SG_DD(8)

📦
version 1.49 · 6 versions
, in arch rolling, nixpkgs unstable

sg_dd(8) · SG3_UTILS · bestie edition, part 1 🎀

✨ NAME ✨

sg_dd — copies data to and from files and devices, and is especially that girl when it comes to SCSI devices 💅

🌸 SYNOPSIS 🌸

sg_dd [bs=BS] [conv=CONV] [count=COUNT] [ibs=BS] [if=IFILE]
[iflag=FLAGS] [obs=BS] [of=OFILE] [oflag=FLAGS] [seek=SEEK] [skip=SKIP]
[--help] [--verbose] [--version]

[blk_sgio={0|1}] [bpt=BPT] [cdbsz={6|10|12|16}] [cdl=CDL]
[coe={0|1|2|3}] [coe_limit=CL] [dio={0|1}] [grpnum=GN] [odir={0|1}]
[of2=OFILE2] [retries=RETR] [sync={0|1}] [time={0|1}[,TO]]
[verbose=VERB] [--dry-run] [--nocopy] [--progress] [--verify]

🦋 DESCRIPTION 🦋

okay so imagine dd(1) but with main character energy: sg_dd copies data to and from any files, and it's lowkey specialized for "files" that are actually Linux SCSI generic (sg) devices, raw devices, or any other device that supports the SG_IO ioctl (which, fun fact, you only find in the lk 2.6 series). The syntax and semantics are similar to the dd(1) command. Slay.

The first group up in the synopsis is the "standard" Unix dd(1) operands, and the second group is the extra options this utility brought to the party. Both groups get defined below, don't worry 🫶

When you give the --verify option, the read side stays exactly the same, but on the write side the WRITE SCSI command gets swapped out for the VERIFY SCSI command. And if any VERIFY command throws a sense key of MISCOMPARE, the verify operation just stops, periodt. The --verify option can only be used when OFILE is either an sg device or a block device that also has oflag=sgio given. When --verify is in play, this utility kinda works like the Unix cmp(1) command. iconic.

Real talk though: this utility is only supported on Linux, whereas most other utilities in the sg3_utils package have been ported to other operating systems. There's a utility called "ddpt" that has similar syntax and functionality to sg_dd — ddpt drops some Linux specific features while adding some other generic features, and that's literally what lets ddpt get ported to other operating systems. iykyk.

For support of NVMe storage devices, see the section below titled: NVME SUPPORT.

