kernel_optimize_test/Documentation/ide/ide-tape.txt

/*
 * IDE ATAPI streaming tape driver.
 *
 * This driver is a part of the Linux ide driver.
 *
 * The driver, in co-operation with ide.c, basically traverses the
 * request-list for the block device interface. The character device
 * interface, on the other hand, creates new requests, adds them
 * to the request-list of the block device, and waits for their completion.
 *
 * Pipelined operation mode is now supported on both reads and writes.
 *
 * The block device major and minor numbers are determined from the
 * tape's relative position in the ide interfaces, as explained in ide.c.
 *
 * The character device interface consists of the following devices:
 *
 * ht0		major 37, minor 0	first  IDE tape, rewind on close.
 * ht1		major 37, minor 1	second IDE tape, rewind on close.
 * ...
 * nht0		major 37, minor 128	first  IDE tape, no rewind on close.
 * nht1		major 37, minor 129	second IDE tape, no rewind on close.
 * ...
 *
 * The general magnetic tape commands compatible interface, as defined by
 * include/linux/mtio.h, is accessible through the character device.
 *
 * General ide driver configuration options, such as the interrupt-unmask
 * flag, can be configured by issuing an ioctl to the block device interface,
 * as any other ide device.
 *
 * Our own ide-tape ioctl's can be issued to either the block device or
 * the character device interface.
 *
 * Maximal throughput with minimal bus load will usually be achieved in the
 * following scenario:
 *
 *	1.	ide-tape is operating in the pipelined operation mode.
 *	2.	No buffering is performed by the user backup program.
 *
 * Testing was done with a 2 GB CONNER CTMA 4000 IDE ATAPI Streaming Tape Drive.
 *
 * Here are some words from the first releases of hd.c, which are quoted
 * in ide.c and apply here as well:
 *
 * | Special care is recommended.  Have Fun!
 *
 *
 * An overview of the pipelined operation mode.
 *
 * In the pipelined write mode, we will usually just add requests to our
 * pipeline and return immediately, before we even start to service them. The
 * user program will then have enough time to prepare the next request while
 * we are still busy servicing previous requests. In the pipelined read mode,
 * the situation is similar - we add read-ahead requests into the pipeline,
 * before the user even requested them.
 *
 * The pipeline can be viewed as a "safety net" which will be activated when
 * the system load is high and prevents the user backup program from keeping up
 * with the current tape speed. At this point, the pipeline will get
 * shorter and shorter but the tape will still be streaming at the same speed.
 * Assuming we have enough pipeline stages, the system load will hopefully
 * decrease before the pipeline is completely empty, and the backup program
 * will be able to "catch up" and refill the pipeline again.
 *
 * When using the pipelined mode, it would be best to disable any type of
 * buffering done by the user program, as ide-tape already provides all the
 * benefits in the kernel, where it can be done in a more efficient way.
 * As we will usually not block the user program on a request, the most
 * efficient user code will then be a simple read-write-read-... cycle.
 * Any additional logic will usually just slow down the backup process.
 *
 * Using the pipelined mode, I get a constant over 400 KBps throughput,
 * which seems to be the maximum throughput supported by my tape.
 *
 * However, there are some downfalls:
 *
 *	1.	We use memory (for data buffers) in proportional to the number
 *		of pipeline stages (each stage is about 26 KB with my tape).
 *	2.	In the pipelined write mode, we cheat and postpone error codes
 *		to the user task. In read mode, the actual tape position
 *		will be a bit further than the last requested block.
 *
 * Concerning (1):
 *
 *	1.	We allocate stages dynamically only when we need them. When
 *		we don't need them, we don't consume additional memory. In
 *		case we can't allocate stages, we just manage without them
 *		(at the expense of decreased throughput) so when Linux is
 *		tight in memory, we will not pose additional difficulties.
 *
 *	2.	The maximum number of stages (which is, in fact, the maximum
 *		amount of memory) which we allocate is limited by the compile
 *		time parameter IDETAPE_MAX_PIPELINE_STAGES.
 *
 *	3.	The maximum number of stages is a controlled parameter - We
 *		don't start from the user defined maximum number of stages
 *		but from the lower IDETAPE_MIN_PIPELINE_STAGES (again, we
 *		will not even allocate this amount of stages if the user
 *		program can't handle the speed). We then implement a feedback
 *		loop which checks if the pipeline is empty, and if it is, we
 *		increase the maximum number of stages as necessary until we
 *		reach the optimum value which just manages to keep the tape
 *		busy with minimum allocated memory or until we reach
 *		IDETAPE_MAX_PIPELINE_STAGES.
 *
 * Concerning (2):
 *
 *	In pipelined write mode, ide-tape can not return accurate error codes
 *	to the user program since we usually just add the request to the
 *      pipeline without waiting for it to be serviced. In case an error
 *      occurs, I will report it on the next user request.
 *
 *	In the pipelined read mode, subsequent read requests or forward
 *	filemark spacing will perform correctly, as we preserve all blocks
 *	and filemarks which we encountered during our excess read-ahead.
 *
 *	For accurate tape positioning and error reporting, disabling
 *	pipelined mode might be the best option.
 *
 * You can enable/disable/tune the pipelined operation mode by adjusting
 * the compile time parameters below.
 *
 *
 *	Possible improvements.
 *
 *	1.	Support for the ATAPI overlap protocol.
 *
 *		In order to maximize bus throughput, we currently use the DSC
 *		overlap method which enables ide.c to service requests from the
 *		other device while the tape is busy executing a command. The
 *		DSC overlap method involves polling the tape's status register
 *		for the DSC bit, and servicing the other device while the tape
 *		isn't ready.
 *
 *		In the current QIC development standard (December 1995),
 *		it is recommended that new tape drives will *in addition*
 *		implement the ATAPI overlap protocol, which is used for the
 *		same purpose - efficient use of the IDE bus, but is interrupt
 *		driven and thus has much less CPU overhead.
 *
 *		ATAPI overlap is likely to be supported in most new ATAPI
 *		devices, including new ATAPI cdroms, and thus provides us
 *		a method by which we can achieve higher throughput when
 *		sharing a (fast) ATA-2 disk with any (slow) new ATAPI device.
 */