illumos: manual page: bufmod.4m

BUFMOD(4M) STREAMS Modules BUFMOD(4M)

NAME

bufmod - STREAMS Buffer Module

SYNOPSIS

#include <sys/bufmod.h>

ioctl(fd, I_PUSH, "bufmod");

DESCRIPTION

bufmod is a STREAMS module that buffers incoming messages, reducing the
number of system calls and the associated overhead required to read and
process them. Although bufmod was originally designed to be used in
conjunction with STREAMS-based networking device drivers, the version
described here is general purpose so that it can be used anywhere STREAMS
input buffering is required.

Read-side Behavior
The behavior of bufmod depends on various parameters and flags that can
be set and queried as described below under IOCTLS. bufmod collects
incoming M_DATA messages into chunks, passing each chunk upstream when
the chunk becomes full or the current read timeout expires. It optionally
converts M_PROTO messages to M_DATA and adds them to chunks as well. It
also optionally adds to each message a header containing a timestamp, and
a cumulative count of messages dropped on the stream read side due to
resource exhaustion or flow control. Thedefault settings of bufmod allow
it to drop messages when flow control sets in or resources are exhausted;
disabling headers and explicitly requesting no drops makes bufmod pass
all messages through. Finally, bufmod is capable of truncating upstream
messages to a fixed, programmable length.

When a message arrives, bufmod processes it in several steps. The
following paragraphs discuss each step in turn.

Upon receiving a message from below, if the SB_NO_HEADER flag is not set,
bufmod immediately timestamps it and saves the current time value for
later insertion in the header described below.

Next, if SB_NO_PROTO_CVT is not set, bufmod converts all leading M_PROTO
blocks in the message to M_DATA blocks, altering only the message type
field and leaving the contents alone.

It then truncates the message to the current snapshot length, which is
set with the SBIOCSSNAP ioctl described below.

Afterwards, if SB_NO_HEADER is not set, bufmod prepends a header to the
converted message. This header is defined as follows.

struct sb_hdr {
uint_t sbh_origlen;
uint_t sbh_msglen;
uint_t sbh_totlen;
uint_t sbh_drops;
#if defined(_LP64) || defined(_I32LPx)
struct timeval32 sbh_timestamp;
#else
struct timeval sbh_timestamp;
#endif /* !_LP64 */
};

The sbh_origlen field gives the message's original length before
truncation in bytes. The sbh_msglen field gives the length in bytes of
the message after the truncation has been done. sbh_totlen gives the
distance in bytes from the start of the truncated message in the current
chunk (described below) to the start of the next message in the chunk;
the value reflects any padding necessary to insure correct data alignment
for the host machine and includes the length of the header itself.
sbh_drops reports the cumulative number of input messages that this
instance of bufmod has dropped due to flow control or resource
exhaustion. In the current implementation message dropping due to flow
control can occur only if the SB_NO_DROPS flag is not set. (Note: this
accounts only for events occurring within bufmod, and does not count
messages dropped by downstream or by upstream modules.) The sbh_timestamp
field contains the message arrival time expressed as a struct timeval.

After preparing a message, bufmod attempts to add it to the end of the
current chunk, using the chunk size and timeout values to govern the
addition. The chunk size and timeout values are set and inspected using
the ioctl() calls described below. If adding the new message would make
the current chunk grow larger than the chunk size, bufmod closes off the
current chunk, passing it up to the next module in line, and starts a new
chunk. If adding the message would still make the new chunk overflow, the
module passes it upward in an over-size chunk of its own. Otherwise, the
module concatenates the message to the end of the current chunk.

To ensure that messages do not languish forever in an accumulating chunk,
bufmod maintains a read timeout. Whenever this timeout expires, the
module closes off the current chunk and passes it upward. The module
restarts the timeout period when it receives a read side data message and
a timeout is not currently active. These two rules insure that bufmod
minimizes the number of chunks it produces during periods of intense
message activity and that it periodically disposes of all messages during
slack intervals, but avoids any timeout overhead when there is no
activity.

bufmod handles other message types as follows. Upon receiving an M_FLUSH
message specifying that the read queue be flushed, the module clears the
currently accumulating chunk and passes the message on to the module or
driver above. (Note: bufmod uses zero length M_CTL messages for internal
synchronization and does not pass them through.) bufmod passes all other
messages through unaltered to its upper neighbor, maintaining message
order for non high priority messages by passing up any accumulated chunk
first.

If the SB_DEFER_CHUNK flag is set, buffering does not begin until the
second message is received within the timeout window.

If the SB_SEND_ON_WRITE flag is set, bufmod passes up the read side any
buffered data when a message is received on the write side.
SB_SEND_ON_WRITE and SB_DEFER_CHUNK are often used together.

Write-side Behavior
bufmod intercepts M_IOCTL messages for the ioctls described below. The
module passes all other messages through unaltered to its lower neighbor.
If SB_SEND_ON_WRITE is set, message arrival on the writer side suffices
to close and transmit the current read side chunk.

IOCTLS

bufmod responds to the following ioctls.

SBIOCSTIME
Set the read timeout value to the value referred to by
the struct timeval pointer given as argument. Setting the
timeout value to zero has the side-effect of forcing the
chunk size to zero as well, so that the module will pass
all incoming messages upward immediately upon arrival.
Negative values are rejected with an EINVAL error.

SBIOCGTIME
Return the read timeout in the struct timeval pointed to
by the argument. If the timeout has been cleared with
the SBIOCCTIME ioctl, return with an ERANGE error.

SBIOCCTIME
Clear the read timeout, effectively setting its value to
infinity. This results in no timeouts being active and
the chunk being delivered when it is full.

SBIOCSCHUNK
Set the chunk size to the value referred to by the uint_t
pointer given as argument. See Notes for a description of
effect on stream head high water mark.

SBIOCGCHUNK
Return the chunk size in the uint_t pointed to by the
argument.

SBIOCSSNAP
Set the current snapshot length to the value given in the
uint_t pointed to by the ioctl's final argument. bufmod
interprets a snapshot length value of zero as meaning
infinity, so it will not alter the message. See Notes for
a description of effect on stream head high water mark.

SBIOCGSNAP
Returns the current snapshot length in the uint_t pointed
to by the ioctl's final argument.

SBIOCSFLAGS
Set the current flags to the value given in the uint_t
pointed to by the ioctl's final argument. Possible values
are a combination of the following.

SB_SEND_ON_WRITE
Transmit the read side chunk on
arrival of a message on the write
side.

SB_NO_HEADER
Do not add headers to read side
messages.

SB_NO_DROPS
Do not drop messages due to flow
control upstream.

SB_NO_PROTO_CVT
Do not convert M_PROTO messages into
M_DATA.

SB_DEFER_CHUNK
Begin buffering on arrival of the
second read side message in a
timeout interval.

SBIOCGFLAGS
Returns the current flags in the uint_t pointed to by the
ioctl's final argument.

NOTES

Older versions of bufmod did not support the behavioral flexibility
controlled by the SBIOCSFLAGS ioctl. Applications that wish to take
advantage of this flexibility can guard themselves against old versions
of the module by invoking the SBIOCGFLAGS ioctl and checking for an
EINVAL error return.

When buffering is enabled by issuing an SBIOCSCHUNK ioctl to set the
chunk size to a non zero value, bufmod sends a SETOPTS message to adjust
the stream head high and low water marks to accommodate the chunked
messages.

When buffering is disabled by setting the chunk size to zero, message
truncation can have a significant influence on data traffic at the
stream head and therefore the stream head high and low water marks are
adjusted to new values appropriate for the smaller truncated message
sizes.

BUGS