💖 OPTIONS 💖

  • blk_sgio={0|1} — when it's set to 0, block devices (e.g. /dev/sda) get treated like normal files, i.e. read(2) and write(2) are used for IO. When it's set to 1, block devices are assumed to accept the SG_IO ioctl and SCSI commands are issued for IO. This is only supported for 2.6 series kernels. Heads up bestie: ATAPI devices (e.g. cd/dvd players) use the SCSI command set but ATA disks do not (unless there's a protocol conversion, as often occurs in the USB mass storage class). If the input or output device is a block device partition (e.g. /dev/sda3) then setting this option causes the partition information to be ignored (since access is directly to the underlying device). Default is 0. See the 'sgio' flag 👀
  • bpt=BPT — each IO transaction will be made using BPT blocks (or less if you're near the end of the copy). Default is 128 for logical block sizes less than 2048 bytes, otherwise the default is 32. So for bs=512 the reads and writes will each convey 64 KiB of data by default (less if near the end of the transfer or memory restrictions). When cd/dvd drives are accessed, the logical block size is typically 2048 bytes and bpt defaults to 32 which again implies 64 KiB transfers. Ngl, the block layer when the blk_sgio=1 option is used has relatively low upper limits for transfer sizes (compared to sg device nodes, see /sys/block/<dev_name>/queue/max_sectors_kb).
  • bs=BS — where BS must be the logical block size of the physical device (if either the input or output files are accessed via SCSI commands). And yes this differs from dd(1), which permits BS to be an integral multiple. Default is 512, which is usually correct for disks but incorrect for cdroms (which normally have 2048 byte blocks). For this utility the maximum size of each individual IO operation is BS * BPT bytes.
  • cdbsz={6|10|12|16} — size of the SCSI READ and/or WRITE commands issued on sg device names (or block devices when 'iflag=sgio' and/or 'oflag=sgio' is given). Default is 10 byte SCSI command blocks (unless calculations indicate that a 4 byte block number may be exceeded or BPT is greater than 16 bits (65535), in which case it defaults to 16 byte SCSI commands).
  • cdl=CDL — allows setting of command duration limits. CDL is either a single value or two values separated by a comma. If one value is given, it applies to both IFILE and OFILE (if they are pass-through devices). If two values are given, the first applies to IFILE while the second applies to OFILE. The value may be from 0 to 7 where 0 is the default and means there are no command duration limits. Command duration limits are only supported by 16 byte READ and WRITE commands (plus READ(32), WRITE(32) and the WRITE SCATTERED command, bit they are used by this utility). If the cdbsz operand is not given and would have a value less than 16, then if CDL is greater than 0, the cdbsz is increased to 16. Command duration limits can be accesses and changed in the Command duration limit A and B mode pages, plus the Command duration limit T2A and T2B mode pages. The sdparm utility may be used to access and change these mode pages ✨
  • coe={0|1|2|3} — set to 1 or more for continue on error ('coe'). Only applies to errors on sg devices or block devices with the 'sgio' flag set. So errors on other files will stop sg_dd. Default is 0 which implies stop on any error. See the 'coe' flag for more information.
  • coe_limit=CL — where CL is the maximum number of consecutive bad blocks stepped over (due to "coe>0") on reads before the copy terminates. This only applies when IFILE is accessed via the SG_IO ioctl. The default is 0 which is interpreted as no limit. This option is meant to stop the copy soon after unrecorded media is detected while still offering "continue on error" capability. thoughtful queen 💜
  • conv=sparse — see the CONVERSIONS section below.
  • count=COUNT — copy COUNT blocks from IFILE to OFILE. Default is the minimum (of IFILE and OFILE) number of blocks that sg devices report from SCSI READ CAPACITY commands or that block devices (or their partitions) report. Normal files are not probed for their size. If skip=SKIP or seek=SEEK are given and the count is derived (i.e. not explicitly given) then the derived count is scaled back so that the copy will not overrun the device. If the file name is a block device partition and COUNT is not given then the size of the partition rather than the size of the whole device is used. If COUNT is not given (or count=-1) and cannot be derived then an error message is issued and no copy takes place. no copy, no drama 😭
  • dio={0|1} — default is 0 which selects indirect (buffered) IO on sg devices. A value of 1 attempts direct IO which, if not available, falls back to indirect IO and notes this at completion. If direct IO is selected and /sys/module/sg/parameters/allow_dio has the value of 0 then a warning is issued (and indirect IO is performed). For finer grain control use 'iflag=dio' or 'oflag=dio'.
  • grpnum=GN — the 'Group number' field is a 6 bit field (0 to 0x3f) found in all SCSI READ and WRITE commands (apart from the deprecated READ(6) and WRITE(6)). GN is either a single value or a pair of values separated by a comma. The default value is 0. If a pair of values is given, the first is used for the IFILE side (i.e. the SCSI READ) and the second is used for the OFILE side (i.e. the SCSI WRITE).

The group number is applied to non-passthrough devices using fcntl(F_SET_FILE_RW_HINT). For the OFILE side, oflag=direct is needed to defeat the write cache because F_SET_FILE_RW_HINT does not change the handling of the write cache (but fcntl(F_SET_RW_HINT) does change the write cache handling) 🧚

