Provided by: freebsd-manpages_10.1~RC1-1_all
sa — SCSI Sequential Access device driver
The sa driver provides support for all SCSI devices of the sequential access class that are attached to the system through a supported SCSI Host Adapter. The sequential access class includes tape and other linear access devices. A SCSI Host adapter must also be separately configured into the system before a SCSI sequential access device can be configured.
The sa driver is based around the concept of a “mount session”, which is defined as the period between the time that a tape is mounted, and the time when it is unmounted. Any parameters set during a mount session remain in effect for the remainder of the session or until replaced. The tape can be unmounted, bringing the session to a close in several ways. These include: 1. Closing a `rewind device', referred to as sub-mode 00 below. An example is /dev/sa0. 2. Using the MTOFFL ioctl(2) command, reachable through the ‘offline’ command of mt(1). It should be noted that tape devices are exclusive open devices, except in the case where a control mode device is opened. In the latter case, exclusive access is only sought when needed (e.g., to set parameters).
Bits 0 and 1 of the minor number are interpreted as ‘sub-modes’. The sub-modes differ in the action taken when the device is closed: 00 A close will rewind the device; if the tape has been written, then a file mark will be written before the rewind is requested. The device is unmounted. 01 A close will leave the tape mounted. If the tape was written to, a file mark will be written. No other head positioning takes place. Any further reads or writes will occur directly after the last read, or the written file mark. 10 A close will rewind the device. If the tape has been written, then a file mark will be written before the rewind is requested. On completion of the rewind an unload command will be issued. The device is unmounted.
SCSI tapes may run in either ‘variable’ or ‘fixed’ block-size modes. Most QIC-type devices run in fixed block-size mode, where most nine-track tapes and many new cartridge formats allow variable block-size. The difference between the two is as follows: Variable block-size: Each write made to the device results in a single logical record written to the tape. One can never read or write part of a record from tape (though you may request a larger block and read a smaller record); nor can one read multiple blocks. Data from a single write is therefore read by a single read. The block size used may be any value supported by the device, the SCSI adapter and the system (usually between 1 byte and 64 Kbytes, sometimes more). When reading a variable record/block from the tape, the head is logically considered to be immediately after the last item read, and before the next item after that. If the next item is a file mark, but it was never read, then the next process to read will immediately hit the file mark and receive an end-of-file notification. Fixed block-size: Data written by the user is passed to the tape as a succession of fixed size blocks. It may be contiguous in memory, but it is considered to be a series of independent blocks. One may never write an amount of data that is not an exact multiple of the blocksize. One may read and write the same data as a different set of records. In other words, blocks that were written together may be read separately, and vice-versa. If one requests more blocks than remain in the file, the drive will encounter the file mark. As there is some data to return (unless there were no records before the file mark), the read will succeed, returning that data. The next read will return immediately with a value of 0. (As above, if the file mark is never read, it remains for the next process to read if in no-rewind mode.)
By default, the driver will NOT accept reads or writes to a tape device that are larger than may be written to or read from the mounted tape using a single write or read request. Because of this, the application author may have confidence that his wishes are respected in terms of the block size written to tape. For example, if the user tries to write a 256KB block to the tape, but the controller can handle no more than 128KB, the write will fail. The previous FreeBSD behavior, prior to FreeBSD 10.0, was to break up large reads or writes into smaller blocks when going to the tape. The problem with that behavior, though, is that it hides the actual on-tape block size from the application writer, at least in variable block mode. If the user would like his large reads and writes broken up into separate pieces, he may set the following loader tunables. Note that these tunables WILL GO AWAY in FreeBSD 11.0. They are provided for transition purposes only. kern.cam.sa.allow_io_split This variable, when set to 1, will configure all sa devices to split large buffers into smaller pieces when needed. kern.cam.sa.%d.allow_io_split This variable, when set to 1, will configure the given sa unit to split large buffers into multiple pieces. This will override the global setting, if it exists. There are several sysctl(8) variables available to view block handling parameters: kern.cam.sa.%d.allow_io_split This variable allows the user to see, but not modify, the current I/O split setting. The user is not permitted to modify this setting so that there is no chance of behavior changing for the application while a tape is mounted. kern.cam.sa.%d.maxio This variable shows the maximum I/O size in bytes that is allowed by the combination of kernel tuning parameters (MAXPHYS, DFLTPHYS) and the capabilities of the controller that is attached to the tape drive. Applications may look at this value for a guide on how large an I/O may be permitted, but should keep in mind that the actual maximum may be restricted further by the tape drive via the SCSI READ BLOCK LIMITS command. kern.cam.sa.%d.cpi_maxio This variable shows the maximum I/O size supported by the controller, in bytes, that is reported via the CAM Path Inquiry CCB (XPT_PATH_INQ). If this is 0, that means that the controller has not reported a maximum I/O size.
FILE MARK HANDLING
The handling of file marks on write is automatic. If the user has written to the tape, and has not done a read since the last write, then a file mark will be written to the tape when the device is closed. If a rewind is requested after a write, then the driver assumes that the last file on the tape has been written, and ensures that there are two file marks written to the tape. The exception to this is that there seems to be a standard (which we follow, but do not understand why) that certain types of tape do not actually write two file marks to tape, but when read, report a `phantom' file mark when the last file is read. These devices include the QIC family of devices. (It might be that this set of devices is the same set as that of fixed block devices. This has not been determined yet, and they are treated as separate behaviors by the driver at this time.)
The sa driver supports all of the ioctls of mtio(4).
/dev/[n][e]sa[0-9] general form: /dev/sa0 Rewind on close /dev/nsa0 No rewind on close /dev/esa0 Eject on close (if capable) /dev/sa0.ctl Control mode device (to examine state while another program is accessing the device, e.g.).
The sa driver was written for the CAM SCSI subsystem by Justin T. Gibbs and Kenneth Merry. Many ideas were gleaned from the st device driver written and ported from Mach 2.5 by Julian Elischer. The current owner of record is Matthew Jacob who has suffered too many years of breaking tape drivers.
This driver lacks many of the hacks required to deal with older devices. Many older SCSI-1 devices may not work properly with this driver yet. Additionally, certain tapes (QIC tapes mostly) that were written under FreeBSD 2.X are not automatically read correctly with this driver: you may need to explicitly set variable block mode or set to the blocksize that works best for your device in order to read tapes written under FreeBSD 2.X. Fine grained density and compression mode support that is bound to specific device names needs to be added. Support for fast indexing by use of partitions is missing.