clk *clk; + int ret; + + glue = devm_kzalloc(&pdev->dev, sizeof(*glue), GFP_KERNEL); + if (!glue) + return -ENOMEM; + + musb = platform_device_alloc("musb-hdrc", PLATFORM_DEVID_AUTO); + if (!musb) { + dev_err(&pdev->dev, "failed to allocate musb device\n"); + return -ENOMEM; + } + + clk = devm_clk_get(&pdev->dev, "udc"); + if (IS_ERR(clk)) { + dev_err(&pdev->dev, "failed to get clock\n"); + ret = PTR_ERR(clk); + goto err_platform_device_put; + } + + ret = clk_prepare_enable(clk); + if (ret) { + dev_err(&pdev->dev, "failed to enable clock\n"); + goto err_platform_device_put; + } + + musb->dev.parent = &pdev->dev; + + glue->dev = &pdev->dev; + glue->musb = musb; + glue->clk = clk; + + return 0; + + err_platform_device_put: + platform_device_put(musb); + return ret; + } + + +The first few lines of the probe function allocate and assign the glue, +musb and clk variables. The GFP_KERNEL flag (line 8) allows the +allocation process to sleep and wait for memory, thus being usable in a +blocking situation. The PLATFORM_DEVID_AUTO flag (line 12) allows +automatic allocation and management of device IDs in order to avoid +device namespace collisions with explicit IDs. With devm_clk_get() +(line 18) the glue layer allocates the clock -- the ``devm_`` prefix +indicates that clk_get() is managed: it automatically frees the +allocated clock resource data when the device is released -- and enable +it. + +Then comes the registration steps: + +:: + + static int jz4740_probe(struct platform_device *pdev) + { + struct musb_hdrc_platform_data *pdata = &jz4740_musb_platform_data; + + pdata->platform_ops = &jz4740_musb_ops; + + platform_set_drvdata(pdev, glue); + + ret = platform_device_add_resources(musb, pdev->resource, + pdev->num_resources); + if (ret) { + dev_err(&pdev->dev, "failed to add resources\n"); + goto err_clk_disable; + } + + ret = platform_device_add_data(musb, pdata, sizeof(*pdata)); + if (ret) { + dev_err(&pdev->dev, "failed to add platform_data\n"); + goto err_clk_disable; + } + + return 0; + + err_clk_disable: + clk_disable_unprepare(clk); + err_platform_device_put: + platform_device_put(musb); + return ret; + } + + +The first step is to pass the device data privately held by the glue +layer on to the controller driver through platform_set_drvdata() (line +7). Next is passing on the device resources information, also privately +held at that point, through platform_device_add_resources() (line 9). + +Finally comes passing on the platform specific data to the controller +driver (line 16). Platform data will be discussed in `Chapter +4 <#device-platform-data>`__, but here we are looking at the +platform_ops function pointer (line 5) in musb_hdrc_platform_data +structure (line 3). This function pointer allows the MUSB controller +driver to know which function to call for device operation: + +:: + + static const struct musb_platform_ops jz4740_musb_ops = { + .init = jz4740_musb_init, + .exit = jz4740_musb_exit, + }; + + +Here we have the minimal case where only init and exit functions are +called by the controller driver when needed. Fact is the JZ4740 MUSB +controller is a basic controller, lacking some features found in other +controllers, otherwise we may also have pointers to a few other +functions like a power management function or a function to switch +between OTG and non-OTG modes, for instance. + +At that point of the registration process, the controller driver +actually calls the init function: + +:: + + static int jz4740_musb_init(struct musb *musb) + { + musb->xceiv = usb_get_phy(USB_PHY_TYPE_USB2); + if (!musb->xceiv) { + pr_err("HS UDC: no transceiver configured\n"); + return -ENODEV; + } + + /* Silicon does not implement ConfigData register. + * Set dyn_fifo to avoid reading EP config from hardware. + */ + musb->dyn_fifo = true; + + musb->isr = jz4740_musb_interrupt; + + return 0; + } + + +The goal of jz4740_musb_init() is to get hold of the transceiver +driver data of the MUSB controller hardware and pass it on to the MUSB +controller driver, as usual. The transceiver is the circuitry inside the +controller hardware responsible for sending/receiving the USB data. +Since it is an implementation of the physical layer of the OSI model, +the transceiver is also referred to as PHY. + +Getting hold of the MUSB PHY driver data is done with