MIDI(4) Kernel Interfaces Manual MIDI(4)
midi -- device-independent MIDI driver layer
midi* at midibus?
midi* at pcppi?
The midi driver is the machine independent layer over anything that can
source or sink a MIDI data stream, whether a physical MIDI IN or MIDI OUT
jack on a soundcard, cabled to some external synthesizer or input
controller, an on-board programmable tone generator, or a single jack,
synthesizer, or controller component within a complex USB or IEEE1394
MIDI device that has several such components and appears as several MIDI
One MIDI data stream is a unidirectional stream of MIDI messages, as
could be carried over one MIDI cable in the MIDI 1.0 specification. Many
MIDI messages carry a four-bit channel number, creating up to 16 MIDI
channels within a single MIDI stream. There may be multiple consumers of
a MIDI stream, each configured to react only to messages on specific
channels; the sets of channels different consumers react to need not be
disjoint. Many modern devices such as multitimbral keyboards and tone
generators listen on all 16 channels, or may be viewed as collections of
16 independent consumers each listening on one channel. MIDI defines
some messages that take no channel number, and apply to all consumers of
the stream on which they are sent. For an inbound stream, midi is a
promiscuous receiver, capturing all messages regardless of channel
number. For an outbound stream, the writer can specify a channel number
per message; there is no notion of binding the stream to one destination
channel in advance.
A single midi device instance is the endpoint of one outbound stream, one
inbound stream, or one of each. In the third case, the write and read
sides are independent MIDI streams. For example, a soundcard driver may
map its MIDI OUT and MIDI IN jacks to the write and read sides of a
single device instance, but those jacks can be cabled to completely
different pieces of gear. Information from dmesg(8), and a diagram of
any external MIDI cabling, will help clarify the mapping.
Underlying drivers and MIDI protocol
Drivers midi can attach include soundcard drivers, many of which support
a UART resembling Roland's MPU401 and handled by mpu(4), USB MIDI devices
via umidi(4), and on-board devices that can make sounds, whether a lowly
PC speaker or a Yamaha OPL. Serial port and IEEE1394 connections are
currently science fiction.
The MIDI protocol permits some forms of message compression such as
running status and hidden note-off. Received messages on inbound streams
are always canonicalized by midi before presentation to higher layers.
Messages for transmission are accepted by midi in any valid form.
Access to midi device instances can be through the raw device nodes,
/dev/rmidiN, or through the sequencer, /dev/music.
Raw MIDI access
A /dev/rmidiN device supports read(2), write(2), ioctl(2),
select(2)/poll(2) and the corresponding kevent(2) filters, and may be
opened only when it is not already open. It may be opened in O_RDONLY,
O_WRONLY, or O_RDWR mode, but a later read(2) or write(2) will return -1
if the device has no associated input or output stream, respectively.
Bytes written are passed as quickly as possible to the underlying driver
as complete MIDI messages; a maximum of two bytes at the end of a
write(2) may remain buffered if they do not complete a message, until
completed by a following write(2).
A read(2) will not block or return EWOULDBLOCK when it could immediately
return any nonzero count, and MIDI messages received are available to
read(2) as soon as they are complete, with a maximum of two received
bytes remaining buffered if they do not complete a message.
As all MIDI messages are three bytes or fewer except for System
Exclusive, which can have arbitrary length, these rules imply that System
Exclusive messages are the only ones of which some bytes can be delivered
before all are available.
System Realtime messages are passed with minimum delay in either
direction, ahead of any possible buffered incomplete message. As a
result, they will never interrupt any MIDI message except possibly System
A read(2) with a buffer large enough to accommodate the first complete
message available will be satisfied with as many complete messages as
will fit. A buffer too small for the first complete message will be
filled to capacity. Therefore, an application that reads from an rmidi
device with buffers of three bytes or larger need never parse across read
boundaries to assemble a received message, except possibly in the case of
a System Exclusive message. However, if the application reads through a
buffering layer such as fread(3), this property will not be preserved.
