xref: /aosp_15_r20/external/sg3_utils/README.details (revision 44704f698541f6367e81f991ef8bb54ccbf3fc18)

                      README.details for sg3_utils
                      ============================
Introduction
============
This package contains low level command line utilities for devices that use
the SCSI command set. Originally the SCSI command set was associated
exclusively with the SCSI Parallel Interface (SPI) transport. SPI has now
almost been completely replaced by the Serial Attached SCSI (SAS) transport
which also accepts the SCSI command set. Additionally many other storage
related transports use the SCSI command set (amongst others); examples are
ATAPI devices (CD/DVDs and tapes), USB mass storage devices (including those
using the newer UAS[P]), Fibre Channel disks, IEEE 1394 storage devices (SBP
protocol), iSCSI, FCoE and SOP devices. Even NVMe which has its own command
set accepts SCSI commands in some contexts; one example is for enclosure
management where NVME-MI has SES Send and SES Receive commands. SES refers
to the SCSI Enclosure Services command set.

This package originally targeted the Linux SCSI subsystem. Since most
operating systems contain a SCSI command pass-through mechanism, many
utilities within this package have been ported. This README mainly
concentrates on Linux: see the README.freebsd file for the FreeBSD port,
README.solaris for the Solaris port, the README.tru64 file for the Tru64
(OSF) port and README.win32 for the Windows ports (of which there are two
variants).

Most utilities within the sg3_utils package work at the SCSI command level.
For example the sg_inq utility issues a SCSI INQUIRY command and decodes the
response. The COVERAGE file has a table containing a row for each SCSI
command issued by this package; to the right of each row is the utility
(sometimes more than one) that issue that SCSI command. The COVERAGE file
has a second table for ATA commands usage.

Some utilities interface at a slightly higher level, for example: sg_dd,
sgm_dd and sgp_dd. These are closely related to the Unix dd command and
typically issue a sequence of SCSI READ and WRITE commands to copy data.
These utilities are relatively tightly bound to Linux and are not ported to
other Operating Systems. A new utility called ddpt (in a package of the same
name) is more generic while still allowing a copy to be done in terms of
SCSI READ and WRITE commands. ddpt has been ported to other OSes.

License
=======
All utilities and libraries have either a "2 clause" BSD license or are
"GPL-2ed". The "2 clause" BSD license is taken from the FreeBSD project but
drops the last paragraph that directly refers to the "FreeBSD project".
That BSD license was updated from the "3 clause" to the newer "2 clause"
version on 20180119. To save space various source code files refer to a
file called "BSD_LICENSE" in the main, src and lib directories. The author's
intention is that users may incorporate all or part of the code in their work
as they please. Attribution is encouraged. Please check the code as other
contributors (apart from the author) may also have copyright notices. For a
list of contributors see the CREDITS file.


Description
===========
A web site supporting the sg3_utils package can be found at
https://sg.danny.cz/sg/sg3_utils.html . That page has a table of released
versions for download. The most recent release or beta of sg3_utils may
be found on this page: https://sg.danny.cz/sg in the News section.

The predecessor to this package was called sg_utils. It is described in
https://sg.danny.cz/sg/uu_index.html and old versions can be downloaded
from the Downloads section of https://sg.danny.cz/sg .

In the Linux 2.4 kernel series these utilities need to use the SCSI generic
(sg) driver to access SCSI devices. The name of this package (i.e. sg3_utils)
refers to version 3 of the SCSI generic (sg) driver which was introduced at
the beginning of the 2.4 Linux kernel series. Significantly this added a new
SCSI command interface structure (i.e. struct sg_io_hdr) that is more
flexible than the older "sg_header" structure found in the sg driver in the
2.2 and earlier Linux kernel series. The sg_io_hdr structure is also more
flexible than the awkward (and limiting) interface to the
SCSI_IOCTL_SEND_COMMAND ioctl supported by the Linux SCSI mid level. The
version 3 sg driver also added the SG_IO ioctl that is synchronous (i.e. it
issues the requested SCSI command and waits for the response (or a timeout)
before the ioctl returns to the user space program that invoked it). The
SG_IO ioctl is now supported in other parts of the Linux kernel in the 2.6
series.