  • ibs=BS — if given, must be the same as BS given to the 'bs=' option.
  • if=IFILE — read from IFILE instead of stdin. If IFILE is '-' then stdin is read. Starts reading at the beginning of IFILE unless SKIP is given.
  • iflag=FLAGS — where FLAGS is a comma separated list of one or more flags outlined below. These flags are associated with IFILE and are ignored when IFILE is stdin.
  • obs=BS — if given, must be the same as BS given to the 'bs=' option.
  • odir={0|1} — when set to one, opens block devices (e.g. /dev/sda) with the O_DIRECT flag. User memory buffers are aligned to the page size when set. The default is 0 (i.e. the O_DIRECT flag is not used). Has no effect on sg, normal or raw files. If blk_sgio is also set then both are honoured: block devices are opened with the O_DIRECT flag and SCSI commands are issued via the SG_IO ioctl 💅
  • of=OFILE — write to OFILE instead of stdout. If OFILE is '-' then writes to stdout. If OFILE is /dev/null then no actual writes are performed. If OFILE is '.' (period) then it is treated the same way as /dev/null (this is a shorthand notation). If OFILE exists then it is not truncated; it is overwritten from the start of OFILE unless 'oflag=append' or SEEK is given.
  • of2=OFILE2 — write output to OFILE2. The default action is not to do this additional write (i.e. when this option is not given). OFILE2 is assumed to be a normal file or a fifo (i.e. a named pipe). OFILE2 is opened for writing, created if necessary, and closed at the end of the transfer. If OFILE2 is a fifo (named pipe) then some other command should be consuming that data (e.g. 'md5sum OFILE2'), otherwise this utility will block. deadlock is not the vibe 💀
  • oflag=FLAGS — where FLAGS is a comma separated list of one or more flags outlined below. These flags are associated with OFILE and are ignored when OFILE is /dev/null, '.' (period), or stdout.
  • retries=RETR — sometimes retries at the host are useful, for example when there is a transport error. When RETR is greater than zero then SCSI READs and WRITEs are retried on error, RETR times. Default value is zero.
  • seek=SEEK — start writing SEEK bs-sized blocks from the start of OFILE. Default is block 0 (i.e. start of file).
  • skip=SKIP — start reading SKIP bs-sized blocks from the start of IFILE. Default is block 0 (i.e. start of file).
  • sync={0|1} — when it’s 1, this does a SYNCHRONIZE CACHE command on OFILE at the end of the transfer. It’s only active when OFILE is an sg device file name or a block device and 'blk_sgio=1' is given. 💾
  • time={0|1}[,TO] — when it’s 1, it times the transfer and does throughput calculation, outputting the results (to stderr) at completion. When it’s 0 (the default), it doesn’t perform timing. ⏱️ If that value is followed by a comma, then TO is the command timeout in seconds for SCSI READ, WRITE or VERIFY commands issued by this utility. The default is 60 seconds.
  • verbose=VERB — as VERB increases, so does the amount of debug output sent to stderr. Default value is zero, which yields the minimum amount of debug output. A value of 1 reports extra information that is not repetitive. A value 2 reports cdbs and responses for SCSI commands that are not repetitive (i.e. other that READ and WRITE). Error processing is not considered repetitive. Values of 3 and 4 yield output for all SCSI commands (and Unix read() and write() calls) so there can be a lot of output. This only occurs for scsi generic (sg) devices and block devices when the 'blk_sgio=1' option is set. 👀
  • -d, --dry-run — does all the command line parsing and preparation but bypasses the actual copy or read. That preparation may include opening IFILE or OFILE to determine their lengths. This option may be useful for testing the syntax of complex command line invocations in advance of executing them. 🧪
  • -h, --help — outputs usage message and exits. 🫶
  • -n, --nocopy — this option causes this utility to exit when it would otherwise attempt to write to OFILE or stdout. It will not block an attempt to verify (i.e. using the --verify option) as in that case OFILE should not be modified. The long option variant may be abbreviated to --nc. The explicit way to instruct this utility to read only IFILE is to use '--of=/dev/null' in the invocation. This option may viewed as extra safety in case when the '--of=/dev/null' is not given or misspelt. 🛟
  • -p, --progress — this option causes a progress report to be output every two minutes until the copy is complete. After the copy is complete a line with "completed" is printed to distinguish the final report from the prior progress reports. When used twice the progress report is every minute, when used three times the progress report is every 30 seconds. If this option is given then the 'time=1' option is set implicitly. 📈
  • -v, --verbose — when used once, this is equivalent to verbose=1. When used twice (e.g. "-vv") this is equivalent to verbose=2, etc. 🗣️
  • -x, --verify — do a verify operation (like Unix command cmp(1)) rather than a copy. Cannot be used with "oflag=sparse". of=OFILE must be given and OFILE must be an sg device or a block device with "oflag=sgio" also given. Uses the SCSI VERIFY command with the BYTCHK field set to 1. The VERIFY command is used instead of WRITE when this option is given. There is no VERIFY(6) command. Stops on the first miscompare unless oflag=coe is given. 🔍
bestie, that verify setup is strict on purpose — no shady mismatches just vibing through. It stops on the first miscompare unless oflag=coe is given. 😭
  • -V, --version — outputs version number information and exits. 👑

