USB device drivers provide two ways of transferring data between host
and peripheral. The first involves USB-specific functionality such as
usbs_start_rx_buffer.
This provides non-blocking I/O: a transfer is started, and some time
later the device driver will call a supplied completion function. The
second uses the conventional I/O model: there are entries in the
device table corresponding to the various endpoints. Standard calls
such as open can then be used to get a suitable
handle. Actual I/O happens via blocking read and
write calls. In practice the blocking operations
are simply implemented using the underlying non-blocking
functionality.
Each endpoint will have its own devtab entry. The exact names are
controlled by the device driver package, but typically the root will
be /dev/usb. This is followed by one or more
decimal digits giving the endpoint number, followed by
c for a control endpoint, r for
a receive endpoint (host to peripheral), and w for
a transmit endpoint (peripheral to host). If the target hardware
involves more than one USB device then different roots should be used,
for example /dev/usb0c and
/dev/usb1_0c. This may require explicit
manipulation of device driver configuration options by the application
developer.
At present the devtab entry for a control endpoint does not support
any I/O operations.
write operations
cyg_io_write and similar functions in
higher-level packages can be used to perform a transfer from
peripheral to host. Successive write operations will not be coalesced.
For example, when doing a 1000 byte write to an endpoint that uses the
bulk transfer protocol this will involve 15 full-size 64-byte packets
and a terminating 40-byte packet. USB device drivers are not expected
to do any locking, and if higher-level code performs multiple
concurrent write operations on a single endpoint then the resulting
behaviour is undefined.
A USB write operation will never transfer less
data than specified. It is the responsibility of higher-level code to
ensure that the amount of data being transferred is acceptable to the
host-side code. Usually this will be defined by a higher-level
protocol. If an attempt is made to transfer more data than the host
expects then the resulting behaviour is undefined.
There are two likely error conditions. EPIPE
indicates that the connection between host and target has been broken.
EAGAIN indicates that the endpoint has been
stalled, either at the request of the host or by other activity
inside the peripheral.
read operations
cyg_io_read and similar functions in higher-level
packages can be used to perform a transfer from host to peripheral.
This should be a complete transfer: higher-level protocols should
define an upper bound on the amount of data being transferred, and the
read operation should involve at least this
amount of data. The return value will indicate the actual transfer
size, which may be less than requested.
Some device drivers may support partial reads, but USB device drivers
are not expected to perform any buffering because that involves both
memory and code overheads. One technique that may work for bulk
transfers is to exploit the fact that such transfers happen in 64-byte
packets. It is possible to read an initial 64
bytes, corresponding to the first packet in the transfer. These 64
bytes can then be examined to determine the total transfer size, and
the remaining data can be transferred in another
read operation. This technique is not guaranteed
to work with all USB hardware. Also, if the delay between accepting
the first packet and the remainder of the transfer is excessive then
this could cause timeout problems for the host-side software. For
these reasons the use of partial reads should be avoided.
There are two likely error conditions. EPIPE
indicates that the connection between host and target has been broken.
EAGAIN indicates that the endpoint has been
stalled, either at the request of the host or by other activity
inside the peripheral.
USB device drivers are not expected to do any locking. If higher-level
code performs multiple concurrent read operations on a single endpoint
then the resulting behaviour is undefined.
select operations
Typical USB device drivers will not provide any support for
select. Consider bulk transfers from the host to
the peripheral. At the USB device driver level there is no way of
knowing in advance how large a transfer will be, so it is not feasible
for the device driver to buffer the entire transfer. It may be
possible to buffer part of the transfer, for example the first 64-byte
packet, and copy this into application space at the start of a
read, but this adds code and memory overheads.
Worse, it means that there is an unknown but potentially long delay
between a peripheral accepting the first packet of a transfer and the
remaining packets, which could confuse or upset the host-side
software.
With some USB hardware it may be possible for the device driver to
detect OUT tokens from the host without actually accepting the data,
and this would indicate that a read is likely to
succeed. However, it would not be reliable since the host-side I/O
operation could time out. A similar mechanism could be used to
implement select for outgoing data, but again
this would not be reliable.
Some device drivers may provide partial support for
select anyway, possibly under the control of a
configuration option. The device driver's documentation should be
consulted for further information. It is also worth noting that the
USB-specific non-blocking API can often be used as an alternative to
select.
get_config and
set_config operations
There are no set_config or
get_config (also known as
ioctl) operations defined for USB devices.
Some device drivers may provide hardware-specific facilities this way.
Note: Currently the USB-specific functions related to halted endpoints cannot be accessed readily
via devtab entries. This functionality should probably be made
available via set_config and
get_config. It may also prove useful to provide
a get_config operation that maps from the
devtab entries to the underlying endpoint data structures.
Presence
The devtab entries are optional. If the USB device is accessed
primarily by class-specific code such as the USB-ethernet package and
that package uses the USB-specific API directly, the devtab entries
are redundant. Even if application code does need to access the USB
device, the non-blocking API may be more convenient than the blocking
I/O provided via the devtab entries. In these cases the devtab entries
serve no useful purpose, but they still impose a memory overhead. It
is possible to suppress the presence of these entries by disabling the
configuration option
CYGGLO_IO_USB_SLAVE_PROVIDE_DEVTAB_ENTRIES.
