freebsd-dev/share/doc/psd/03.iosys/iosys

.\" Copyright (C) Caldera International Inc. 2001-2002.  All rights reserved.
.\" 
.\" Redistribution and use in source and binary forms, with or without
.\" modification, are permitted provided that the following conditions are
.\" met:
.\" 
.\" Redistributions of source code and documentation must retain the above
.\" copyright notice, this list of conditions and the following
.\" disclaimer.
.\" 
.\" Redistributions in binary form must reproduce the above copyright
.\" notice, this list of conditions and the following disclaimer in the
.\" documentation and/or other materials provided with the distribution.
.\" 
.\" All advertising materials mentioning features or use of this software
.\" must display the following acknowledgement:
.\" 
.\" This product includes software developed or owned by Caldera
.\" International, Inc.  Neither the name of Caldera International, Inc.
.\" nor the names of other contributors may be used to endorse or promote
.\" products derived from this software without specific prior written
.\" permission.
.\" 
.\" USE OF THE SOFTWARE PROVIDED FOR UNDER THIS LICENSE BY CALDERA
.\" INTERNATIONAL, INC.  AND CONTRIBUTORS ``AS IS'' AND ANY EXPRESS OR
.\" IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
.\" WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
.\" DISCLAIMED.  IN NO EVENT SHALL CALDERA INTERNATIONAL, INC. BE LIABLE
.\" FOR ANY DIRECT, INDIRECT INCIDENTAL, SPECIAL, EXEMPLARY, OR
.\" CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
.\" SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
.\" BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
.\" WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
.\" OR OTHERWISE) RISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
.\" IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.\" 
.\"	@(#)iosys	8.1 (Berkeley) 6/8/93
.\"
.\" $FreeBSD$
.EH 'PSD:3-%''The UNIX I/O System'
.OH 'The UNIX I/O System''PSD:3-%'
.TL
The UNIX I/O System
.AU
Dennis M. Ritchie
.AI
AT&T Bell Laboratories
Murray Hill, NJ
.PP
This paper gives an overview of the workings of the UNIX\(dg
.FS
\(dgUNIX is a Trademark of Bell Laboratories.
.FE
I/O system.
It was written with an eye toward providing
guidance to writers of device driver routines,
and is oriented more toward describing the environment
and nature of device drivers than the implementation
of that part of the file system which deals with
ordinary files.
.PP
It is assumed that the reader has a good knowledge
of the overall structure of the file system as discussed
in the paper ``The UNIX Time-sharing System.''
A more detailed discussion
appears in
``UNIX Implementation;''
the current document restates parts of that one,
but is still more detailed.
It is most useful in
conjunction with a copy of the system code,
since it is basically an exegesis of that code.
.SH
Device Classes
.PP
There are two classes of device:
.I block
and
.I character.
The block interface is suitable for devices
like disks, tapes, and DECtape
which work, or can work, with addressible 512-byte blocks.
Ordinary magnetic tape just barely fits in this category,
since by use of forward
and
backward spacing any block can be read, even though
blocks can be written only at the end of the tape.
Block devices can at least potentially contain a mounted
file system.
The interface to block devices is very highly structured;
the drivers for these devices share a great many routines
as well as a pool of buffers.
.PP
Character-type devices have a much
more straightforward interface, although
more work must be done by the driver itself.
.PP
Devices of both types are named by a
.I major
and a
.I minor
device number.
These numbers are generally stored as an integer
with the minor device number
in the low-order 8 bits and the major device number
in the next-higher 8 bits;
macros
.I major
and
.I minor
are available to access these numbers.
The major device number selects which driver will deal with
the device; the minor device number is not used
by the rest of the system but is passed to the
driver at appropriate times.
Typically the minor number
selects a subdevice attached to
a given controller, or one of
several similar hardware interfaces.
.PP
The major device numbers for block and character devices
are used as indices in separate tables;
they both start at 0 and therefore overlap.
.SH
Overview of I/O
.PP
The purpose of
the
.I open
and
.I creat
system calls is to set up entries in three separate
system tables.
The first of these is the
.I u_ofile
table,
which is stored in the system's per-process
data area
.I u.
This table is indexed by
the file descriptor returned by the
.I open
or
.I creat,
and is accessed during
a
.I read,
.I write,
or other operation on the open file.
An entry contains only
a pointer to the corresponding
entry of the
.I file
table,
which is a per-system data base.