🦋 CONVERSIONS 🦋

One or more conversions can be given to the "conv=" option. If more than one is given, they should be comma separated. sg_dd does not perform the traditional dd conversions (e.g. ASCII to EBCDIC). Recently added conversions overlap somewhat with the flags so some conversions are now supported by sg_dd. ✨

  • nocreat — this conversion has the same effect as "oflag=nocreat", namely: OFILE must exist, it will not be created. 🚫
  • noerror — this conversion is very close to "iflag=coe" and is treated as such. See the "coe" flag. Note that an error on OFILE will stop the copy. ⚠️
  • notrunc — this conversion is accepted for compatibility with dd and ignored since the default action of this utility is not to truncate OFILE. 🎀
  • null — has no affect, just a placeholder. 😌
  • sparse — FreeBSD supports "conv=sparse" so the same syntax is supported in sg_dd. See "sparse" in the FLAGS sections for more information. 🌸
  • sync — is ignored by sg_dd. With dd it means supply zero fill (rather than skip) and is typically used like this "conv=noerror,sync" to have the same functionality as sg_dd's "iflag=coe". 💅

✨ FLAGS ✨

okay so here's the flag lore, bestie — every single one, no skips, no cap. 💅

  • 00 — this one is only active with iflag=, and when you give it, it replaces if=IFILE. If you give both, you get an error, periodt. The input will be a stream of zeros, similar to using if=/dev/zero alone (but a little quicker), apart from the following case. If iflag=00,ff is given, then the block address (lower 32 bits, in 4 bytes, big endian) is placed, multiple times, in each block. The block address takes into account the skip=SKIP setting. The sgp_dd utility has a --chkaddr option that complements this option. ✨
  • append — causes the O_APPEND flag to be added to the open of OFILE. For regular files this will lead to data appended to the end of any existing data. Cannot be used together with the seek=SEEK option as they conflict. The default action of this utility is to overwrite any existing data from the beginning of the file or, if SEEK is given, starting at block SEEK. Note that attempting to “append” to a device file (e.g. a disk) will usually be ignored or may cause an error to be reported. 🫶
  • coe — continue on error. Only active for sg devices and block devices that have the sgio flag set. iflag=coe oflag=coe and coe=1 are equivalent. Use this flag twice (e.g. iflag=coe,coe) to have the same action as the coe=2. A medium, hardware or blank check error while reading will re-read blocks prior to the bad block, then try to recover the bad block, supplying zeros if that fails, and finally re-read the blocks after the bad block. A medium, hardware or blank check error while writing is noted and ignored. A miscompare sense key during a VERIFY command (i.e. --verify given) is noted and ignored when oflag=coe. The recovery of the bad block when reading uses the SCSI READ LONG command if coe given twice or more (also with the command line option coe=2). Further, the READ LONG will set its CORRCT bit if coe given thrice. SCSI disks may automatically try and remap faulty sectors (see the AWRE and ARRE in the read write error recovery mode page (the sdparm utility can access and possibly change these attributes)). Errors occurring on other file types will stop sg_dd. Error messages are sent to stderr. This flag is similar to conv=noerror,sync in the dd(1) utility. See note about READ LONG below. 😭
  • dio — requests that the sg device node associated with this flag does direct IO. If direct IO is not available, falls back to indirect IO and notes this at completion. If direct IO is selected and /sys/module/sg/parameters/allow_dio has the value of 0 then a warning is issued (and indirect IO is performed). 🔥
  • direct — causes the O_DIRECT flag to be added to the open of IFILE and/or OFILE. This flag requires some memory alignment on IO. Hence user memory buffers are aligned to the page size. Has no effect on sg, normal or raw files. If iflag=sgio and/or oflag=sgio is also set then both are honoured: block devices are opened with the O_DIRECT flag and SCSI commands are issued via the SG_IO ioctl. 💅
  • dpo — set the DPO bit (disable page out) in SCSI READ and WRITE commands. Not supported for 6 byte cdb variants of READ and WRITE. Indicates that data is unlikely to be required to stay in device (e.g. disk) cache. May speed media copy and/or cause a media copy to have less impact on other device users. 👑
  • dsync — causes the O_SYNC flag to be added to the open of IFILE and/or OFILE. The d is prepended to lower confusion with the sync=0|1 option which has another action (i.e. a synchronisation to media at the end of the transfer). 🌈
  • excl — causes the O_EXCL flag to be added to the open of IFILE and/or OFILE. 🎀