usb_get_phy() +which returns a pointer to the structure containing the driver instance +data. The next couple of instructions (line 12 and 14) are used as a +quirk and to setup IRQ handling respectively. Quirks and IRQ handling +will be discussed later in `Chapter 5 <#device-quirks>`__ and `Chapter +3 <#handling-irqs>`__. + +:: + + static int jz4740_musb_exit(struct musb *musb) + { + usb_put_phy(musb->xceiv); + + return 0; + } + + +Acting as the counterpart of init, the exit function releases the MUSB +PHY driver when the controller hardware itself is about to be released. + +Again, note that init and exit are fairly simple in this case due to the +basic set of features of the JZ4740 controller hardware. When writing an +musb glue layer for a more complex controller hardware, you might need +to take care of more processing in those two functions. + +Returning from the init function, the MUSB controller driver jumps back +into the probe function: + +:: + + static int jz4740_probe(struct platform_device *pdev) + { + ret = platform_device_add(musb); + if (ret) { + dev_err(&pdev->dev, "failed to register musb device\n"); + goto err_clk_disable; + } + + return 0; + + err_clk_disable: + clk_disable_unprepare(clk); + err_platform_device_put: + platform_device_put(musb); + return ret; + } + + +This is the last part of the device registration process where the glue +layer adds the controller hardware device to Linux kernel device +hierarchy: at this stage, all known information about the device is +passed on to the Linux USB core stack. + +:: + + static int jz4740_remove(struct platform_device *pdev) + { + struct jz4740_glue *glue = platform_get_drvdata(pdev); + + platform_device_unregister(glue->musb); + clk_disable_unprepare(glue->clk); + + return 0; + } + + +Acting as the counterpart of probe, the remove function unregister the +MUSB controller hardware (line 5) and disable the clock (line 6), +allowing it to be gated. + +Handling IRQs +============= + +Additionally to the MUSB controller hardware basic setup and +registration, the glue layer is also responsible for handling the IRQs: + +:: + + static irqreturn_t jz4740_musb_interrupt(int irq, void *__hci) + { + unsigned long flags; + irqreturn_t retval = IRQ_NONE; + struct musb *musb = __hci; + + spin_lock_irqsave(&musb->lock, flags); + + musb->int_usb = musb_readb(musb->mregs, MUSB_INTRUSB); + musb->int_tx = musb_readw(musb->mregs, MUSB_INTRTX); + musb->int_rx = musb_readw(musb->mregs, MUSB_INTRRX); + + /* + * The controller is gadget only, the state of the host mode IRQ bits is + * undefined. Mask them to make sure that the musb driver core will + * never see them set + */ + musb->int_usb &= MUSB_INTR_SUSPEND | MUSB_INTR_RESUME | + MUSB_INTR_RESET | MUSB_INTR_SOF; + + if (musb->int_usb || musb->int_tx || musb->int_rx) + retval = musb_interrupt(musb); + + spin_unlock_irqrestore(&musb->lock, flags); + + return retval; + } + + +Here the glue layer mostly has to read the relevant hardware registers +and pass their values on to the controller driver which will handle the +actual event that triggered the IRQ. + +The interrupt handler critical section is protected by the +spin_lock_irqsave() and counterpart spin_unlock_irqrestore() +functions (line 7 and 24 respectively), which prevent the interrupt +handler code to be run by two different threads at the same time. + +Then the relevant interrupt registers are read (line 9 to 11): + +- MUSB_INTRUSB: indicates which USB interrupts are currently active, + +- MUSB_INTRTX: indicates which of the interrupts for TX endpoints are + currently active, + +- MUSB_INTRRX: indicates which of the interrupts for TX endpoints are + currently active. + +Note that musb_readb() is used to read 8-bit registers at most, while +musb_readw() allows us to read at most 16-bit registers. There are +other functions that can be used depending on the size of your device +registers. See musb_io.h for more information. + +Instruction on line 18 is another quirk specific to the JZ4740 USB +device controller, which will be discussed later in `Chapter +5 <#device-quirks>`__. + +The glue layer still needs to register the IRQ handler though. Remember +the instruction on line 14 of the init function: + +:: + + static int jz4740_musb_init(struct musb *musb) + { + musb->isr = jz4740_musb_interrupt; + + return 