sg_raw [--binary] [--cmdfile=CF] [--cmdset=CS] [--enumerate] [--help] [--infile=IFILE] [--nosense] [--nvm] [--outfile=OFILE] [--raw] [--readonly] [--request=RLEN] [--scan=FO,LO] [--send=SLEN] [--skip=KLEN] [--timeout=SECS] [--verbose] [--version] DEVICE [CDB0 CDB1 ...]
This utility sends an arbitrary SCSI command (between 6 and 256 bytes) to the DEVICE. There may be no associated data transfer; or data may be read from a file and sent to the DEVICE; or data may be received from the DEVICE and then displayed or written to a file. If supported by the pass through, bidirectional commands may be sent (i.e. containing both data to be sent to the DEVICE and received from the DEVICE).
The SCSI command may be between 6 and 256 bytes long. Each command byte is specified in plain hex format (00..FF) without a prefix or suffix. The command can be given either on the command line or via the --cmdfile=CF option. See Examples section below.
The commands pass through a generic SCSI interface which is implemented for several operating systems including Linux, FreeBSD and Windows.
Experimental support has been added to send NVMe Admin and NVM commands to the DEVICE. Since all NVMe commands are 64 bytes long it is more convenient to use the --cmdfile=CF option rather than type the 64 bytes of the NVMe command on the command line. See the section on NVME below. A heuristic based on command length is used to decide if the given command is SCSI or NVMe, to override this heuristic use the --cmdset=CS option.
Arguments to long options are mandatory for short options as well. The options are arranged in alphabetical order based on the long option name.
- -b, --binary
Dump data in binary form, even when writing to stdout.
- -c, --cmdfile=CF
CF is the name of a file which contains the command to be executed. Without this option the command must be given on the command line, after the options and the DEVICE.
- -C, --cmdset=CS
CS is a number to indicate which command set (i.e. SCSI or NVMe) to use. 0, the default, causes a heuristic based on command length to be used. Use a CS of 1 to override that heuristic and choose the SCSI command set. Use a CS of 2 to override that heuristic and choose the NVMe command set.
- -h, --help
Display usage information and exit.
- -i, --infile=IFILE
Read data from IFILE instead of stdin. This option is ignored if --send is not specified.
- -n, --nosense
Don't display SCSI Sense information.
- -N, --nvm
When sending NVMe commands, the Admin command set is assumed. To send the NVM command set (e.g. the Read and Write (user data) commands) this option needs to be given.
- -o, --outfile=OFILE
Write data received from the DEVICE to OFILE. The data is written in binary. By default, data is dumped in hex format to stdout. If OFILE is '-' then data is dumped in binary to stdout. This option is ignored if --request is not specified.
- -w, --raw
interpret CF (i.e. the command file) as containing binary. The default is to assume that it contains ASCII hexadecimal.
- -R, --readonly
Open DEVICE read-only. The default (without this option) is to open it read-write.
- -r, --request=RLEN
Expect to receive up to RLEN bytes of data from the DEVICE. RLEN may be suffixed with 'k' to use kilobytes (1024 bytes) instead of bytes. RLEN is decimal unless it has a leading '0x' or a trailing 'h'.
If RLEN is too small (i.e. either smaller than indicated by the cdb (typically the "allocation length" field) and/or smaller than the DEVICE tries to send back) then the HBA driver may complain. Making RLEN larger than required should cause no problems. Most SCSI "data-in" commands return a data block that contains (in its early bytes) a length that the DEVICE would "like" to send back if the "allocation length" field in the cdb is large enough. In practice, the DEVICE will return no more bytes than indicated in the "allocation length" field of the cdb.
- -Q, --scan=FO,LO
Scan a range of opcodes (i.e. first byte of each command). The first opcode in the scan is FO (which is decimal unless it has a '0x' prefix or 'h' suffix). The last opcode in the scan is LO. The maximum value of LO is 255. The remaining bytes of the SCSI/NVMe command are as supplied at invocation.
Warning: this option can be dangerous. Sending somewhat arbitrary commands to a device can have unexpected results. It is recommended that this option is used with the --cmdset=CS option where CS is 1 or 2 in order to stop the command set possibly changing during the scan.
- -s, --send=SLEN
Read SLEN bytes of data, either from stdin or from a file, and send them to the DEVICE. In the SCSI transport, SLEN becomes the length (in bytes) of the "data-out" buffer. SLEN is decimal unless it has a leading '0x' or a trailing 'h'.
It is the responsibility of the user to make sure that the "data-out" length implied or stated in the cdb matches SLEN. Note that some common SCSI commands such as WRITE(10) have a "transfer length" field whose units are logical blocks (which are usually 512 or 4096 bytes long).
- -k, --skip=KLEN
Skip the first KLEN bytes of the input file or stream. This option is ignored if --send is not specified. If --send is given and this option is not given, then zero bytes are skipped.
- -t, --timeout=SECS
Wait up to SECS seconds for command completion (default: 20). Note that if a command times out the operating system may start by aborting the command and if that is unsuccessful it may attempt to reset the device.
- -v, --verbose
Increase level of verbosity. Can be used multiple times.
- -V, --version
Display version and license information and exit.
The sg_inq utility can be used to send an INQUIRY command to a device to determine its peripheral device type (e.g. '1' for a streaming device (tape drive)) which determines which SCSI command sets a device should support (e.g. SPC and SSC). The sg_vpd utility reads and decodes a device's Vital Product Pages which may contain useful information.
The ability to send more than a 16 byte CDB (in some cases 12 byte CDB) may be restricted by the pass-through interface, the low level driver or the transport. In the Linux series 3 kernels, the bsg driver can handle longer CDBs, block devices (e.g. /dev/sdc) accessed via the SG_IO ioctl cannot handle CDBs longer than 16 bytes, and the sg driver can handle longer CDBs from lk 3.17 .