In sg3_utils version 1.27 support has been added for the Linux bsg driver
which use the sg version 4 interface. There seems no point in renaming
this package sg4_utils. The existing utilities just silently support either.
Currently the source build must be able to see the /usr/include/linux/bsg.h
file. Then at run time the /proc/devices pseudo file needs to have an entry
for the bsg driver (appeared around lk 2.6.28). With this in place each
utility at run time checks the device it has been given and if it is a char
device whose major number matches the bsg entry in /proc/devices then the
sg v4 interface is used. Otherwise the sg v3 interface is used.

Utilities that wish to use the asynchronous SCSI command interface (i.e. via
a write() read() sequence) or issue special "commands" (e.g. bus and device
resets) still need to use the Linux sg driver. Note that various
drivers (e.g. cdrom/sr) have different open() flag and permissions policies
that the user may need to take into account.

If users have problems or questions about them please contact the author.
Documentation for the Linux sg device driver can be found at:
https://sg.danny.cz/sg/p/sg_v3_ho.html . This is written in DocBook and the
original xml can be found in the same directory with the ".xml" extension.
Postscript and pdf renderings are also in that directory. Older documentation
for the sg version 3 driver can be found at:
https://sg.danny.cz/sg/p/scsi_generic_v3.txt .

To save the repetition of common code (e.g. SCSI error processing) and
reduce the size of the executable files, a shared library called
libsgutils<num>.so (its Linux name) is created during the build process.
That library is built from the contents of the include and lib
subdirectories. The header files in the include subdirectory can be seen
as the API of libsgutils and are commented with that in mind. The SCSI
pass-through code for the supported operating systems is found in the lib
subdirectory with names like sg_pt_linux.c and sg_pt_win32.c .

Various distributions (of Linux mainly) distribute sg3_utils as 3
installable packages. One is a package containing the shared library
discussed above (e.g. libsgutils2-2_1.33-0.1_i386.deb). A second package
contains the utilities (e.g. sg3-utils_1.33-0.1_i386.deb) and depends on the
first package). Finally there is an optional package that contains header
files and a static library (e.g. libsgutils2-dev_1.33-0.1_i386.deb). This
final package is only needed to build other packages (e.g. sdparm) that
wish to use the sg3_utils shared library.

All the utilities in the src subdirectory have "man" pages that are
placed in the doc subdirectory. There is also a sg3_utils (8) man page that
summarizes common facilities including exit statuses. Additional
information (including each utility's version number) can be found towards
the top of each ".c" file corresponding to the utility name.

The sg driver in Linux can be seen as having 3 distinct versions:

   v1   lk < 2.2.6     sg_header based relatively unchanged since 1992
   v2   lk >= 2.2.6    enhanced sg_header interface structure [1999/4/16]
   v3   lk >= 2.4      additional sg_io_hdr interface structure [2001/1/4]
   v3   lk >= 2.6      same interface as found in lk 2.4 [2.6.0: 2003/12/18]

and the bsg driver supports the sg v4 interface and was added around
lk 2.6.28 . This package is targeted at "v3" and "v4". Another package called
"sg_utils" is targeted at "v2" and to a lesser extent "v1". The "sg_utils"
package has a subset of the utilities found in this package.

In Linux some sg driver ioctls (notably SG_IO) are defined for many block
devices in lk 2.6 series. In practice this means all SCSI block devices,
ATAPI block devices (mainly CD, DVD and BD optical devices) but _not_ ATA
disks, depending on which kernel configuration options, can be accessed by
the utilities in this package. SATA disks that use the libata kernel library
(or some other SCSI to ATA Translation (SAT) Layer (SATL)) accept SCSI
commands and thus are supported. Support for the SG_IO as been added to the
scsi tape driver (st) in lk 2.6.6 .

In the src directory the bulk of the utilities are written in relatively
clean POSIX compliant C code with Linux specific system calls and structures
removed and placed in Linux specific files in the lib directory. A small
number of utilities in the src directory do contain Linux specific logic
and are not ported to other OSes (e.g. sg_dd). One utility, sg_scan, has
two separate implementations, one for Linux (sg_scan_linux.c) and one for
Windows (sg_scan_win32.c). The src-lib directory split approach allows
FreeBSD, Solaris, Tru64 and Windows specific code to be isolated to a few
files in the lib directory whose interfaces match those of the Linux
specific code.