0; + } + + +This instruction sets a pointer to the glue layer IRQ handler function, +in order for the controller hardware to call the handler back when an +IRQ comes from the controller hardware. The interrupt handler is now +implemented and registered. + +Device Platform Data +==================== + +In order to write an MUSB glue layer, you need to have some data +describing the hardware capabilities of your controller hardware, which +is called the platform data. + +Platform data is specific to your hardware, though it may cover a broad +range of devices, and is generally found somewhere in the arch/ +directory, depending on your device architecture. + +For instance, platform data for the JZ4740 SoC is found in +arch/mips/jz4740/platform.c. In the platform.c file each device of the +JZ4740 SoC is described through a set of structures. + +Here is the part of arch/mips/jz4740/platform.c that covers the USB +Device Controller (UDC): + +:: + + /* USB Device Controller */ + struct platform_device jz4740_udc_xceiv_device = { + .name = "usb_phy_gen_xceiv", + .id = 0, + }; + + static struct resource jz4740_udc_resources[] = { + [0] = { + .start = JZ4740_UDC_BASE_ADDR, + .end = JZ4740_UDC_BASE_ADDR + 0x10000 - 1, + .flags = IORESOURCE_MEM, + }, + [1] = { + .start = JZ4740_IRQ_UDC, + .end = JZ4740_IRQ_UDC, + .flags = IORESOURCE_IRQ, + .name = "mc", + }, + }; + + struct platform_device jz4740_udc_device = { + .name = "musb-jz4740", + .id = -1, + .dev = { + .dma_mask = &jz4740_udc_device.dev.coherent_dma_mask, + .coherent_dma_mask = DMA_BIT_MASK(32), + }, + .num_resources = ARRAY_SIZE(jz4740_udc_resources), + .resource = jz4740_udc_resources, + }; + + +The jz4740_udc_xceiv_device platform device structure (line 2) +describes the UDC transceiver with a name and id number. + +At the time of this writing, note that "usb_phy_gen_xceiv" is the +specific name to be used for all transceivers that are either built-in +with reference USB IP or autonomous and doesn't require any PHY +programming. You will need to set CONFIG_NOP_USB_XCEIV=y in the +kernel configuration to make use of the corresponding transceiver +driver. The id field could be set to -1 (equivalent to +PLATFORM_DEVID_NONE), -2 (equivalent to PLATFORM_DEVID_AUTO) or +start with 0 for the first device of this kind if we want a specific id +number. + +The jz4740_udc_resources resource structure (line 7) defines the UDC +registers base addresses. + +The first array (line 9 to 11) defines the UDC registers base memory +addresses: start points to the first register memory address, end points +to the last register memory address and the flags member defines the +type of resource we are dealing with. So IORESOURCE_MEM is used to +define the registers memory addresses. The second array (line 14 to 17) +defines the UDC IRQ registers addresses. Since there is only one IRQ +register available for the JZ4740 UDC, start and end point at the same +address. The IORESOURCE_IRQ flag tells that we are dealing with IRQ +resources, and the name "mc" is in fact hard-coded in the MUSB core in +order for the controller driver to retrieve this IRQ resource by +querying it by its name. + +Finally, the jz4740_udc_device platform device structure (line 21) +describes the UDC itself. + +The "musb-jz4740" name (line 22) defines the MUSB driver that is used +for this device; remember this is in fact the name that we used in the +jz4740_driver platform driver structure in `Chapter +2 <#linux-musb-basics>`__. The id field (line 23) is set to -1 +(equivalent to PLATFORM_DEVID_NONE) since we do not need an id for the +device: the MUSB controller driver was already set to allocate an +automatic id in `Chapter 2 <#linux-musb-basics>`__. In the dev field we +care for DMA related information here. The dma_mask field (line 25) +defines the width of the DMA mask that is going to be used, and +coherent_dma_mask (line 26) has the same purpose but for the +alloc_coherent DMA mappings: in both cases we are using a 32 bits mask. +Then the resource field (line 29) is simply a pointer to the resource +structure defined before, while the num_resources field (line 28) keeps +track of the number of arrays defined in the resource structure (in this +case there were two resource arrays defined before). + +With this quick overview of