The UNIX I/O System

Dennis M. Ritchie

      This paper gives an overview of the workings of the UNIX** 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.

      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.

Device Classes

      There are two classes of device: block and 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.

      Character-type devices have a much more straightforward interface, although more work must be done by the driver itself.

      Devices of both types are named by a major and a 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 major and 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.

      The major device numbers for block and character devices are used as indices in separate tables; they both start at 0 and therefore overlap.

Overview of I/O

      The purpose of the open and creat system calls is to set up entries in three separate system tables. The first of these is the u_ofile table, which is stored in the system's per-process data area u. This table is indexed by the file descriptor returned by the open or creat, and is accessed during a read, write, or other operation on the open file. An entry contains only a pointer to the corresponding entry of the file table, which is a per-system data base. There is one entry in the file table for each instance of open or 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 forks after the file is opened. A 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 inode table, which contains a copy of the file's i-node.

      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.

      An entry in the file table corresponds precisely to an instance of open or 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 inode table for a given file. Also, a file may enter the 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.

      An entry in the 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 long quantities.

      During the processing of an open or creat call for a special file, the system always calls the device's 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 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.

      When a read or write takes place, the user's arguments and the file table entry are used to set up the variables u.u_base, u.u_count, and 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 bmap routine. In any event, the resulting physical block number is used, as discussed below, to read or write the appropriate device.

Character Device Drivers

      The 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 ioctl system call). Any of these may be missing. If a call on the routine should be ignored, (e.g. open on non-exclusive devices that require no setup) the cdevsw entry can be given as nulldev; if it should be considered an error, (e.g. write on read-only devices) nodev is used. For terminals, the cdevsw structure also contains a pointer to the tty structure associated with the terminal.

      The 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.

      The 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 close.

      When write is called, it is supplied the device as argument. The per-user variable u.u_count has been set to the number of characters indicated by the user; for character devices, this number may be 0 initially. 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 u.u_segflg is supplied that indicates, if on, that u.u_base refers to the system address space instead of the user's.

      The write routine should copy up to u.u_count characters from the user's buffer to the device, decrementing u.u_count for each character passed. For most drivers, which work one character at a time, the routine cpass( ) is used to pick up characters from the user's buffer. Successive calls on it return the characters to be written until u.u_count goes to 0 or an error occurs, when it returns -1. Cpass takes care of interrogating u.u_segflg and updating u.u_count.

      Write routines which want to transfer a probably large number of characters into an internal buffer may also use the routine iomove(buffer, offset, count, flag) which is faster when many characters must be moved. Iomove transfers up to count characters into the buffer starting offset bytes from the start of the buffer; flag should be 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, iomove is much slower if any of buffer+offset, count or u.u_base is odd.

      The device's read routine is called under conditions similar to write, except that u.u_count is guaranteed to be non-zero. To return characters to the user, the routine passc(c) is available; it takes care of housekeeping like cpass and returns -1 as the last character specified by u.u_count is returned to the user; before that time, 0 is returned. Iomove is also usable as with write; the flag should be B_READ but the same cautions apply.

      The ``special-functions'' routine is invoked by the stty and gtty system calls as follows: (*p) (dev, v) where p is a pointer to the device's routine, dev is the device number, and v is a vector. In the 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 stty case, v is 0; the device should take up to 3 words of control information from the array u.u_arg[0...2].

      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.

      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 getc and putc. A queue header has the structure

struct {
int c_cc; /* character count */
char *c_cf; /* first character */
char *c_cl; /* last character */
} queue;
A character is placed on the end of a queue by putc(c, &queue) where c is the character and queue is the queue header. The routine returns -1 if there is no space to put the character, 0 otherwise. The first character on the queue may be retrieved by getc(&queue) which returns either the (non-negative) character or -1 if the queue is empty.

      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.

      The other major help available to device handlers is the sleep-wakeup mechanism. The call sleep(event, priority) causes the process to wait (allowing other processes to run) until the event occurs; at that time, the process is marked ready-to-run and the call will return when there is no process with higher priority.

      The call wakeup(event) indicates that the event has happened, that is, causes processes sleeping on the event to be awakened. The 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.

      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.

      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 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 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 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 ``u.'' should be touched, let alone changed, by an interrupt routine.