Darwin is not supported because the Apple folks do not want to give their
users a pass-through SCSI interface. The author has read about creative
hackers using a VM containing a real OS to circumvent the Apple restriction.

C standard is C11
==================
The C code in this package is written for portability rather than speed.
It assumes a level of C99 compliance (the C standard prior to C11) and
favours POSIX system and library calls over OS specific calls.

The C code is written in a C++ friendly way and is checked from time to
time that it compiles clean with C++. To accommodate C++ certain C99
constructs such as designated initializers cannot be used. To build
with C++, C++11 (i.e. the C++ standard from 2011) or later is required.
Finding a common C and C++ syntax for zeroing stack variables (including
aggregates) may need to wait until C23 allows this syntax:
   struct example_t ex1 {};
which C++ introduced in C++11.  In the meantime the SG_C_CPP_ZERO_INIT
define (hack) does this.

The author has not seriously attempted to build this code on MSVC (aka
Visual Studio). There are a few roadblocks (that may be overcome in the
future) that include MSVC being basically a C++ compiler, not a C/C++
compiler. For some reason MSVC only claims C89 compliance (i.e. the first
C standard from 1989). MSVC 2013 and 2015 are moving closer to C99
compliance and may be sufficient to compile this package. Another problem
is the assumption of the availability of basic Unix system calls such as
open(). Nearly 20 years ago Microsoft indicated (promised ?) that it
would move in the direction of POSIX compliance, but very little ever
happened. "Talk is cheap, there should be a tax on it."

Building
========
This package is designed to be built with the usual:
    "./configure ; make ; make install"
sequence. In some situations that may need to be prefixed by a call to
the "./autogen.sh" script which invokes autoconf and automake. That in turn
may require packages containing those utilities to be installed. The
libtool utility is also required. Naturally a C compiler is required
and due to the vagaries of libtool a C++ compiler also.

The "./configure" takes many command line options with the defaults
being usually sufficient to start with. One quirk is that the location
of the installation is under the /usr/local directory. So the sg_inq
utility will be installed at /usr/local/bin/sg_inq . This is controlled
by the "--prefix=<directory>" option which defaults to
"--prefix=/usr/local". As an example to install the executables in /usr/bin
and disable the creation of the shared library (libsgutils<num>.so) this
invocation could be used: "./configure --prefix=/usr --disable-shared".
To reduce the size of an executable as well try this:
"./configure --prefix=/usr --disable-shared --disable-scsistrings".
Also --disable-shared will produce (relatively) "static" executables in
the src directory that are easier to debug. And
"./configure --enable-debug" will compile with more debug type options,
including more compiler checks and defining "DEBUG" within the src and
lib source files. Most utilities in the src directory set '-vv' (i.e.
equivalent to calling "--verbose" twice) when "DEBUG" is set.

In Linux there are package build files for "rpm" based and for "deb" based
systems. The 'sg3_utils.spec' file in the main directory can be used like
this: 'rpmbuild -ba sg3_utils.spec' in a rpmbuild tree SPECS directory.
To cross build or make a more widely distributable package then the --target
option may be useful: 'rpmbuild --target=i386 -ba sg3_utils.spec' or
'rpmbuild --target=x86_64 -ba sg3_utils.spec' . The sg3_utils.spec file
in the main directory targets Red Hat systems, an alternative "spec" file
for Suse systems has been placed under the 'suse' directory.

The 'build_debian.sh' script should build several "deb" packages and place
them in the parent directory. In debian based systems doing
a 'apt-get install build-essential' is one way to get most of build
environment needed if it has not already been loaded. There are now some
problems with this script and the superseded Debian 4.0 ("etch"). See
debian/README.debian4 for a workaround. Amongst other things debian
builds are sensitive to the value in the debian/compat file. If it
contains "7" then it works on lenny and gives warning on squeeze (but
fails on the earlier etch).