the UDC platform data at the arch/ level now +done, let's get back to the MUSB glue layer specific platform data in +drivers/usb/musb/jz4740.c: + +:: + + static struct musb_hdrc_config jz4740_musb_config = { + /* Silicon does not implement USB OTG. */ + .multipoint = 0, + /* Max EPs scanned, driver will decide which EP can be used. */ + .num_eps = 4, + /* RAMbits needed to configure EPs from table */ + .ram_bits = 9, + .fifo_cfg = jz4740_musb_fifo_cfg, + .fifo_cfg_size = ARRAY_SIZE(jz4740_musb_fifo_cfg), + }; + + static struct musb_hdrc_platform_data jz4740_musb_platform_data = { + .mode = MUSB_PERIPHERAL, + .config = &jz4740_musb_config, + }; + + +First the glue layer configures some aspects of the controller driver +operation related to the controller hardware specifics. This is done +through the jz4740_musb_config musb_hdrc_config structure. + +Defining the OTG capability of the controller hardware, the multipoint +member (line 3) is set to 0 (equivalent to false) since the JZ4740 UDC +is not OTG compatible. Then num_eps (line 5) defines the number of USB +endpoints of the controller hardware, including endpoint 0: here we have +3 endpoints + endpoint 0. Next is ram_bits (line 7) which is the width +of the RAM address bus for the MUSB controller hardware. This +information is needed when the controller driver cannot automatically +configure endpoints by reading the relevant controller hardware +registers. This issue will be discussed when we get to device quirks in +`Chapter 5 <#device-quirks>`__. Last two fields (line 8 and 9) are also +about device quirks: fifo_cfg points to the USB endpoints configuration +table and fifo_cfg_size keeps track of the size of the number of +entries in that configuration table. More on that later in `Chapter +5 <#device-quirks>`__. + +Then this configuration is embedded inside jz4740_musb_platform_data +musb_hdrc_platform_data structure (line 11): config is a pointer to +the configuration structure itself, and mode tells the controller driver +if the controller hardware may be used as MUSB_HOST only, +MUSB_PERIPHERAL only or MUSB_OTG which is a dual mode. + +Remember that jz4740_musb_platform_data is then used to convey +platform data information as we have seen in the probe function in +`Chapter 2 <#linux-musb-basics>`__ + +Device Quirks +============= + +Completing the platform data specific to your device, you may also need +to write some code in the glue layer to work around some device specific +limitations. These quirks may be due to some hardware bugs, or simply be +the result of an incomplete implementation of the USB On-the-Go +specification. + +The JZ4740 UDC exhibits such quirks, some of which we will discuss here +for the sake of insight even though these might not be found in the +controller hardware you are working on. + +Let's get back to the init function first: + +:: + + static int jz4740_musb_init(struct musb *musb) + { + musb->xceiv = usb_get_phy(USB_PHY_TYPE_USB2); + if (!musb->xceiv) { + pr_err("HS UDC: no transceiver configured\n"); + return -ENODEV; + } + + /* Silicon does not implement ConfigData register. + * Set dyn_fifo to avoid reading EP config from hardware. + */ + musb->dyn_fifo = true; + + musb->isr = jz4740_musb_interrupt; + + return 0; + } + + +Instruction on line 12 helps the MUSB controller driver to work around +the fact that the controller hardware is missing registers that are used +for USB endpoints configuration. + +Without these registers, the controller driver is unable to read the +endpoints configuration from the hardware, so we use line 12 instruction +to bypass reading the configuration from silicon, and rely on a +hard-coded table that describes the endpoints configuration instead: + +:: + + static struct musb_fifo_cfg jz4740_musb_fifo_cfg[] = { + { .hw_ep_num = 1, .style = FIFO_TX, .maxpacket = 512, }, + { .hw_ep_num = 1, .style = FIFO_RX, .maxpacket = 512, }, + { .hw_ep_num = 2, .style = FIFO_TX, .maxpacket = 64, }, + }; + + +Looking at the configuration table above, we see that each endpoints is +described by three fields: hw_ep_num is the endpoint number, style is +its direction (either FIFO_TX for the controller driver to send packets +in the controller hardware, or FIFO_RX to receive packets from +hardware), and maxpacket defines the maximum size of each data