6.2. Character Devices Almost all peripherals on the system, except network interfaces, have a character-device interface. A character device usually maps the hardware interface into a byte stream, similar to that of the filesystem. Character devices of this type include terminals (e.g., /dev/tty00), line printers (e.g, /dev/lp0), an interface to physical main memory (/dev/mem), and a bottomless sink for data and an endless source of end-of-file markers (/dev/null). Some of these character devices, such as terminal devices, may display special behavior on line boundaries but in general are still treated as byte streams. Devices emulating terminals use buffers that are smaller than those used for disks. This buffering system involves small (usually 64-byte) blocks of characters kept in linked lists. Although all free character buffers are kept in a single free list, most device drivers that use them limit the number of characters that can be queued at one time for a single terminal port. Devices such as high-speed graphics interfaces may have their own buffers or may always do I/O directly into the address space of the user; they too are classed as character devices. Some of these drivers may recognize special types of records and thus be further from the plain byte-stream model. The character interface for disks is also called the raw device interface; it provides an unstructured interface to the device. Its primary task is to arrange for direct I/O to and from the device. The disk driver handles the asynchronous nature of I/O by maintaining and ordering an active queue of pending transfers. Each entry in the queue specifies whether it is for reading or writing, the main-memory address for the transfer, the device address for the transfer (usually a disk sector number), and the transfer size (in bytes). All other restrictions of the underlying hardware are passed through the character interface to its clients, making character-device interfaces the furthest from the byte-stream model. Thus, the user process must abide by the sectoring restrictions imposed by the underlying hardware. For magnetic disks, the file offset and transfer size must be a multiple of the sector size. The character interface does not copy the user data into a kernel buffer before putting them on an I/O queue. Instead, it arranges to have the I/O done directly to or from the address space of the process. The size and alignment of the transfer is limited by the physical device. However, the transfer size is not restricted by the maximum size of the internal buffers of the system because these buffers are not used. The character interface is typically used by only those system utility programs that have an intimate knowledge of the data structures on the disk. The character interface also allows user-level prototyping; for example, the 4.2BSD filesystem implementation was written and largely tested as a user process that used a raw disk interface, before the code was moved into the kernel. Character devices are described by entries in the cdevsw structure. The entry points in this structure (see Table 6.1 on page 220) are used to support raw access to block-oriented devices such as disk, as well as normal access to character-oriented devices through the terminal driver. Raw devices support a subset of the entry points that correspond to those entry points found in block-oriented devices. The base set of entry points for all device drivers is described in this section; the additional set of entry points for block-oriented devices is given in Section 6.3. Table 6.1. Entry points for character and raw device drivers.Entry point | Function |
|---|
open() | open the device | close() | close the device | read() | do an input operation | write() | do an output operation | ioctl() | do an I/O control operation | poll() | poll device for I/O readiness | stop() | stop output on the device | mmap() | map device offset to memory location | reset() | reinitialize device after a bus reset |
Raw Devices and Physical I/O Most raw devices differ from filesystems only in the way that they do I/O. Whereas filesystems read and write data to and from kernel buffers, raw devices transfer data to and from user buffers. Bypassing kernel buffers eliminates the memory-to-memory copy that must be done by filesystems but also denies applications the benefits of data caching. In addition, for devices that support both raw- and filesystem access, applications must take care to preserve consistency between data in the kernel buffers and data written directly to the device. The raw device should be used only when the filesystem is unmounted or mounted readonly. Raw-device access is used by many filesystem utilities, such as the filesystem check program, fsck, and by programs that read and write backup media for example, tar, dump, and restore. Because raw devices bypass kernel buffers, they are responsible for managing their own buffer structures. Most devices borrow swap buffers to describe their I/O. The read and write routines use the physio() routine to start a raw I/O operation (see Figure 6.2). The strategy parameter identifies a block-device strategy routine that starts I/O operations on the device. The buffer is used by physio() in constructing the request(s) made to the strategy routine. The device, read-write flag, and uio parameters completely specify the I/O operation that should be done. The maximum transfer size for the device is checked by physio () to adjust the size of each I/O transfer before the latter is passed to the strategy routine. This check allows the transfer to be done in sections according to the maximum transfer size supported by the device. Figure 6.2. Algorithm for physical I/O. void physio( device dev, struct uio *uio, int ioflag); { allocate a swap buffer; while (uio is not exhausted) { mark the buffer busy for physical I/O; set up the buffer for a maximum-sized transfer; use device maximum I/O size to bound the transfer size; check user read/write access at uio location; lock the part of the user address space involved in the transfer into RAM; map the user pages into the buffer; call dev->strategy() to start the transfer; wait for the transfer to complete; unmap the user pages from the buffer; unlock the part of the address space previously locked; deduct the transfer size from the total number of data to transfer; } free swap buffer; } Raw-device I/O operations request the hardware device to transfer data directly to or from the data buffer in the user program's address space described by the uio parameter. Thus, unlike I/O operations that do direct memory access (DMA) from buffers in the kernel address space, raw I/O operations must check that the user's buffer is accessible by the device and must lock it into memory for the duration of the transfer. Character-oriented Devices Character-oriented I/O devices are typified by terminal ports, although they also include printers and other character- or line-oriented devices. These devices are usually accessed through the terminal driver, described in Chapter 10. The close tie to the terminal driver has heavily influenced the structure of character device drivers. For example, several entry points in the cdevsw structure exist for communication between the generic terminal handler and the terminal multiplexer hardware drivers. Entry Points for Character Device Drivers A device driver for a character device is defined by its entries in a cdevsw structure. open | Open the device in preparation for I/O operations. A device's open entry point will be called for each open system call on a special-device file, or, internally, when a device is prepared for mounting a filesystem with the mount system call. The open() routine will commonly verify the integrity of the associated medium. For example, it will verify that the device was identified during the autoconfiguration phase and, for disk drives, that a medium is present and ready to accept commands. | close | Close a device. The close() routine is called after the final client interested in using the device terminates. These semantics are defined by the higher-level I/O facilities. Disk devices have nothing to do when a device is closed and thus use a null close() routine. Devices that support access to only a single client must mark the device as available once again. | read | Read data from a device. For raw devices, this entry point normally just calls the physio () routine with device-specific parameters. For terminal-oriented devices, a read request is passed immediately to the terminal driver. For other devices, a read request requires that the specified data be copied into the kernel's address space, typically with the uiomove() routine (see the end of Section 6.4), and then be passed to the device. | write | Write data to a device. This entry point is a direct parallel of the read entry point: Raw devices use physio (), terminal-oriented devices call the terminal driver to do this operation, and other devices handle the request internally. | ioctl | Do an operation other than a read or write. This entry point originally provided a mechanism to get and set device parameters for terminal devices; its use has expanded to other types of devices as well. Historically, ioctl() operations have varied widely from device to device. | poll | Check the device to see whether data are available for reading, or space is available for writing, data. The poll entry point is used by the select and poll system calls in checking file descriptors associated with device special files. For raw devices, a poll operation is meaningless, since data are not buffered. Here, the entry point is set to seltrue(), a routine that returns true for any poll request. For devices used with the terminal driver, this entry point is set to ttselect(), a routine described in Chapter 10. | mmap | Map a device offset into a memory address. This entry point is called by the virtual-memory system to convert a logical mapping to a physical address. For example, it converts an offset in /dev/mem to a kernel address. | kqfilter | Add the device to the kernel event list for the calling thread. |
|