Warning
=======
Many devices use SCSI command sets over transport protocols not normally
associated with SCSI (as defined at https://www.t10.org ). Some of these
devices react poorly (e.g. lock up) when sent SCSI commands that they don't
support. Even sending a supported SCSI command with a field set to an
unexpected value can cause problems. [The author is talking about billions
of USB devices with horrible SCSI implementations.]

For example, all "SCSI" devices must support the INQUIRY command which the
SCSI-2 standard says should request a 36 byte response. However later SCSI
standards (e.g. SPC-2) have increased that length but some SCSI devices lock
up when they receive a request for anything other than a 36 byte response.

Any well implemented "SCSI" device should react sensibly when a utility in
sg3_utils sends a SCSI command that it doesn't support. Unfortunately this
cannot be guaranteed.

Prior to lk 2.6.29 USB mass storage limited sense data to 18 bytes which
caused problems for certain types of descriptor based sense data. An
example of this is the SCSI ATA PASS-THROUGH command with the CK_COND bit
set.


Utilities
=========
Here is list in alphabetical order of utilities found in the 'src'
subdirectory of the sg3_utils package:
    sginfo, sg_bt_ctl, sg_compare_and_write, sg_copy_results, sgm_dd, sgp_dd,
    sg_dd, sg_decode_sense, sg_emc_trespass, sg_format, sg_get_config,
    sg_get_elem_status, sg_get_lba_status, sg_ident, sg_inq, sg_logs,
    sg_luns, sg_map, sg_map26, sg_modes, sg_opcodes, sg_persist, sg_prevent,
    sg_raw, sg_rbuf, sg_rdac, sg_read, sg_read_attr, sg_readcap,
    sg_read_block_limits, sg_read_buffer, sg_read_long, sg_reassign,
    sg_referrals, sg_rem_rest_elem, sg_rep_density, sg_rep_pip, sg_rep_zones,
    sg_request, sg_reset, sg_rmsn, sg_rtpg, sg_safte, sg_sanitize,
    sg_sat_identify, sg_sat_phy_event, sg_sat_read_gplog, sg_sat_set_features,
    sg_scan, sg_seek, sg_senddiag, sg_ses, sg_ses_microcode, sg_start,
    sg_stpg, sg_stream_ctl, sg_sync, sg_test_rwbuff, sg_timestamp, sg_turs,
    sg_unmap, sg_verify, sg_vpd, sg_write_buffer, sg_write_long,
    sg_write_same, sg_write_verify, sg_write_x, sg_wr_mode, sg_xcopy, sg_zone,
    sg_z_act_query

Each of the above utilities depends on header files found in the 'include'
subdirectory and library code found in the 'lib' subdirectory. Associated
man pages are found in the 'doc' subdirectory. Additional programs found
in the 'archive', 'examples' and 'utils' subdirectories in not build by the
top level build infrastructure. Linux binary distributions of the sg3_utils
package (e.g. "rpm" and debian packages) typically contain the shared
library, the utilities found in the 'src' subdirectory, their associated man
pages and some documentation files (e.g. README, INSTALL, CREDITS, COPYING
and COVERAGE). See the INSTALL file for generic instructions about building
with autotools (e.g. ./configure ).

Man pages can be read (without building and installing the package) by
going to the 'doc' subdirectory and executing something like this:
 $ man ./sg_dd.8

To see which SCSI commands (and ATA commands) are used by these utilities
refer to the COVERAGE file.

Here is a list in alphabetical order of utilities found in the 'examples'
subdirectory:
  - sg_excl, scsi_inquiry, sg_sat_chk_power, sg__sat_identify,
    sg__sat_phy_event, sg__sat_set_features, sg_sat_smart_rd_data,
    sg_simple1, sg_simple2, sg_simple3, sg_simple4, sg_simple5,
    sg_simple16

Also in that subdirectory is a script to test sg_persist, an example data
file for sg_persist (called "transport_ids.txt") and an example data file for
sg_reassign (called "reassign_addr.txt"). There are several scripts
for 'sg_senddiag -pf -raw=-' that will put some SAS disk phys into
a "compliant jitter tolerance pattern" (CJTPAT).