packet +that can be transmitted over that endpoint. Reading from the table, the +controller driver knows that endpoint 1 can be used to send and receive +USB data packets of 512 bytes at once (this is in fact a bulk in/out +endpoint), and endpoint 2 can be used to send data packets of 64 bytes +at once (this is in fact an interrupt endpoint). + +Note that there is no information about endpoint 0 here: that one is +implemented by default in every silicon design, with a predefined +configuration according to the USB specification. For more examples of +endpoint configuration tables, see musb_core.c. + +Let's now get back to the interrupt handler function: + +:: + + static irqreturn_t jz4740_musb_interrupt(int irq, void *__hci) + { + unsigned long flags; + irqreturn_t retval = IRQ_NONE; + struct musb *musb = __hci; + + spin_lock_irqsave(&musb->lock, flags); + + musb->int_usb = musb_readb(musb->mregs, MUSB_INTRUSB); + musb->int_tx = musb_readw(musb->mregs, MUSB_INTRTX); + musb->int_rx = musb_readw(musb->mregs, MUSB_INTRRX); + + /* + * The controller is gadget only, the state of the host mode IRQ bits is + * undefined. Mask them to make sure that the musb driver core will + * never see them set + */ + musb->int_usb &= MUSB_INTR_SUSPEND | MUSB_INTR_RESUME | + MUSB_INTR_RESET | MUSB_INTR_SOF; + + if (musb->int_usb || musb->int_tx || musb->int_rx) + retval = musb_interrupt(musb); + + spin_unlock_irqrestore(&musb->lock, flags); + + return retval; + } + + +Instruction on line 18 above is a way for the controller driver to work +around the fact that some interrupt bits used for USB host mode +operation are missing in the MUSB_INTRUSB register, thus left in an +undefined hardware state, since this MUSB controller hardware is used in +peripheral mode only. As a consequence, the glue layer masks these +missing bits out to avoid parasite interrupts by doing a logical AND +operation between the value read from MUSB_INTRUSB and the bits that +are actually implemented in the register. + +These are only a couple of the quirks found in the JZ4740 USB device +controller. Some others were directly addressed in the MUSB core since +the fixes were generic enough to provide a better handling of the issues +for others controller hardware eventually. + +Conclusion +========== + +Writing a Linux MUSB glue layer should be a more accessible task, as +this documentation tries to show the ins and outs of this exercise. + +The JZ4740 USB device controller being fairly simple, I hope its glue +layer serves as a good example for the curious mind. Used with the +current MUSB glue layers, this documentation should provide enough +guidance to get started; should anything gets out of hand, the linux-usb +mailing list archive is another helpful resource to browse through. + +Acknowledgements +================ + +Many thanks to Lars-Peter Clausen and Maarten ter Huurne for answering +my questions while I was writing the JZ4740 glue layer and for helping +me out getting the code in good shape. + +I would also like to thank the Qi-Hardware community at large for its +cheerful guidance and support. + +Resources +========= + +USB Home Page: http://www.usb.org + +linux-usb Mailing List Archives: http://marc.info/?l=linux-usb + +USB On-the-Go Basics: +http://www.maximintegrated.com/app-notes/index.mvp/id/1822 + +Writing USB Device Drivers: +https://www.kernel.org/doc/htmldocs/writing_usb_driver/index.html + +Texas Instruments USB Configuration Wiki Page: +http://processors.wiki.ti.com/index.php/Usbgeneralpage + +Analog Devices Blackfin MUSB Configuration: +http://docs.blackfin.uclinux.org/doku.php?id=linux-kernel:drivers:musb diff --git a/Documentation/driver-api/usb/writing_usb_driver.rst b/Documentation/driver-api/usb/writing_usb_driver.rst new file mode 100644 index 000000000000..c18dbd74152b --- /dev/null +++ b/Documentation/driver-api/usb/writing_usb_driver.rst @@ -0,0 +1,344 @@ +========================== +Writing USB Device Drivers +========================== + +:Author: Greg Kroah-Hartman + +Introduction +============ + +The Linux USB subsystem has grown from supporting only two different +types of devices in the 2.2.7 kernel (mice and keyboards), to over 20 +different types of devices in the 2.4 kernel. Linux currently supports +almost all USB class devices (standard types of