sort: spawn fewer threads for small inputs

[platform/upstream/coreutils.git] / NEWS

GNU coreutils NEWS                                    -*- outline -*-

* Noteworthy changes in release ?.? (????-??-??) [?]

** Bug fixes

  cut could segfault when invoked with a user-specified output
  delimiter and an unbounded range like "-f1234567890-".
  [bug introduced in coreutils-5.3.0]

  du would infloop when given --files0-from=DIR
  [bug introduced in coreutils-7.1]

  sort no longer spawns 7 worker threads to sort 16 lines
  [bug introduced in coreutils-8.6]

  wc would dereference a NULL pointer upon an early out-of-memory error
  [bug introduced in coreutils-7.1]

** New features

  dd now accepts the 'nocache' flag to the iflag and oflag options,
  which will discard any cache associated with the files, or
  processed portion thereof.


* Noteworthy changes in release 8.10 (2011-02-04) [stable]

** Bug fixes

  du would abort with a failed assertion when two conditions are met:
  part of the hierarchy being traversed is moved to a higher level in the
  directory tree, and there is at least one more command line directory
  argument following the one containing the moved sub-tree.
  [bug introduced in coreutils-5.1.0]

  join --header now skips the ordering check for the first line
  even if the other file is empty.  [bug introduced in coreutils-8.5]

  rm -f no longer fails for EINVAL or EILSEQ on file systems that
  reject file names invalid for that file system.

  uniq -f NUM no longer tries to process fields after end of line.
  [bug introduced in coreutils-7.0]

** New features

  cp now copies sparse files efficiently on file systems with FIEMAP
  support (ext4, btrfs, xfs, ocfs2).  Before, it had to read 2^20 bytes
  when copying a 1MiB sparse file.  Now, it copies bytes only for the
  non-sparse sections of a file.  Similarly, to induce a hole in the
  output file, it had to detect a long sequence of zero bytes.  Now,
  it knows precisely where each hole in an input file is, and can
  reproduce them efficiently in the output file.  mv also benefits
  when it resorts to copying, e.g., between file systems.

  join now supports -o 'auto' which will automatically infer the
  output format from the first line in each file, to ensure
  the same number of fields are output for each line.

** Changes in behavior

  join no longer reports disorder when one of the files is empty.
  This allows one to use join as a field extractor like:
  join -a1 -o 1.3,1.1 - /dev/null


* Noteworthy changes in release 8.9 (2011-01-04) [stable]

** Bug fixes

  split no longer creates files with a suffix length that
  is dependent on the number of bytes or lines per file.
  [bug introduced in coreutils-8.8]


* Noteworthy changes in release 8.8 (2010-12-22) [stable]

** Bug fixes

  cp -u no longer does unnecessary copying merely because the source
  has finer-grained time stamps than the destination.

  od now prints floating-point numbers without losing information, and
  it no longer omits spaces between floating-point columns in some cases.

  sort -u with at least two threads could attempt to read through a
  corrupted pointer. [bug introduced in coreutils-8.6]

  sort with at least two threads and with blocked output would busy-loop
  (spinlock) all threads, often using 100% of available CPU cycles to
  do no work.  I.e., "sort < big-file | less" could waste a lot of power.
  [bug introduced in coreutils-8.6]

  sort with at least two threads no longer segfaults due to use of pointers
  into the stack of an expired thread. [bug introduced in coreutils-8.6]

  sort --compress no longer mishandles subprocesses' exit statuses,
  no longer hangs indefinitely due to a bug in waiting for subprocesses,
  and no longer generates many more than NMERGE subprocesses.

  sort -m -o f f ... f no longer dumps core when file descriptors are limited.

** Changes in behavior

  sort will not create more than 8 threads by default due to diminishing
  performance gains.  Also the --parallel option is no longer restricted
  to the number of available processors.

** New features

  split accepts the --number option to generate a specific number of files.