      If a device driver wishes to wait for some event for which it is inconvenient or impossible to supply a wakeup, (for example, a device going on-line, which does not generally cause an interrupt), the call sleep(&lbolt, priority) may be given. Lbolt is an external cell whose address is awakened once every 4 seconds by the clock interrupt routine.

      The routines spl4( ), spl5( ), spl6( ), spl7( ) are available to set the processor priority level as indicated to avoid inconvenient interrupts from the device.

      If a device needs to know about real-time intervals, then timeout(func, arg, interval) will be useful. This routine arranges that after interval sixtieths of a second, the func will be called with arg as argument, in the style (*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 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 func is called at clock-interrupt time, so it should conform to the requirements of interrupt routines in general.

The Block-device Interface

      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 buf. Each buffer header contains a pair of pointers (b_forw, b_back) which maintain a doubly-linked list of the buffers associated with a particular block device, and a pair of pointers (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.

      Seven routines constitute the most important part of the interface with the rest of the system. Given a device and block number, both bread and getblk return a pointer to a buffer header for the block; the difference is that bread is guaranteed to return a buffer actually containing the current data for the block, while 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 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. 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.

      The breada routine is used to implement read-ahead. it is logically similar to 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.

      Given a pointer to a buffer, the brelse routine makes the buffer again available to other processes. It is called, for example, after data has been extracted following a 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. 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. 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. 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.

      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. 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. Bdwrite is used when there is doubt that the write is needed at the moment. For example, bdwrite is called when the last byte of a write system call falls short of the end of a block, on the assumption that another write will be given soon which will re-use the same block. On the other hand, as the end of a block is passed, 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.

      In any event, notice that the routines getblk and bread dedicate the given block exclusively to the use of the caller, and make others wait, while one of brelse, bwrite, bawrite, or bdwrite must eventually be called to free the block for use by others.

      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.

This bit is set when the buffer is handed to the device strategy routine (see below) to indicate a read operation. The symbol B_WRITE is defined as 0 and does not define a flag; it is provided as a mnemonic convenience to callers of routines like swap which have a separate argument which indicates read or write.
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 getblk to indicate if 1 that the returned buffer actually contains the data in the requested block.
This bit may be set to 1 when B_DONE is set to indicate that an I/O or other error occurred. If it is set the b_error byte of the buffer header may contain an error code if it is non-zero. If b_error is 0 the nature of the error is not specified. Actually no driver at present sets b_error; the latter is provided for a future improvement whereby a more detailed error-reporting scheme may be implemented.
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 getblk (or 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.
This bit is set for raw I/O transactions that need to allocate the Unibus map on an 11/70.
This bit is set on buffers that have the Unibus map allocated, so that the iodone routine knows to deallocate the map.
This flag is used in conjunction with the B_BUSY bit. Before sleeping as described just above, getblk sets this flag. Conversely, when the block is freed and the busy bit goes down (in brelse) a wakeup is given for the block header whenever B_WANTED is on. This strategem avoids the overhead of having to call wakeup every time a buffer is freed on the chance that someone might want it.
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.
This bit is set by 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 bwrite and 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 relse should be called for the buffer on completion.
This bit is set by bdwrite before releasing the buffer. When 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.

Block Device Drivers

      The bdevsw table contains the names of the interface routines and that of a table for each block device.

      Just as for character devices, block device drivers may supply an open and a 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 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 B_DONE (and possibly the B_ERROR) bits should be set. Then if the B_ASYNC bit is set, brelse should be called; otherwise, 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.

      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 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.

      The device's table specified by 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 (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 (av_forw, av_back) are also used to contain the pointers which maintain the device queues.

      A couple of routines are provided which are useful to block device drivers. iodone(bp) arranges that the buffer to which 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 B_ERROR bit has presumably been set.)

      The routine 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 (B_DONE has been set).

Raw Block-device I/O

      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 read and 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 open and close routines may be provided but this is usually unnecessary. A special-function routine might come in handy, especially for magtape.

      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 physio(strat, bp, dev, rw) whose arguments are the name of the strategy routine strat, the buffer pointer bp, the device number dev, and a read-write flag rw whose value is either B_READ or B_WRITE. 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.

Table of Contents