devices like keyboards, +mice, modems, printers and speakers) and an ever-growing number of +vendor-specific devices (such as USB to serial converters, digital +cameras, Ethernet devices and MP3 players). For a full list of the +different USB devices currently supported, see Resources. + +The remaining kinds of USB devices that do not have support on Linux are +almost all vendor-specific devices. Each vendor decides to implement a +custom protocol to talk to their device, so a custom driver usually +needs to be created. Some vendors are open with their USB protocols and +help with the creation of Linux drivers, while others do not publish +them, and developers are forced to reverse-engineer. See Resources for +some links to handy reverse-engineering tools. + +Because each different protocol causes a new driver to be created, I +have written a generic USB driver skeleton, modelled after the +pci-skeleton.c file in the kernel source tree upon which many PCI +network drivers have been based. This USB skeleton can be found at +drivers/usb/usb-skeleton.c in the kernel source tree. In this article I +will walk through the basics of the skeleton driver, explaining the +different pieces and what needs to be done to customize it to your +specific device. + +Linux USB Basics +================ + +If you are going to write a Linux USB driver, please become familiar +with the USB protocol specification. It can be found, along with many +other useful documents, at the USB home page (see Resources). An +excellent introduction to the Linux USB subsystem can be found at the +USB Working Devices List (see Resources). It explains how the Linux USB +subsystem is structured and introduces the reader to the concept of USB +urbs (USB Request Blocks), which are essential to USB drivers. + +The first thing a Linux USB driver needs to do is register itself with +the Linux USB subsystem, giving it some information about which devices +the driver supports and which functions to call when a device supported +by the driver is inserted or removed from the system. All of this +information is passed to the USB subsystem in the usb_driver structure. +The skeleton driver declares a usb_driver as: + +:: + + static struct usb_driver skel_driver = { + .name = "skeleton", + .probe = skel_probe, + .disconnect = skel_disconnect, + .fops = &skel_fops, + .minor = USB_SKEL_MINOR_BASE, + .id_table = skel_table, + }; + + +The variable name is a string that describes the driver. It is used in +informational messages printed to the system log. The probe and +disconnect function pointers are called when a device that matches the +information provided in the id_table variable is either seen or +removed. + +The fops and minor variables are optional. Most USB drivers hook into +another kernel subsystem, such as the SCSI, network or TTY subsystem. +These types of drivers register themselves with the other kernel +subsystem, and any user-space interactions are provided through that +interface. But for drivers that do not have a matching kernel subsystem, +such as MP3 players or scanners, a method of interacting with user space +is needed. The USB subsystem provides a way to register a minor device +number and a set of file_operations function pointers that enable this +user-space interaction. The skeleton driver needs this kind of +interface, so it provides a minor starting number and a pointer to its +file_operations functions. + +The USB driver is then registered with a call to usb_register, usually +in the driver's init function, as shown here: + +:: + + static int __init usb_skel_init(void) + { + int result; + + /* register this driver with the USB subsystem */ + result = usb_register(&skel_driver); + if (result < 0) { + err("usb_register failed for the "__FILE__ "driver." + "Error number %d", result); + return -1; + } + + return 0; + } + module_init(usb_skel_init); + + +When the driver is unloaded from the system, it needs to deregister +itself with the USB subsystem. This is done with the usb_deregister +function: + +:: + + static void __exit usb_skel_exit(void) + { + /* deregister this driver with the USB subsystem */ + usb_deregister(&skel_driver); + } + module_exit(usb_skel_exit); + + +To enable the linux-hotplug system to load the driver automatically when +the device is plugged in, you need