  • ff — this flag is only active with iflag= and when given replaces if=IFILE. If both are given an error is generated. The input will be a stream of 0xff bytes (or all bits set), apart from the following case. If iflag=00,ff is given then the block address (lower 32 bits, in 4 bytes, big endian) is placed, multiple times, in each block. The block address takes into account the skip=SKIP setting. ⭐
  • flock — after opening the associated file (i.e. IFILE and/or OFILE) an attempt is made to get an advisory exclusive lock with the flock() system call. The flock arguments are "FLOCK_EX | FLOCK_NB" which will cause the lock to be taken if available else a “temporarily unavailable” error is generated. An exit status of 90 is produced in the latter case and no copy is done. 🙈
  • fua — causes the FUA (force unit access) bit to be set in SCSI READ and/or WRITE commands. This only has an effect with sg devices or block devices that have the sgio flag set. The 6 byte variants of the SCSI READ and WRITE commands do not support the FUA bit. 🌸
  • nocache — use posix_fadvise() to advise corresponding file there is no need to fill the file buffer with recently read or written blocks. 🧚
  • nocreat — this flag is only active in oflag=FLAGS. If present then OFILE will be opened if it exists. If OFILE doesn't exist then an error is generated. Without this flag a regular (empty) file named OFILE will be created (and then filled). For production quality scripts where OFILE is a device node (e.g. /dev/sdc) this flag is recommended. It guards against the remote possibility of dev/sdc disappearing temporarily (e.g. a USB memory key removed) resulting in a large regular file called /dev/sdc being created. 💀
  • null — has no affect, just a placeholder. 🍓
  • pt — has the same meaning as the sgio flag. Added for compatibility with the ddpt utility. 🦋
  • random — this flag is only active with iflag= and when given replaces if=IFILE. If both are given an error is generated. The input will be a stream of pseudo random bytes. The Linux getrandom(2) system call is used to create a seed and thereafter mrand48(3) is used to generate a pseudo random sequence, 4 bytes at a time. The quality of the randomness can be viewed with the ent(1) utility. This is not a high quality random number generator, it is built for speed, not quality. One application is checking the correctness of the copy and verify operations of this utility. ✨
  • sgio — causes block devices to be accessed via the SG_IO ioctl rather than standard UNIX read() and write() commands. When the SG_IO ioctl is used the SCSI READ and WRITE commands are used directly to move data. sg devices always use the SG_IO ioctl. This flag offers finer grain control compared to the otherwise identical blk_sgio=1 option. 💜
  • sparse — after each BS * BPT byte segment is read from the input, it is checked for being all zeros. If so, nothing is written to the output file unless this is the last segment of the transfer. This flag is only active with the oflag option. It cannot be used when the output is not seekable (e.g. stdout). It is ignored if the output file is /dev/null. Note that this utility does not remove the OFILE prior to starting to write to it. Hence it may be advantageous to manually remove the OFILE if it is large prior to using oflag=sparse. The last segment is always written so regular files will show the same length and so programs like md5sum and sha1sum will generate the same value regardless of whether oflag=sparse is given or not. This option may be used when the OFILE is a raw device but is probably only useful if the device is known to contain zeros (e.g. a SCSI disk after a FORMAT command). 😌

🌸 RETIRED OPTIONS 🌸

here are some retired options that are still present — yes, they're still here, living their best life. 🦋

  • append=0 | 1 — when set, equivalent to oflag=append. When clear the action is to overwrite the existing file (if it exists); this is the default. See the append flag.
  • fua=0 | 1 | 2 | 3 — force unit access bit. When 3, fua is set on both IFILE and OFILE; when 2, fua is set on IFILE;, when 1, fua is set on OFILE; when 0 (default), fua is cleared on both. See the fua flag. 💖

🗒️ NOTES 🗒️

okay bestie, real talk about block devices 💅 — you can hand /dev/sda and /dev/hda (and friends) to IFILE. If none of -iflag=direct, iflag=sgio, or blk_sgio=1 is given, then normal block IO involving buffering and caching is performed. If only -iflag=direct is given, then the buffering and caching is bypassed — and yes, that applies to both SCSI devices and ATA disks. If iflag=sgio or blk_sgio=1 is given, then the SG_IO ioctl is used on the given file, causing SCSI commands to be sent to the device, and that also bypasses most of the actions performed by the block layer — but this is only applicable to SCSI devices, not ATA disks. The same applies for block devices given for OFILE. 🫶