The 'testing' subdirectory contains source and a Makefiles to test
kernel pass-through and associated drivers, mainly for Linux. There is
both C code (with the extension ".c") and C++ code (with the extension
".cpp"). There is a "Makefile" to build the C + C++ code. The Makefile
depends on some object files from the "lib" subdirectory. So a sequence
like this may be required prior to invoking make: "cd <top_of_package> ;
./configure ; cd lib ; make ; cd ../testing".

Here is a list in alphabetical order of utilities found in the 'testing'
subdirectory:
  - bsg_queue_tst, sgh_dd (C++), sg_iovec_tst, sg_queue_tst, sg_sense_tst,
    sg_tst_async (C++), sg_tst_context (C++), sg_tst_excl (C++),
    sg_tst_excl2 (C++), sg_tst_excl3 (C++)

The 'utils' subdirectory contains source and a Makefile to build "hxascdmp"
which accepts binary data from stdin (or a file on the command line) and
outputs an ASCII-HEX and ASCII representation of it. It is similar to the
Unix od command. There is also code to sg_chk_asc.c which checks a given
text file (typically a copy of https://www.t10.org/lists/asc-num.txt ) and
checks it against the asc/ascq text strings held in sg_lib_data.c .

The 'doc' subdirectory contains a README file containing the urls of
various related documents.

The 'scripts' subdirectory contains some Bourne (bash) shell scripts that
rely on utilities in the main directory. One script uses the sdparm utility.
These scripts are described in the scripts/README file and have usage
messages.


Notes for utilities without man pages
=====================================
These utils are found in the 'examples' subdirectory.

The "scsi_inquiry" program shows the use of the SCSI_IOCTL_SEND_COMMAND
ioctl to send a SCSI INQUIRY command. That ioctl() is supported by the
SCSI sub system mid level and so is common to all sd, sr, st and sg devices.
That ioctl is deprecated in the lk 2.6 series. This program has been placed
in the "examples" subdirectory.

"sg_simple1" and "sg_simple2" are example programs demonstrating calls
to the SCSI INQUIRY and TEST UNIT READY commands. They only differ in their
error processing: sg_simple1 uses sg_lib.[hc] for error processing while
sg_simple2 does its own more primitive checks.

"sg_simple3" tests out user space scatter gather added to the version 3
sg driver.

"sg_simple4" shows the INQUIRY command using mmap-ed IO to obtain its
response buffer.

"sg_simple5" also sends and INQUIRY and TEST UNIT READY commands. It
uses the generic pass through mechanism based on sg_pt.h . It will
currently build in Linux and FreeBSD (with "make -f Makefile.freebsd").
It has extensive error checking code.

"sg_simple16" attempts to send a 16 byte SCSI command, READ_16, to the
scsi device. This is only supported for lk >= 2.4.15 and for adapter
drivers that indicate that they have 16 byte CDB capability (otherwise
DID_ABORT will appear in the host_status).

"sg_sat_chk_power" attempts to push an ATA CHECK POWER MODE command
through the SAT-defined ATA PASS_THROUGH (16) SCSI command. That
ATA command needs to read the "FIS" registers after the command is
completed which involves using the ATA Status Return (sense data)
descriptor (as defined in SAT).

"sg_sat_smart_rd_data" attempts to push an ATA SMART/READ DATA command
through the SAT-defined ATA PASS_THROUGH (16) SCSI command. If
successful, the 256 word (512 byte) response is output.

"sg_tst_excl" and "sg_tst_excl2" use multiple threads to bombard the
given device with O_EXCL open flags, so only one should succeed at a
time. While holding O_EXCL control a thread attempts a double increment
on an integer in the given LBA. If the integer starts even (after the
first read) then it should remain even if the O_EXCL flag is doing its job.
The "sg_tst_excl" variant uses the Linux SG_IO v3 interface while the
"sg_tst_excl2" uses the more generic sg_pt infrastructure.

"sg_tst_excl3" is a variant of "sg_tst_excl2". "sg_tst_excl3" only does
the double increment from the first thread, each time using O_EXCL on
open. The remaining threads check the value is even, each time doing
an open without the O_EXCL flag.