The midi driver itself supports the ioctl(2) operations FIOASYNC,
FIONBIO, and FIONREAD. Underlying devices may support others. The value
returned for FIONREAD reflects the size in bytes of complete messages (or
System Exclusive chunks) ready to read. If the ioctl(2) returns n and a
read(2) of size n is issued, n bytes will be read, but if a read(2) of
size m < n is issued, fewer than m bytes may be read if m does not fall
on a message/chunk boundary.
Raw MIDI access can be used to receive bulk dumps from synthesizers,
download bulk data to them, and so on. Simple patching of one device to
another can be done at the command line, as with
$ cat -u 0<>/dev/rmidi0 1>&0
which will loop all messages received on the input stream of rmidi0 input
stream back to its output stream in real time. However, an attempt to
record and play back music with
$ cat /dev/rmidiN >foo; cat foo >/dev/rmidiN
will be disappointing. The file foo will contain all of the notes that
were played, but because MIDI messages carry no explicit timing, the
`playback' will reproduce them all at once, as fast as they can be
transmitted. To preserve timing information, the sequencer device can be
The MIDI protocol includes a keepalive function called Active Sensing.
In any receiver that has not received at least one Active Sense MIDI
message, the feature is suppressed and no timeout applies. If at least
one such message has been received, the lapse of any subsequent 300 ms
interval without receipt of any message reflects loss of communication,
and the receiver should silence any currently sounding notes and return
to non-Active-Sensing behavior. A sender using Active Sensing generally
avoids 300 ms gaps in transmission by sending Active Sense messages
(which have no other effect) as needed when there is no other traffic to
send in the interval. This feature can be important for MIDI, which
relies on separate Note On and Note Off messages, to avoid notes stuck on
indefinitely if communication is interrupted before a Note Off message
This protocol is supported in midi. An outbound stream will be kept
alive by sending Active Sense messages as needed, beginning after any
real traffic is sent on the stream, and continuing until the stream is
closed. On an inbound stream, if any Active Sense has been received,
then a process reading an rmidi device will see an end-of-file indication
if the input timeout elapses. The stream remains open, the driver
reverts to enforcing no timeout, and the process may continue to read for
more input. Subsequent receipt of an Active Sense message will re-arm
the timeout. As received Active Sense messages are handled by midi, they
are not included among messages read from the /dev/rmidiN device.
These rules support end-to-end Active Sensing behavior in simple cases
without special action in an application. For example, in
$ cat -u /dev/rmidi0 >/dev/rmidi1
if the input stream to rmidi0 is lost, the cat(1) command exits; on the
close(2) of rmidi1, midi ceases to send Active Sense messages, and the
receiving device will detect the loss and silence any outstanding notes.
Access through the sequencer
To play music using the raw MIDI API would require an application to
issue many small writes with very precise timing. The sequencer device,
/dev/music, can manage the timing of MIDI data in the kernel, to avoid
such demanding real-time constraints on a user process.
The /dev/music device can be opened only when it is not already open.
When opened, the sequencer internally opens all MIDI instances existing
in the system that are not already open at their raw nodes; any attempts
to open them at their raw nodes while the sequencer is open will fail.
All access to the corresponding MIDI streams will then be through the
Reads and writes of /dev/music pass eight-byte event structures defined
in <sys/midiio.h> (which see for their documentation and examples of
use). Some events correspond to MIDI messages, and carry an integer
device field to identify one of the MIDI devices opened by the sequencer.
Other events carry timing information interpreted or generated by the
A message received on an input stream is wrapped in a sequencer event
along with the device index of the stream it arrived on, and queued for
the reader of /dev/music. If a measurable time interval passed since the
last preceding message, a timing event that represents a delay for that
interval is queued ahead of the received event. The sequencer handles
output events by interpreting any timing event, and routing any MIDI
message event at the proper time to an underlying output stream according
to its device index. Therefore
$ cat /dev/music >foo; cat foo >/dev/music
can be expected to capture and reproduce an input performance including
The process of playing back a complex MIDI file is illustrated below.