Various numeric arguments (e.g. SKIP) may include multiplicative suffixes or be given in hexadecimal. Go peek at the "NUMERIC ARGUMENTS" section in the sg3_utils(8) man page for that tea. ☕

The COUNT, SKIP, and SEEK arguments can take 64 bit values (i.e. very big numbers — main character energy 🔥). Other values are limited to what can fit in a signed 32 bit number.

Data usually gets to user space in a 2 stage process: first the SCSI adapter DMAs into kernel buffers, and then the sg driver copies this data into user memory (write operations reverse this sequence). This is called "indirect IO", and there is a 'dio' option to select "direct IO", which will DMA directly into user memory. Due to some issues, "direct IO" is disabled in the sg driver and needs a configuration change to activate it. This is typically done with:

echo 1 > /sys/module/sg/parameters/allow_dio

All informative, warning, and error output is sent to stderr so that dd's output file can be stdout and remain unpolluted. If no options are given, then the usage message is output and nothing else happens. 🙈

Even if READ LONG succeeds on a “bad” block when 'coe=2' (or 'coe=3') is given, the recovered data may not be useful. There are no guarantees that the user data will appear "as is" in the first 512 bytes. 😭

A raw device must be bound to a block device prior to using sg_dd. See raw(8) for more information about binding raw devices. To be safe, the sg device mapping to SCSI block devices should be checked with sg_map before use. 🌸

Disk partition information can often be found with fdisk(8) [the "-ul" argument is useful in this respect].

For sg devices (and block devices when blk_sgio=1 is given) this utility issues SCSI READ and WRITE (SBC) commands which are appropriate for disks and reading from CD/DVD/HD-DVD/BD drives. Those commands are not formatted correctly for tape devices, so sg_dd should not be used on tape devices. If the largest block address of the requested transfer exceeds a 32 bit block number (i.e. 0xffff), then a warning is issued and the sg device is accessed via SCSI READ(16) and WRITE(16) commands. 🎀

The attributes of a block device (partition) are ignored when 'blk_sgio=1' is used. Hence the whole device is read (rather than just the second partition) by this invocation:

sg_dd if=/dev/sdb2 blk_sgio=1 of=t bs=512

🧚 NVME SUPPORT 🧚

Some support for copying from and to NVMe devices in Linux has been added. There are two varieties of NVME “char” devices, examples: /dev/nvme<cid> and /dev/ng<cid>n<nsid>, where <cid> is the controller identifier (starting at 0) and <nsid> is the NVMe namespace identifier (starting at 1). The latter form is called "nvme-generic" in /proc/devices. The NVMe block devices have the form: /dev/nvme<cid>n<nsid>[p<part_id>], where the optional <part_id> is the partition identifier (starting at 1). In the case of NVMe block devices, the default action of sg_dd is to access them using standard Unix IO (i.e. read(2) or write(2)). To access them using SCSI/NVMe commands, the flag "sgio" or "pt" needs to be given (e.g. iflag=pt if=/dev/nvme0n1). ✨

🚨 Just as with partitions on SCSI disks, it is dangerous to access NVMe partitions (e.g. iflag=pt if=/dev/nvme0n1p2) using a pass-through interface. The reason is that both SCSI and NVMe commands are oblivious to any partitioning arrangement which Linux (or any other OS) has set up on those disks. In other words, in the above example the "p2" at the end of if=/dev/nvme0n1p2 is ignored when a pass-through is being used. 💀

The READ and WRITE commands in NVMe are in the "NVM" command set, which requires an NVMe namespace to be specified. This means that the char device /dev/nvme0 is not useful for the utility — only administrative commands can be sent to it. That leaves the /dev/ng<cid>n<nsid> as the only NVMe “char” device that can be used with this utility. Since it is not a block device, the standard Unix read(2) and write(2) will not work with it; so the pass-through interface, issuing SCSI/NVME commands, is assumed by this utility. 🌈