There is one entry in the
.I file
table for each
instance of
.I open
or
.I creat.
This table is per-system because the same instance
of an open file must be shared among the several processes
which can result from
.I forks
after the file is opened.
A
.I file
table entry contains
flags which indicate whether the file
was open for reading or writing or is a pipe, and
a count which is used to decide when all processes
using the entry have terminated or closed the file
(so the entry can be abandoned).
There is also a 32-bit file offset
which is used to indicate where in the file the next read
or write will take place.
Finally, there is a pointer to the
entry for the file in the
.I inode
table,
which contains a copy of the file's i-node.
.PP
Certain open files can be designated ``multiplexed''
files, and several other flags apply to such
channels.
In such a case, instead of an offset,
there is a pointer to an associated multiplex channel table.
Multiplex channels will not be discussed here.
.PP
An entry in the
.I file
table corresponds precisely to an instance of
.I open
or
.I creat;
if the same file is opened several times,
it will have several
entries in this table.
However,
there is at most one entry
in the
.I inode
table for a given file.
Also, a file may enter the
.I inode
table not only because it is open,
but also because it is the current directory
of some process or because it
is a special file containing a currently-mounted
file system.
.PP
An entry in the
.I inode
table differs somewhat from the
corresponding i-node as stored on the disk;
the modified and accessed times are not stored,
and the entry is augmented
by a flag word containing information about the entry,
a count used to determine when it may be
allowed to disappear,
and the device and i-number
whence the entry came.
Also, the several block numbers that give addressing
information for the file are expanded from
the 3-byte, compressed format used on the disk to full
.I long
quantities.
.PP
During the processing of an
.I open
or
.I creat
call for a special file,
the system always calls the device's
.I open
routine to allow for any special processing
required (rewinding a tape, turning on
the data-terminal-ready lead of a modem, etc.).
However,
the
.I close
routine is called only when the last
process closes a file,
that is, when the i-node table entry
is being deallocated.
Thus it is not feasible
for a device to maintain, or depend on,
a count of its users, although it is quite
possible to
implement an exclusive-use device which cannot
be reopened until it has been closed.
.PP
When a
.I read
or
.I write
takes place,
the user's arguments
and the
.I file
table entry are used to set up the
variables
.I u.u_base,
.I u.u_count,
and
.I u.u_offset
which respectively contain the (user) address
of the I/O target area, the byte-count for the transfer,
and the current location in the file.
If the file referred to is
a character-type special file, the appropriate read
or write routine is called; it is responsible
for transferring data and updating the
count and current location appropriately
as discussed below.
Otherwise, the current location is used to calculate
a logical block number in the file.
If the file is an ordinary file the logical block
number must be mapped (possibly using indirect blocks)
to a physical block number; a block-type
special file need not be mapped.
This mapping is performed by the
.I bmap
routine.
In any event, the resulting physical block number
is used, as discussed below, to
read or write the appropriate device.
.SH
Character Device Drivers
.PP
The
.I cdevsw
table specifies the interface routines present for
character devices.
Each device provides five routines:
open, close, read, write, and special-function
(to implement the
.I ioctl
system call).
Any of these may be missing.
If a call on the routine
should be ignored,
(e.g.
.I open
on non-exclusive devices that require no setup)
the
.I cdevsw
entry can be given as
.I nulldev;
if it should be considered an error,
(e.g.
.I write
on read-only devices)
.I nodev
is used.
For terminals,
the
.I cdevsw
structure also contains a pointer to the
.I tty
structure associated with the terminal.
.PP
The
.I open
routine is called each time the file
is opened with the full device number as argument.
The second argument is a flag which is
non-zero only if the device is to be written upon.
.PP
The
.I close
routine is called only when the file
is closed for the last time,
that is when the very last process in
which the file is open closes it.
This means it is not possible for the driver to
maintain its own count of its users.
The first argument is the device number;
the second is a flag which is non-zero
if the file was open for writing in the process which
performs the final
.I close.
.PP
When
.I write
is called, it is supplied the device
as argument.
The per-user variable
.I u.u_count
has been set to
the number of characters indicated by the user;
for character devices, this number may be 0
initially.
.I u.u_base
is the address supplied by the user from which to start
taking characters.
The system may call the
routine internally, so the
flag
.I u.u_segflg
is supplied that indicates,
if
.I on,
that
.I u.u_base
refers to the system address space instead of
the user's.