The file may contain several tracks--four, in this example--of MIDI
events, each marked with a device index and a time stamp, that may
overlap in time. In the example, a, b, and c are device indices of the
three output MIDI streams; the left-hand digit in each input event
represents a MIDI channel on the selected stream, and the right-hand
digit represents a time for the event's occurrence. As illustrated, the
input tracks are not firmly associated with output streams; any track may
contain events for any stream.
| | a2|4 |
a0|3 | c1|3 c0|3
| b0|2 b1|2 |
| b1|1 | c0|1
a0|0 | b0|0 |
v v v v
| merge to 1 ordered stream |
| user code, eg midiplay(1) |
| /dev/music | kernel
| (sequencer) |
| 1 0
+-----' | '-----.
0 0 |
v v v
+-------+ +--------+ +---------+
|midi(4)| |midi(4) | |midi(4) |
|rmidia | |rmidib | |rmidic |
+-------+ +--------+ +---------+
| mpu(4)| |umidi(4)| |midisyn |
+-------+ +--------+ +---------+
| HW | | | opl(4) |
| MIDI | U +---------+
| UART | S | internal|
+-------+ B | tone |
| | |generator|
v | +---------+
MIDI device external
A user process must merge the tracks into a single stream of sequencer
MIDI and timing events in order by desired timing. The sequencer obeys
the timing events and distributes the MIDI events to the three
destinations, in this case two external devices connected to a sound card
UART and a USB interface, and an OPL tone generator on a sound card.
Use of select(2)/poll(2) with the sequencer is supported, however, there
is no guarantee that a write(2) will not block or return EWOULDBLOCK if
it begins with a timer-wait event, even if select(2)/poll(2) reported the
The delivery of a realtime message ahead of buffered bytes of an
incomplete message may cause the realtime message to seem, in a saved
byte stream, to have arrived up to 640 us earlier than it really did, at
MIDI 1.0 data rates. Higher data rates make the effect less significant.
Another sequencer device, /dev/sequencer, is provided only for backward
compatibility with an obsolete OSS interface in which some sequencer
events were four-byte records. It is not further documented here, and
the /dev/music API should be used in new code. The /dev/sequencer
emulation is implemented only for writing, and that might not be
Some hardware devices supporting midi lack transmit-ready interrupts, and
some have the capability in hardware but currently lack driver support.
They can be recognized by the annotation (CPU-intensive output) in
dmesg(8). While suitable for music playback, they may have an
objectionable impact on system responsiveness during bulk transmission
such as patch downloads, and are best avoided for that purpose if other
suitable devices are present.
Buffer space in midi itself is adequate for about 200 ms of traffic at
MIDI 1.0 data rates, per stream.
Event counters record bytes and messages discarded because of protocol
errors or buffer overruns, and can be viewed with vmstat -e. They can be
useful in diagnosing flaky cables and other communication problems.
A raw sound generator uses the midisyn layer to present a MIDI message-
driven interface attachable by midi.
While midi accepts messages for transmission in any valid mixture of
compressed or canonical form, they are always presented to an underlying
driver in the form it prefers. Drivers for simple UART-like devices
register their preference for a compressed byte stream, while those like
umidi(4), which uses a packet protocol, or midisyn, which interprets
complete messages, register for intact canonical messages. This design
eliminates the need for compression and canonicalization logic from all
layers above and below midi itself.
In addition to other errors documented for the write(2) family of system
calls, EPROTO can be returned if the bytes to be written on a raw midi
device would violate MIDI protocol.
midiplay(1), ioctl(2), ossaudio(3), audio(4), mpu(4), opl(4), umidi(4)
For ports using the ISA bus: cms(4), pcppi(4), sb(4)
For ports using the PCI bus: autri(4), clcs(4), eap(4)
The midi driver first appeared in NetBSD 1.4. It was overhauled and this
manual page rewritten for NetBSD 4.0.