The CDB command name defined by T10 for the given CDB is shown if the '-vv' option is given. The command line syntax still needs to be correct, so /dev/null may be used for the DEVICE since the CDB command name decoding is done before the DEVICE is checked.
The intention of the --scan=FO,LO option is to slightly simplify the process of finding hidden or undocumented commands. It should be used with care; for example checking for vendor specific SCSI commands: 'sg_raw --cmdset=1 --scan=0xc0,0xff /dev/sg1 0 0 0 0 0 0'.
Support for NVMe (a.k.a. NVM Express) is currently experimental. NVMe concepts map reasonably well to the SCSI architecture. A SCSI logical unit (LU) is similar to a NVMe namespace (although LUN 0 is very common in SCSI while namespace IDs start at 1). A SCSI target device is similar to a NVMe controller. SCSI commands vary from 6 to 260 bytes long (although SCSI command descriptor blocks (cdb_s) longer than 32 bytes are uncommon) while all NVMe commands are currently 64 bytes long. The SCSI architecture makes a clear distinction between an initiator (often called a HBA) and a target (device) while (at least on the PCIe transport) the NVMe controller plays both roles. This utility defaults to assuming the user provided 64 byte command belongs to NVMe's Admin command set. To issue commands from the "NVM" command set, the --nvm option must be given. Admin and NVM commands are sent to submission queue 0.
One significant difference is that SCSI uses a big endian representation for integers that are longer than 8 bits (i.e. longer than 1 byte) while NVMe uses a little endian representation (like most things that have originated from the Intel organisation). NVMe specifications talk about Words (16 bits), Double Words (32 bits) and sometimes Quad Words (64 bits) and has tighter alignment requirements than SCSI.
One difference that impacts this utility is that NVMe places pointers to host memory in its commands while SCSI leaves this detail to whichever transport it is using (e.g. SAS, iSCSI, SRP). Since this utility takes the command from the user (either on the command line or in a file named CF) but this utility allocates a data-in or data-out buffer as required, the user does not know in advance what the address of that buffer will be. Some special addresses have been introduced to help with this problem: the address 0xfffffffffffffffe is interpreted as "use the data-in buffer's address" while 0xfffffffffffffffd is interpreted as "use the data-out buffer's address". Since NVMe uses little endian notation then that first address appears in the NVMe command byte stream as "fe" followed by seven "ff"s. A similar arrangement is made for the length of that buffer (in bytes), but since that is a 32 byte quantity, the first 4 bytes (all "ff"s) are removed.
Several command file examples can be found in the examples directory of this package's source tarball: nvme_identify_ctl.hex, nvme_dev_self_test.hex, nvme_read_ctl.hex and nvme_write_ctl.hex .
Beware: the NVMe standard often refers to some of its fields as "0's based". They are typically counts of something like the number of blocks to be read. For example in NVMe Read command, a "0's based" number of blocks field containing the value 3 means to read 4 blocks! No, this is not a joke.
These examples, apart from the last one, use Linux device names. For suitable device names in other supported Operating Systems see the sg3_utils(8) man page.
- sg_raw /dev/scd0 1b 00 00 00 02 00
Eject the medium in CD drive /dev/scd0.
- sg_raw -r 1k /dev/sg0 12 00 00 00 60 00
Perform an INQUIRY on /dev/sg0 and dump the response data (up to 1024 bytes) to stdout.
- sg_raw -s 512 -i i512.bin /dev/sda 3b 02 00 00 00 00 00 02 00 00
Showing an example of writing 512 bytes to a sector on a disk is a little dangerous. Instead this example will read i512.bin (assumed to be 512 bytes long) and use the SCSI WRITE BUFFER command to send it to the "data" buffer (that is mode 2). This is a safe operation.
- sg_raw -r 512 -o o512.bin /dev/sda 3c 02 00 00 00 00 00 02 00 00
This will use the SCSI READ BUFFER command to read 512 bytes from the "data" buffer (i.e. mode 2) then write it to the o512.bin file. When used in conjunction with the previous example, if both commands work then 'cmp i512.bin o512.bin' should show a match.
- sg_raw --infile=urandom.bin --send=512 --request=512 --outfile=out.bin "/dev/bsg/7:0:0:0" 53 00 00 00 00 00 00 00 01 00
This is a bidirectional XDWRITEREAD(10) command being sent via a Linux bsg device. Note that data is being read from "urandom.bin" and sent to the device (data-out) while resulting data (data-in) is placed in the "out.bin" file. Also note the length of both is 512 bytes which corresponds to the transfer length of 1 (block) in the cdb (i.e. the second last byte). urandom.bin can be produced like this:
dd if=/dev/urandom bs=512 count=1 of=urandom.bin
- sg_raw.exe PhysicalDrive1 a1 0c 0e 00 00 00 00 00 00 e0 00 00
This example is from Windows and shows a ATA STANDBY IMMEDIATE command being sent to PhysicalDrive1. That ATA command is contained within the SCSI ATA PASS-THROUGH(12) command (see the SAT or SAT-2 standard at http://www.t10.org). Notice that the STANDBY IMMEDIATE command does not send or receive any additional data, however if it fails sense data should be returned and displayed.
The exit status of sg_raw is 0 when it is successful. Otherwise see the sg3_utils(8) man page.
Written by Ingo van Lil
Report bugs to <inguin at gmx dot de> or to <dgilbert at interlog dot com>.
Copyright © 2001-2021 Ingo van Lil
This software is distributed under the GPL version 2. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
sg_inq, sg_vpd, sg3_utils (sg3_utils), plscsi