.PP
The
.I write
routine
should copy up to
.I u.u_count
characters from the user's buffer to the device,
decrementing
.I u.u_count
for each character passed.
For most drivers, which work one character at a time,
the routine
.I "cpass( )"
is used to pick up characters
from the user's buffer.
Successive calls on it return
the characters to be written until
.I u.u_count
goes to 0 or an error occurs,
when it returns \(mi1.
.I Cpass
takes care of interrogating
.I u.u_segflg
and updating
.I u.u_count.
.PP
Write routines which want to transfer
a probably large number of characters into an internal
buffer may also use the routine
.I "iomove(buffer, offset, count, flag)"
which is faster when many characters must be moved.
.I Iomove
transfers up to
.I count
characters into the
.I buffer
starting
.I offset
bytes from the start of the buffer;
.I flag
should be
.I B_WRITE
(which is 0) in the write case.
Caution:
the caller is responsible for making sure
the count is not too large and is non-zero.
As an efficiency note,
.I iomove
is much slower if any of
.I "buffer+offset, count"
or
.I u.u_base
is odd.
.PP
The device's
.I read
routine is called under conditions similar to
.I write,
except that
.I u.u_count
is guaranteed to be non-zero.
To return characters to the user, the routine
.I "passc(c)"
is available; it takes care of housekeeping
like
.I cpass
and returns \(mi1 as the last character
specified by
.I u.u_count
is returned to the user;
before that time, 0 is returned.
.I Iomove
is also usable as with
.I write;
the flag should be
.I B_READ
but the same cautions apply.
.PP
The ``special-functions'' routine
is invoked by the
.I stty
and
.I gtty
system calls as follows:
.I "(*p) (dev, v)"
where
.I p
is a pointer to the device's routine,
.I dev
is the device number,
and
.I v
is a vector.
In the
.I gtty
case,
the device is supposed to place up to 3 words of status information
into the vector; this will be returned to the caller.
In the
.I stty
case,
.I v
is 0;
the device should take up to 3 words of
control information from
the array
.I "u.u_arg[0...2]."
.PP
Finally, each device should have appropriate interrupt-time
routines.
When an interrupt occurs, it is turned into a C-compatible call
on the devices's interrupt routine.
The interrupt-catching mechanism makes
the low-order four bits of the ``new PS'' word in the
trap vector for the interrupt available
to the interrupt handler.
This is conventionally used by drivers
which deal with multiple similar devices
to encode the minor device number.
After the interrupt has been processed,
a return from the interrupt handler will
return from the interrupt itself.
.PP
A number of subroutines are available which are useful
to character device drivers.
Most of these handlers, for example, need a place
to buffer characters in the internal interface
between their ``top half'' (read/write)
and ``bottom half'' (interrupt) routines.
For relatively low data-rate devices, the best mechanism
is the character queue maintained by the
routines
.I getc
and
.I putc.
A queue header has the structure
.DS
struct {
	int	c_cc;	/* character count */
	char	*c_cf;	/* first character */
	char	*c_cl;	/* last character */
} queue;
.DE
A character is placed on the end of a queue by
.I "putc(c, &queue)"
where
.I c
is the character and
.I queue
is the queue header.
The routine returns \(mi1 if there is no space
to put the character, 0 otherwise.
The first character on the queue may be retrieved
by
.I "getc(&queue)"
which returns either the (non-negative) character
or \(mi1 if the queue is empty.
.PP
Notice that the space for characters in queues is
shared among all devices in the system
and in the standard system there are only some 600
character slots available.
Thus device handlers,
especially write routines, must take
care to avoid gobbling up excessive numbers of characters.
.PP
The other major help available
to device handlers is the sleep-wakeup mechanism.
The call
.I "sleep(event, priority)"
causes the process to wait (allowing other processes to run)
until the
.I event
occurs;
at that time, the process is marked ready-to-run
and the call will return when there is no
process with higher
.I priority.
.PP
The call
.I "wakeup(event)"
indicates that the
.I event
has happened, that is, causes processes sleeping
on the event to be awakened.
The
.I event
is an arbitrary quantity agreed upon
by the sleeper and the waker-up.
By convention, it is the address of some data area used
by the driver, which guarantees that events
are unique.
.PP
Processes sleeping on an event should not assume
that the event has really happened;
they should check that the conditions which
caused them to sleep no longer hold.