to create a MODULE_DEVICE_TABLE. +The following code tells the hotplug scripts that this module supports a +single device with a specific vendor and product ID: + +:: + + /* table of devices that work with this driver */ + static struct usb_device_id skel_table [] = { + { USB_DEVICE(USB_SKEL_VENDOR_ID, USB_SKEL_PRODUCT_ID) }, + { } /* Terminating entry */ + }; + MODULE_DEVICE_TABLE (usb, skel_table); + + +There are other macros that can be used in describing a usb_device_id +for drivers that support a whole class of USB drivers. See usb.h for +more information on this. + +Device operation +================ + +When a device is plugged into the USB bus that matches the device ID +pattern that your driver registered with the USB core, the probe +function is called. The usb_device structure, interface number and the +interface ID are passed to the function: + +:: + + static int skel_probe(struct usb_interface *interface, + const struct usb_device_id *id) + + +The driver now needs to verify that this device is actually one that it +can accept. If so, it returns 0. If not, or if any error occurs during +initialization, an errorcode (such as ``-ENOMEM`` or ``-ENODEV``) is +returned from the probe function. + +In the skeleton driver, we determine what end points are marked as +bulk-in and bulk-out. We create buffers to hold the data that will be +sent and received from the device, and a USB urb to write data to the +device is initialized. + +Conversely, when the device is removed from the USB bus, the disconnect +function is called with the device pointer. The driver needs to clean +any private data that has been allocated at this time and to shut down +any pending urbs that are in the USB system. + +Now that the device is plugged into the system and the driver is bound +to the device, any of the functions in the file_operations structure +that were passed to the USB subsystem will be called from a user program +trying to talk to the device. The first function called will be open, as +the program tries to open the device for I/O. We increment our private +usage count and save a pointer to our internal structure in the file +structure. This is done so that future calls to file operations will +enable the driver to determine which device the user is addressing. All +of this is done with the following code: + +:: + + /* increment our usage count for the module */ + ++skel->open_count; + + /* save our object in the file's private structure */ + file->private_data = dev; + + +After the open function is called, the read and write functions are +called to receive and send data to the device. In the skel_write +function, we receive a pointer to some data that the user wants to send +to the device and the size of the data. The function determines how much +data it can send to the device based on the size of the write urb it has +created (this size depends on the size of the bulk out end point that +the device has). Then it copies the data from user space to kernel +space, points the urb to the data and submits the urb to the USB +subsystem. This can be seen in the following code: + +:: + + /* we can only write as much as 1 urb will hold */ + bytes_written = (count > skel->bulk_out_size) ? skel->bulk_out_size : count; + + /* copy the data from user space into our urb */ + copy_from_user(skel->write_urb->transfer_buffer, buffer, bytes_written); + + /* set up our urb */ + usb_fill_bulk_urb(skel->write_urb, + skel->dev, + usb_sndbulkpipe(skel->dev, skel->bulk_out_endpointAddr), + skel->write_urb->transfer_buffer, + bytes_written, + skel_write_bulk_callback, + skel); + + /* send the data out the bulk port */ + result = usb_submit_urb(skel->write_urb); + if (result) { + err("Failed submitting write urb, error %d", result); + } + + +When the write urb is filled up with the proper information using the +usb_fill_bulk_urb function, we point the urb's completion callback to +call our own skel_write_bulk_callback function. This function is +called when the urb is finished by the USB subsystem. The callback +function is called in interrupt context, so caution must be taken not to +do very much processing at that time. Our implementation of +skel_write_bulk_callback merely reports if the urb was completed +successfully or not and then returns. + +The read function works a