The term "SCSI/NVME command" refers to the fact that this utility issues SCSI commands and a lower level, within libsgutils, converts them to the corresponding NVMe commands. This action is known as a SCSI to NVMe Translation Layer (SNTL), and the NVMe consortium wrote a white paper on it early in the development of NVMe. 👑

🌟 EXAMPLES 🌟

Looks quite similar in usage to dd:

sg_dd if=/dev/sg0 of=t bs=512 count=1MB

This will copy 1 million 512 byte blocks from the device associated with /dev/sg0 (which should have 512 byte blocks) to a file called t. Assuming /dev/sda and /dev/sg0 are the same device, then the above is equivalent to:

dd if=/dev/sda iflag=direct of=t bs=512 count=1000000

although dd's speed may improve if bs was larger and count was suitably reduced. The use of the 'iflag=direct' option bypasses the buffering and caching that is usually done on a block device. 💨

Using a raw device to do something similar on an ATA disk:

raw /dev/raw/raw1 /dev/hda
sg_dd if=/dev/raw/raw1 of=t bs=512 count=1MB

To copy a SCSI disk partition to an ATA disk partition:

raw /dev/raw/raw2 /dev/hda3
sg_dd if=/dev/sg0 skip=10123456 of=/dev/raw/raw2 bs=512

This assumes a valid partition is found on the SCSI disk at the given skip block address (past the 5 GB point of that disk), and that the partition goes to the end of the SCSI disk. An explicit count is probably a safer option. The partition is copied to /dev/hda3, which is an offset into the ATA disk /dev/hda. The exact number of blocks read from /dev/sg0 are written to /dev/hda (i.e. no padding). 🍓

To time a streaming read of the first 1 GB (2 ** 30 bytes) on a disk, this utility could be used:

sg_dd if=/dev/sg0 of=/dev/null bs=512 count=2m time=1

On completion, this will output a line like: "time to transfer data was 18.779506 secs, 57.18 MB/sec". The "MB/sec" in this case is 1,000,000 bytes per second. ⏱️

The 'of2=' option can be used to copy data and take a md5sum of it without needing to re-read the data:

mkfifo fif
md5sum fif &
sg_dd if=/dev/sg3 iflag=coe of=sg3.img oflag=sparse of2=fif bs=512

This will image /dev/sg3 (e.g. an unmounted disk) and place the contents in the (sparse) file sg3.img . Without re-reading the data, it will also perform a md5sum calculation on the image. 🧮

📡 SIGNALS 📡

The signal handling has been borrowed from dd:

  • SIGINT, SIGQUIT, and SIGPIPE — output the number of remaining blocks to be transferred and the records in + out counts; then they have their default action.
  • SIGUSR1 — causes the same information to be output yet the copy continues.

All output caused by signals is sent to stderr. 🫡

🚪 EXIT STATUS 🚪

The exit status of sg_dd is 0 when it is successful. Otherwise, see the sg3_utils(8) man page. Since this utility works at a higher level than individual commands, and there are 'coe' and 'retries' flags, individual SCSI command failures do not necessarily cause the process to exit.

  • 90 — an additional exit status of 90 is generated if the flock flag is given and some other process holds the advisory exclusive lock. 🔒

✍️ AUTHORS ✍️

Written by Douglas Gilbert and Peter Allworth. 👯

🐛 REPORTING BUGS 🐛

Report bugs to <dgilbert at interlog dot com>. 📬

Copyright © 2000-2026 Douglas Gilbert.

This software is distributed under the GPL version 2. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. ⚖️

👀 SEE ALSO 👀

  • cmp(1)
  • There is a web page discussing sg_dd at https://sg.danny.cz/sg/sg_dd.html
  • A POSIX threads version of this utility called sgp_dd is in the sg3_utils package. Another version from that package is called sgm_dd, and it uses memory mapped IO to speed transfers from sg devices.
  • The lmbench package contains lmdd, which is also interesting. For moving data to and from tapes, see dt, which is found at https://www.scsifaq.org/RMiller_Tools/index.html
  • To change mode parameters that effect a SCSI device's caching and error recovery, see sdparm(sdparm)
  • To verify the data on the media or to verify it against some other copy of the data, see sg_verify(sg3_utils)
  • See also raw(8), dd(1), ddrescue(GNU), ddpt

That’s the whole vibe, bestie — catch you in the sg3_utils-1.49 era, April 2026, on the SG_DD(8) page 💖✨

📦 where this page comes from