.PP
Priorities can range from 0 to 127;
a higher numerical value indicates a less-favored
scheduling situation.
A distinction is made between processes sleeping
at priority less than the parameter
.I PZERO
and those at numerically larger priorities.
The former cannot
be interrupted by signals, although it
is conceivable that it may be swapped out.
Thus it is a bad idea to sleep with
priority less than PZERO on an event which might never occur.
On the other hand, calls to
.I sleep
with larger priority
may never return if the process is terminated by
some signal in the meantime.
Incidentally, it is a gross error to call
.I sleep
in a routine called at interrupt time, since the process
which is running is almost certainly not the
process which should go to sleep.
Likewise, none of the variables in the user area
``\fIu\fB.\fR''
should be touched, let alone changed, by an interrupt routine.
.PP
If a device driver
wishes to wait for some event for which it is inconvenient
or impossible to supply a
.I wakeup,
(for example, a device going on-line, which does not
generally cause an interrupt),
the call
.I "sleep(&lbolt, priority)
may be given.
.I Lbolt
is an external cell whose address is awakened once every 4 seconds
by the clock interrupt routine.
.PP
The routines
.I "spl4( ), spl5( ), spl6( ), spl7( )"
are available to
set the processor priority level as indicated to avoid
inconvenient interrupts from the device.
.PP
If a device needs to know about real-time intervals,
then
.I "timeout(func, arg, interval)
will be useful.
This routine arranges that after
.I interval
sixtieths of a second, the
.I func
will be called with
.I arg
as argument, in the style
.I "(*func)(arg).
Timeouts are used, for example,
to provide real-time delays after function characters
like new-line and tab in typewriter output,
and to terminate an attempt to
read the 201 Dataphone
.I dp
if there is no response within a specified number
of seconds.
Notice that the number of sixtieths of a second is limited to 32767,
since it must appear to be positive,
and that only a bounded number of timeouts
can be going on at once.
Also, the specified
.I func
is called at clock-interrupt time, so it should
conform to the requirements of interrupt routines
in general.
.SH
The Block-device Interface
.PP
Handling of block devices is mediated by a collection
of routines that manage a set of buffers containing
the images of blocks of data on the various devices.
The most important purpose of these routines is to assure
that several processes that access the same block of the same
device in multiprogrammed fashion maintain a consistent
view of the data in the block.
A secondary but still important purpose is to increase
the efficiency of the system by
keeping in-core copies of blocks that are being
accessed frequently.
The main data base for this mechanism is the
table of buffers
.I buf.
Each buffer header contains a pair of pointers
.I "(b_forw, b_back)"
which maintain a doubly-linked list
of the buffers associated with a particular
block device, and a
pair of pointers
.I "(av_forw, av_back)"
which generally maintain a doubly-linked list of blocks
which are ``free,'' that is,
eligible to be reallocated for another transaction.
Buffers that have I/O in progress
or are busy for other purposes do not appear in this list.
The buffer header
also contains the device and block number to which the
buffer refers, and a pointer to the actual storage associated with
the buffer.
There is a word count
which is the negative of the number of words
to be transferred to or from the buffer;
there is also an error byte and a residual word
count used to communicate information
from an I/O routine to its caller.
Finally, there is a flag word
with bits indicating the status of the buffer.
These flags will be discussed below.
.PP
Seven routines constitute
the most important part of the interface with the
rest of the system.
Given a device and block number,
both
.I bread
and
.I getblk
return a pointer to a buffer header for the block;
the difference is that
.I bread
is guaranteed to return a buffer actually containing the
current data for the block,
while
.I getblk
returns a buffer which contains the data in the
block only if it is already in core (whether it is
or not is indicated by the
.I B_DONE
bit; see below).
In either case the buffer, and the corresponding
device block, is made ``busy,''
so that other processes referring to it
are obliged to wait until it becomes free.
.I Getblk
is used, for example,
when a block is about to be totally rewritten,
so that its previous contents are
not useful;
still, no other process can be allowed to refer to the block
until the new data is placed into it.
.PP
The
.I breada
routine is used to implement read-ahead.
it is logically similar to
.I bread,
but takes as an additional argument the number of
a block (on the same device) to be read asynchronously
after the specifically requested block is available.
.PP
Given a pointer to a buffer,
the
.I brelse
routine
makes the buffer again available to other processes.
It is called, for example, after
data has been extracted following a
.I bread.