"bsg_queue_tst" sends an INQUIRY command via the Linux SG_IO v4 interface
which is used by the bsg driver. So it will take device names like
"/dev/bsg/6:0:0:0". It tests if sending repeated INQUIRYs with
the BSG_FLAG_Q_AT_HEAD or BSG_FLAG_Q_AT_TAIL flag makes any difference.

"sg_tst_async" is a test harness for the Linux sg driver. It is multi
threaded, submitting either TEST UNIT READY, READ(16) or WRITE(16) SCSI
commands asynchronously. Each thread opens a file descriptor and submits
those commands up to the queue limit (sg driver has a per file descriptor
queue limit of 16). Multiple threads doing the same thing act as a
multiplier to that queue limit.


NVME Support
============
Firstly the author has no intention of extending this package to contain
general purpose NVMe utilities. That leaves the areas where SCSI overlaps
with NVMe. There was a SCSI to NVMe Translation Layer (SNTL) driver in the
Linux kernel based on a white paper from NVM Express. Intel has withdrawn
that driver and T10 (SCSI) and NVM Express have made no further attempts
to standardize a SNTL. Given the SCSI to ATA Translation Layer (SATL) which
is standardized by T10, it is pretty clear what a SNTL should do.

The NVMe Management Interface (NVME-MI) committee have decided to use SES-3
standard from T10 via the newly added SES Send and SES Receive MI commands.
So the sg_ses utility and this package's library have been extended to use
these commands when a NVMe device (typically a disk enclosure) is detected.
This has been tested by a disk vendor who is happy with the results. Other
user reports are welcome as the author does not have equipment to test
this.

Other utilities in this package that use the SES Send and Receive commands,
or the SNTL in the library are sg_senddiag, sg_inq, sg_raw and sg_readcap.


Command line processing
=======================
These utilities can be divided into 3 groups when their handling of command
line arguments is considered:
  - ad hoc, typically in a short form only, sometimes longer (e.g.
    "sg_logs -pcb /dev/sdc")
  - inspired by the dd Unix command (e.g. sg_dd, sgm_dd, sgp_dd, sg_read)
  - recent utilities use "getopt_long" (see "man getopt_long")
    type command lines. These have short form (starting with "-")
    and corresponding longer form (starting with "--") options.

The older utilities that use ad hoc options, in alphabetical order:
  - sg_emc_trespass, sginfo(1/2), sg_inq, sg_logs, sg_map, sg_modes,
    sg_opcodes, sg_rbuf, sg_rdac, sg_readcap, sg_reset, sg_scan (Linux),
    sg_senddiag, sg_start, sg_test_rwbuf, sg_turs
In sg3_utils version 1.23 the following utilities from this group were
converted to have a dual getopt_long/ad_hoc interface, defaulting to
the getop_long interface:
  - sg_inq, sg_logs, sg_modes, sg_opcodes, sg_rbuf, sg_readcap,
    sg_senddiag, sg_start, sg_turs
These can be switched back to the older (backward compatible) ad hoc
interface by defining the SG3_UTILS_OLD_OPTS environment variable
or using '-O' as the first command line option.

The more recent utilities that use "getopt_long" only are:
  - sg_bt_ctl, sg_compare_and_write, sg_decode_sense, sg_format,
    sg_get_config, sg_get_lba_status, sg_ident, sg_luns, sg_map26,
    sg_persist, sg_prevent, sg_raw, sg_read_attr, sg_read_block_limits,
    sg_read_buffer, sg_read_long, sg_reassign, sg_referrals, sg_rep_pip,
    sg_rep_zones, sg_requests, sg_rmsn, sg_rtpg, sg_safte, sg_sanitize,
    sg_sat_identify, sg_sat_phy_event, sg_sat_read_gplog,
    sg_sat_set_features, sg_scan(w), sg_seek, sg_ses, sg_ses_microcode,
    sg_stpg, sg_stream_ctl, sg_sync, sg_test_rwbuf, sg_timestamp, sg_unmap,
    sg_verify, sg_vpd, sg_write_buffer, sg_write_long, sg_write_same,
    sg_write_verify, sg_write_x, sg_wr_mode, sg_zone, sg_z_act_query