bit differently from the write function in +that we do not use an urb to transfer data from the device to the +driver. Instead we call the usb_bulk_msg function, which can be used +to send or receive data from a device without having to create urbs and +handle urb completion callback functions. We call the usb_bulk_msg +function, giving it a buffer into which to place any data received from +the device and a timeout value. If the timeout period expires without +receiving any data from the device, the function will fail and return an +error message. This can be shown with the following code: + +:: + + /* do an immediate bulk read to get data from the device */ + retval = usb_bulk_msg (skel->dev, + usb_rcvbulkpipe (skel->dev, + skel->bulk_in_endpointAddr), + skel->bulk_in_buffer, + skel->bulk_in_size, + &count, HZ*10); + /* if the read was successful, copy the data to user space */ + if (!retval) { + if (copy_to_user (buffer, skel->bulk_in_buffer, count)) + retval = -EFAULT; + else + retval = count; + } + + +The usb_bulk_msg function can be very useful for doing single reads or +writes to a device; however, if you need to read or write constantly to +a device, it is recommended to set up your own urbs and submit them to +the USB subsystem. + +When the user program releases the file handle that it has been using to +talk to the device, the release function in the driver is called. In +this function we decrement our private usage count and wait for possible +pending writes: + +:: + + /* decrement our usage count for the device */ + --skel->open_count; + + +One of the more difficult problems that USB drivers must be able to +handle smoothly is the fact that the USB device may be removed from the +system at any point in time, even if a program is currently talking to +it. It needs to be able to shut down any current reads and writes and +notify the user-space programs that the device is no longer there. The +following code (function :c:func:`skel_delete()`) is an example of +how to do this: + +:: + + static inline void skel_delete (struct usb_skel *dev) + { + kfree (dev->bulk_in_buffer); + if (dev->bulk_out_buffer != NULL) + usb_free_coherent (dev->udev, dev->bulk_out_size, + dev->bulk_out_buffer, + dev->write_urb->transfer_dma); + usb_free_urb (dev->write_urb); + kfree (dev); + } + + +If a program currently has an open handle to the device, we reset the +flag ``device_present``. For every read, write, release and other +functions that expect a device to be present, the driver first checks +this flag to see if the device is still present. If not, it releases +that the device has disappeared, and a -ENODEV error is returned to the +user-space program. When the release function is eventually called, it +determines if there is no device and if not, it does the cleanup that +the skel_disconnect function normally does if there are no open files +on the device (see Listing 5). + +Isochronous Data +================ + +This usb-skeleton driver does not have any examples of interrupt or +isochronous data being sent to or from the device. Interrupt data is +sent almost exactly as bulk data is, with a few minor exceptions. +Isochronous data works differently with continuous streams of data being +sent to or from the device. The audio and video camera drivers are very +good examples of drivers that handle isochronous data and will be useful +if you also need to do this. + +Conclusion +========== + +Writing Linux USB device drivers is not a difficult task as the +usb-skeleton driver shows. This driver, combined with the other current +USB drivers, should provide enough examples to help a beginning author +create a working driver in a minimal amount of time. The linux-usb-devel +mailing list archives also contain a lot of helpful information. + +Resources +========= + +The Linux USB Project: +`http://www.linux-usb.org/ <http://www.linux-usb.org>`__ + +Linux Hotplug Project: +`http://linux-hotplug.sourceforge.net/ <http://linux-hotplug.sourceforge.net>`__ + +Linux USB Working Devices List: +`http://www.qbik.ch/usb/devices/ <http://www.qbik.ch/usb/devices>`__ + +linux-usb-devel Mailing List Archives: +http://marc.theaimsgroup.com/?l=linux-usb-devel + +Programming Guide for Linux USB Device Drivers: +http://usb.cs.tum.edu/usbdoc + +USB Home Page: http://www.usb.org |