There are three subtly-different write routines,
all of which take a buffer pointer as argument,
and all of which logically release the buffer for
use by others and place it on the free list.
.I Bwrite
puts the
buffer on the appropriate device queue,
waits for the write to be done,
and sets the user's error flag if required.
.I Bawrite
places the buffer on the device's queue, but does not wait
for completion, so that errors cannot be reflected directly to
the user.
.I Bdwrite
does not start any I/O operation at all,
but merely marks
the buffer so that if it happens
to be grabbed from the free list to contain
data from some other block, the data in it will
first be written
out.
.PP
.I Bwrite
is used when one wants to be sure that
I/O takes place correctly, and that
errors are reflected to the proper user;
it is used, for example, when updating i-nodes.
.I Bawrite
is useful when more overlap is desired
(because no wait is required for I/O to finish)
but when it is reasonably certain that the
write is really required.
.I Bdwrite
is used when there is doubt that the write is
needed at the moment.
For example,
.I bdwrite
is called when the last byte of a
.I write
system call falls short of the end of a
block, on the assumption that
another
.I write
will be given soon which will re-use the same block.
On the other hand,
as the end of a block is passed,
.I bawrite
is called, since probably the block will
not be accessed again soon and one might as
well start the writing process as soon as possible.
.PP
In any event, notice that the routines
.I "getblk"
and
.I bread
dedicate the given block exclusively to the
use of the caller, and make others wait,
while one of
.I "brelse, bwrite, bawrite,"
or
.I bdwrite
must eventually be called to free the block for use by others.
.PP
As mentioned, each buffer header contains a flag
word which indicates the status of the buffer.
Since they provide
one important channel for information between the drivers and the
block I/O system, it is important to understand these flags.
The following names are manifest constants which
select the associated flag bits.
.IP B_READ 10
This bit is set when the buffer is handed to the device strategy routine
(see below) to indicate a read operation.
The symbol
.I B_WRITE
is defined as 0 and does not define a flag; it is provided
as a mnemonic convenience to callers of routines like
.I swap
which have a separate argument
which indicates read or write.
.IP B_DONE 10
This bit is set
to 0 when a block is handed to the the device strategy
routine and is turned on when the operation completes,
whether normally as the result of an error.
It is also used as part of the return argument of
.I getblk
to indicate if 1 that the returned
buffer actually contains the data in the requested block.
.IP B_ERROR 10
This bit may be set to 1 when
.I B_DONE
is set to indicate that an I/O or other error occurred.
If it is set the
.I b_error
byte of the buffer header may contain an error code
if it is non-zero.
If
.I b_error
is 0 the nature of the error is not specified.
Actually no driver at present sets
.I b_error;
the latter is provided for a future improvement
whereby a more detailed error-reporting
scheme may be implemented.
.IP B_BUSY 10
This bit indicates that the buffer header is not on
the free list, i.e. is
dedicated to someone's exclusive use.
The buffer still remains attached to the list of
blocks associated with its device, however.
When
.I getblk
(or
.I bread,
which calls it) searches the buffer list
for a given device and finds the requested
block with this bit on, it sleeps until the bit
clears.
.IP B_PHYS 10
This bit is set for raw I/O transactions that
need to allocate the Unibus map on an 11/70.
.IP B_MAP 10
This bit is set on buffers that have the Unibus map allocated,
so that the
.I iodone
routine knows to deallocate the map.
.IP B_WANTED 10
This flag is used in conjunction with the
.I B_BUSY
bit.
Before sleeping as described
just above,
.I getblk
sets this flag.
Conversely, when the block is freed and the busy bit
goes down (in
.I brelse)
a
.I wakeup
is given for the block header whenever
.I B_WANTED
is on.
This strategem avoids the overhead
of having to call
.I wakeup
every time a buffer is freed on the chance that someone
might want it.
.IP B_AGE
This bit may be set on buffers just before releasing them; if it
is on,
the buffer is placed at the head of the free list, rather than at the
tail.
It is a performance heuristic
used when the caller judges that the same block will not soon be used again.
.IP B_ASYNC 10
This bit is set by
.I bawrite
to indicate to the appropriate device driver
that the buffer should be released when the
write has been finished, usually at interrupt time.
The difference between
.I bwrite
and
.I bawrite
is that the former starts I/O, waits until it is done, and
frees the buffer.
The latter merely sets this bit and starts I/O.
The bit indicates that
.I relse
should be called for the buffer on completion.