Dangerous code
==============
This C code snippet:
    unsigned char uc = 0x80;
    uint64_t ull;
    ull = (uc << 24);
Somewhat surprisingly sets ull to:
    ull: 0xffffffff80000000
This result is due to the 'unary conversion' of uc to a (32 bit signed)
'int' before the shift. The resultant type from the shift is also an int
and it has its top bit set so there is sign extension when it is assigned
into a 64 bit unsigned integer. Making sure there is no conversion to 'int'
solves the problem. In this case if uc is declared as unsigned int the
result will be as expected (i.e. 0x80000000).


Bypassing the somewhat dangerous shift operators
================================================
The shift operators in C are "<<" and ">>". They can be dangerous (as shown
in the above section) or tedious and hence error prone to use. However they
are often needed to cope with the translation of integers on the host OS to
the corresponding representation within a SCSI command or parameter data
moved to or from a SCSI device. The Logical Block Address (LBA) is a good
example; it is either 32 or 64 bits long typically (i.e. 4 or 8 bytes
respectively). The host machine representation may be big or little endian
and may prefer or require alignment to a particular memory address boundary
(e.g. module 4 (or in 'C' code: "(lba % 4) == 0")). For SCSI commands and
the parameter data moved to or from a SCSI device, the integer
representation is big endian and it is unaligned.

Recent versions of this package have replaced the explicit use of the C
shift operators with a group of functions modelled on those found in the
Linux kernel. These functions contain either "get_unaligned" or
"put_unaligned" in their names and are found in the asm/unaligned.h
header. This package contains the sg_unaligned.h header that implements
a similar set of functions. The current implementation favours correctness
over speed. The functions in the package use a "sg_" prefix but otherwise
use the same function name as the Linux kernel for the same action.

An example of the change made to a snippet of sg_write_buffer.c may
clarify this change. The old code was:

   wbufCmdBlk[3] = (unsigned char)((buffer_offset >> 16) & 0xff);
   wbufCmdBlk[4] = (unsigned char)((buffer_offset >> 8) & 0xff);
   wbufCmdBlk[5] = (unsigned char)(buffer_offset & 0xff);

and it has been replaced by:

   sg_put_unaligned_be24(buffer_offset, wbufCmdBlk + 3);

The Linux kernel only supplies "unaligned" functions for 16, 32 and 64
bit quantities. SCSI commands also have cases of 24 and 48 bit numbers
so sg_unaligned.h contains support for those plus a variant where the
byte length is passed as an argument.

The unaligned functions are inlined for speed (at the possible expense of
space) and now have specializations depending whether the host is big or
(more likely) little endian. These functions can be broken down to a
memcpy() and optionally a byte-swap for 16, 32 and 64 bit operations.
The memcpy() takes care of alignment while the byte-swap (bswap_16(),
bswap_32() and bswap_64() ) addresses integer endianness. If the host is
little endian and a little endian variant of the unaligned functions is
requested, then no byte-swap is required. These specializations can be
"compiled out" with this configure option: './configure
--disable-fast-lebe' in which case the classic "C shifting" technique is
used to implement all the unaligned functions.

Associated with the above change, fixed length integer types seem a better
fit for SCSI command and parameter integers than the traditional integer
types in the C language. Fixed length integer types were standardized in
C99 and require the inclusion of <stdint.h>. For example this means for
an integer that will represent a 64 bit LBA, to favour using "uint64_t"
over the "unsigned long long" type. Also "unsigned char" has mostly been
replaced by "uint8_t" as the 8 bit (unsigned) byte type; "char" is still
used for ASCII text.


Coding Style
============
Everyone has their own C/C++ coding style and the author is no different.
In terms of the GNU indent command:
     indent -i4 -il0 -nut -br -npcs -ncs -ce
is pretty close. That is similar to the Linux kernel coding style but
with 4 space indentations and no tabs.


Other SCSI and storage tools
============================
See https://sg.danny.cz/sg/tools.html


Douglas Gilbert		[email protected]
26th August 2022