Some OSS sequencer events and ioctl(2) operations are unimplemented, as
OSS source-compatible sequencer macros should be added to
<sys/soundcard.h>, implemented with the NetBSD ones in <sys/midiio.h>, so
sources written for OSS can be easily compiled.
The sequencer blocks (or returns EWOULDBLOCK) only when its buffer
physically fills, which can represent an arbitrary latency because of
buffered timing events. As a result, interrupting a process writing the
sequencer may not interrupt music playback for a considerable time. The
sequencer could enforce a reasonable latency bound by examining timing
events as they are enqueued and blocking appropriately.
FIOASYNC enables signal delivery to the calling process only; FIOSETOWN
is not supported.
The sequencer can only be a timing master, but does not send timing
messages to synchronize any slave device; it cannot be slaved to timing
messages received on any interface (which would presumably require a PLL
algorithm similar to NTP's, and expertise in that area to implement it).
The sequencer ignores timing messages received on any interface and does
not pass them along to the reading process, and the OSS operations to
change that behavior are unimplemented.
The SEQUENCER_TMR_TIMEBASE ioctl(2) will report successfully setting any
timebase up to ridiculously high resolutions, though the actual
resolution, and therefore jitter, is constrained by hz(9). Comparable
sequencer implementations typically allow a selection from available
sources of time interrupts that may be programmable.
The device number in a sequencer event is treated on write(2) as index
into the array of MIDI devices the sequencer has opened, but on read(2)
as the unit number of the source MIDI device; these are usually the same
if the sequencer has opened all the MIDI devices (that is, none was
already open at its raw node when the sequencer was opened), but might
not be the same otherwise.
There is at present no way to make reception nonpromiscuous, should
anyone have a reason to want to.
There should be ways to override default Active Sense behavior. As one
obvious case, if an application is seen to send Active Sense explicitly,
midi should refrain from adding its own. On receive, there should be an
option to pass Active Sense through rather than interpreting it, for apps
that wish to handle or ignore it themselves and never see EOF.
When a midi stream is open by the sequencer, Active Sense messages
received on the stream are passed to the sequencer and not interpreted by
midi. The sequencer at present neither does anything itself with Active
Sense messages received, nor supports the OSS API for making them
available to the user process.
System Exclusive messages can be received by reading a raw device, but
not by reading the sequencer; they are discarded on receipt when the
stream is open by the sequencer, rather than being presented as the OSS-
defined sequencer events.
midisyn is too rudimentary at present to get satisfactory results from
any onboard synth. It lacks the required special interpretation of the
General MIDI percussion channel in GM mode. More devices should be
supported; some sound cards with synthesis capability have NetBSD drivers
that implement the audio(4) but not the midisyn interface. Voice
stealing algorithm does not follow the General MIDI Developer Guidelines.
ALSA sequencer compatibility is lacking, but becoming important to
applications. It would require the function of merging multiple tracks
into a single ordered stream to be moved from user space into the
sequencer. Assuming the sequencer driven by periodic interrupts, timing
wheels could be used as in hardclock(9) itself. Similar functionality
will be in OSS4; with the right infrastructure it should be possible to
support both. When merging MIDI streams, a notion of transaction is
needed to group critical message sequences. If ALSA or OSS4 have no such
notion, it should be provided as an upward-compatible extension.
I would rather have open(2) itself return an error (by the POSIX
description ENODEV looks most appropriate) if a read or write mode is
requested that is not supported by the instance, rather than letting
open(2) succeed and read(2) or write(2) return -1, but so help me, the
latter seems the more common UNIX practice.
NetBSD 6.1.5 May 6, 2006 NetBSD 6.1.5