.IP B_DELWRI 10
This bit is set by
.I bdwrite
before releasing the buffer.
When
.I getblk,
while searching for a free block,
discovers the bit is 1 in a buffer it would otherwise grab,
it causes the block to be written out before reusing it.
.SH
Block Device Drivers
.PP
The
.I bdevsw
table contains the names of the interface routines
and that of a table for each block device.
.PP
Just as for character devices, block device drivers may supply
an
.I open
and a
.I close
routine
called respectively on each open and on the final close
of the device.
Instead of separate read and write routines,
each block device driver has a
.I strategy
routine which is called with a pointer to a buffer
header as argument.
As discussed, the buffer header contains
a read/write flag, the core address,
the block number, a (negative) word count,
and the major and minor device number.
The role of the strategy routine
is to carry out the operation as requested by the
information in the buffer header.
When the transaction is complete the
.I B_DONE
(and possibly the
.I B_ERROR)
bits should be set.
Then if the
.I B_ASYNC
bit is set,
.I brelse
should be called;
otherwise,
.I wakeup.
In cases where the device
is capable, under error-free operation,
of transferring fewer words than requested,
the device's word-count register should be placed
in the residual count slot of
the buffer header;
otherwise, the residual count should be set to 0.
This particular mechanism is really for the benefit
of the magtape driver;
when reading this device
records shorter than requested are quite normal,
and the user should be told the actual length of the record.
.PP
Although the most usual argument
to the strategy routines
is a genuine buffer header allocated as discussed above,
all that is actually required
is that the argument be a pointer to a place containing the
appropriate information.
For example the
.I swap
routine, which manages movement
of core images to and from the swapping device,
uses the strategy routine
for this device.
Care has to be taken that
no extraneous bits get turned on in the
flag word.
.PP
The device's table specified by
.I bdevsw
has a
byte to contain an active flag and an error count,
a pair of links which constitute the
head of the chain of buffers for the device
.I "(b_forw, b_back),"
and a first and last pointer for a device queue.
Of these things, all are used solely by the device driver
itself
except for the buffer-chain pointers.
Typically the flag encodes the state of the
device, and is used at a minimum to
indicate that the device is currently engaged in
transferring information and no new command should be issued.
The error count is useful for counting retries
when errors occur.
The device queue is used to remember stacked requests;
in the simplest case it may be maintained as a first-in
first-out list.
Since buffers which have been handed over to
the strategy routines are never
on the list of free buffers,
the pointers in the buffer which maintain the free list
.I "(av_forw, av_back)"
are also used to contain the pointers
which maintain the device queues.
.PP
A couple of routines
are provided which are useful to block device drivers.
.I "iodone(bp)"
arranges that the buffer to which
.I bp
points be released or awakened,
as appropriate,
when the
strategy module has finished with the buffer,
either normally or after an error.
(In the latter case the
.I B_ERROR
bit has presumably been set.)
.PP
The routine
.I "geterror(bp)"
can be used to examine the error bit in a buffer header
and arrange that any error indication found therein is
reflected to the user.
It may be called only in the non-interrupt
part of a driver when I/O has completed
.I (B_DONE
has been set).
.SH
Raw Block-device I/O
.PP
A scheme has been set up whereby block device drivers may
provide the ability to transfer information
directly between the user's core image and the device
without the use of buffers and in blocks as large as
the caller requests.
The method involves setting up a character-type special file
corresponding to the raw device
and providing
.I read
and
.I write
routines which set up what is usually a private,
non-shared buffer header with the appropriate information
and call the device's strategy routine.
If desired, separate
.I open
and
.I close
routines may be provided but this is usually unnecessary.
A special-function routine might come in handy, especially for
magtape.
.PP
A great deal of work has to be done to generate the
``appropriate information''
to put in the argument buffer for
the strategy module;
the worst part is to map relocated user addresses to physical addresses.
Most of this work is done by
.I "physio(strat, bp, dev, rw)
whose arguments are the name of the
strategy routine
.I strat,
the buffer pointer
.I bp,
the device number
.I dev,
and a read-write flag
.I rw
whose value is either
.I B_READ
or
.I B_WRITE.
.I Physio
makes sure that the user's base address and count are
even (because most devices work in words)
and that the core area affected is contiguous
in physical space;
it delays until the buffer is not busy, and makes it
busy while the operation is in progress;
and it sets up user error return information.