Rationale

The rationale can be roughly summarized with the following points.

1. To increase performance by providing a mechanism to carry out I/O without blocking. Theoretically, if I/O would never block, neither at the software nor at the hardware level, the overhead of I/O would become zero, and processes would no longer be I/O bound.
2. To segregate the different I/O operations into logically distinctive procedures. Unlike with the standard stdio(3), the aio interface separates queuing and submitting I/O operations to the kernel, and receiving notifications of operation completion from the kernel.
3. To provide an uniform and standardized framework for asynchronous I/O. For instance, aio avoids the need for (and the overhead of) extra worker threads sometimes used to perform asynchronous I/O.

Asynchronous I/O Control Block

The Asynchronous I/O Control Block is the basic operational unit behind aio. This is required since an arbitrary number of operations can be started at once, and because each operation can be either input or output. This block is represented by the aiocb structure, which is defined in the <aio.h> header. The following fields are available for user applications:

off_t		 aio_offset;
void		*aio_buf;
size_t		 aio_nbytes;
int		 aio_fildes;
int		 aio_lio_opcode;
int		 aio_reqprio;
struct sigevent	 aio_sigevent;

The fields are:

1. The aio_offset specifies the implicit file offset at which the I/O operations are performed. This cannot be expected to be the actual read/write offset of the file descriptor.
2. The aio_buf member is a pointer to the buffer to which data is going to be written or to which the read operation stores data.
3. The aio_nbytes specifies the length of aio_buf.
4. The aio_fildes specifies the used file descriptor.
5. The aio_lio_opcode is used by the lio_listio() function to initialize a list of I/O requests with a single call.
6. The aio_reqprio member can be used to lower the scheduling priority of an aio operation. This is only available if _POSIX_PRIORITIZED_IO and _POSIX_PRIORITY_SCHEDULING are defined, and the associated file descriptor supports it.
7. The aio_sigevent member is used to specify how the calling process is notified once an aio operation completes.

The members aio_buf, aio_fildes, and aio_nbytes are conceptually similar to the parameters ‘buf’, ‘fildes’, and ‘nbytes’ used in the standard read(2) and write(2) functions. For example, the caller can read aio_nbytes from a file associated with the file descriptor aio_fildes into the buffer aio_buf. All appropriate fields should be initialized by the caller before aio_read() or aio_write() is called.

File Offsets

Asynchronous I/O operations are not strictly sequential; operations are carried out in arbitrary order and more than one operation for one file descriptor can be started. The requested read or write operation starts from the absolute position specified by aio_offset, as if lseek(2) would have been called with SEEK_SET immediately prior to the operation. The POSIX standard does not specify what happens after an aio operation has been successfully completed. Depending on the implementation, the actual file offset may or may not be updated.

Errors and Completion

Asynchronous I/O operations are said to be complete when:

• An error is detected.
• The I/O transfer is performed successfully.
• The operation is canceled.

If an error condition is detected that prevents an operation from being started, the request is not enqueued. In this case the read and write functions, aio_read() and aio_write(), return immediately, setting the global errno to indicate the cause of the error.

After an operation has been successfully enqueued, aio_error() and aio_return() must be used to determine the status of the operation and to determine any error conditions. This includes the conditions reported by the standard read(2), write(2), and fsync(2). The request remains enqueued and consumes process and system resources until aio_return() is called.

Waiting for Completion

The aio interface supports both polling and notification models. The first can be implemented by simply repeatedly calling the aio_error() function to test the status of an operation. Once the operation has completed, aio_return() is used to free the aiocb structure for re-use.

The notification model is implemented by using the aio_sigevent member of the Asynchronous I/O Control Block. The operational model and the used structure are described in sigevent(3).

The aio_suspend() function can be used to wait for the completion of one or more operations. It is possible to set a timeout so that the process can continue the execution and take recovery actions if the aio operations do not complete as expected.

Cancellation and Synchronization

The aio_cancel() function can be used to request cancellation of an asynchronous I/O operation. Note however that not all of them can be canceled. The same aiocb used to start the operation may be used as a handle for identification. It is also possible to request cancellation of all operations pending for a file.

Comparable to fsync(2), the aio_fsync() function can be used to synchronize the contents of permanent storage when multiple asynchronous I/O operations are outstanding for the file or device. The synchronization operation includes only those requests that have already been successfully enqueued.