* Noteworthy changes in release 8.7 (2010-11-13) [stable]

** Bug fixes

  cp, install, mv, and touch no longer crash when setting file times
  on Solaris 10 Update 9 [Solaris PatchID 144488 and newer expose a
  latent bug introduced in coreutils 8.1, and possibly a second latent
  bug going at least as far back as coreutils 5.97]

  csplit no longer corrupts heap when writing more than 999 files,
  nor does it leak memory for every chunk of input processed
  [the bugs were present in the initial implementation]

  tail -F once again notices changes in a currently unavailable
  remote directory [bug introduced in coreutils-7.5]

** Changes in behavior

  cp --attributes-only now completely overrides --reflink.
  Previously a reflink was needlessly attempted.

  stat's %X, %Y, and %Z directives once again print only the integer
  part of seconds since the epoch.  This reverts a change from
  coreutils-8.6, that was deemed unnecessarily disruptive.
  To obtain a nanosecond-precision time stamp for %X use %.X;
  if you want (say) just 3 fractional digits, use %.3X.
  Likewise for %Y and %Z.

  stat's new %W format directive would print floating point seconds.
  However, with the above change to %X, %Y and %Z, we've made %W work
  the same way as the others.


* Noteworthy changes in release 8.6 (2010-10-15) [stable]

** Bug fixes

  du no longer multiply counts a file that is a directory or whose
  link count is 1, even if the file is reached multiple times by
  following symlinks or via multiple arguments.

  du -H and -L now consistently count pointed-to files instead of
  symbolic links, and correctly diagnose dangling symlinks.

  du --ignore=D now ignores directory D even when that directory is
  found to be part of a directory cycle.  Before, du would issue a
  "NOTIFY YOUR SYSTEM MANAGER" diagnostic and fail.

  split now diagnoses read errors rather than silently exiting.
  [bug introduced in coreutils-4.5.8]

  tac would perform a double-free when given an input line longer than 16KiB.
  [bug introduced in coreutils-8.3]

  tail -F once again notices changes in a currently unavailable directory,
  and works around a Linux kernel bug where inotify runs out of resources.
  [bugs introduced in coreutils-7.5]

  tr now consistently handles case conversion character classes.
  In some locales, valid conversion specifications caused tr to abort,
  while in all locales, some invalid specifications were undiagnosed.
  [bugs introduced in coreutils 6.9.90 and 6.9.92]

** New features

  cp now accepts the --attributes-only option to not copy file data,
  which is useful for efficiently modifying files.

  du recognizes -d N as equivalent to --max-depth=N, for compatibility
  with FreeBSD.

  sort now accepts the --debug option, to highlight the part of the
  line significant in the sort, and warn about questionable options.

  sort now supports -d, -f, -i, -R, and -V in any combination.

  stat now accepts the %m format directive to output the mount point
  for a file.  It also accepts the %w and %W format directives for
  outputting the birth time of a file, if one is available.

** Changes in behavior

  df now consistently prints the device name for a bind mounted file,
  rather than its aliased target.

  du now uses less than half as much memory when operating on trees
  with many hard-linked files.  With --count-links (-l), or when
  operating on trees with no hard-linked files, there is no change.

  ls -l now uses the traditional three field time style rather than
  the wider two field numeric ISO style, in locales where a style has
  not been specified.  The new approach has nicer behavior in some
  locales, including English, which was judged to outweigh the disadvantage
  of generating less-predictable and often worse output in poorly-configured
  locales where there is an onus to specify appropriate non-default styles.
  [The old behavior was introduced in coreutils-6.0 and had been removed
   for English only using a different method since coreutils-8.1]

  rm's -d now evokes an error;  before, it was silently ignored.

  sort -g now uses long doubles for greater range and precision.

  sort -h no longer rejects numbers with leading or trailing ".", and
  no longer accepts numbers with multiple ".".  It now considers all
  zeros to be equal.

  sort now uses the number of available processors to parallelize
  the sorting operation.  The number of sorts run concurrently can be
  limited with the --parallel option or with external process
  control like taskset for example.

  stat now provides translated output when no format is specified.

  stat no longer accepts the --context (-Z) option.  Initially it was
  merely accepted and ignored, for compatibility.  Starting two years
  ago, with coreutils-7.0, its use evoked a warning.  Printing the
  SELinux context of a file can be done with the %C format directive,
  and the default output when no format is specified now automatically
  includes %C when context information is available.

  stat no longer accepts the %C directive when the --file-system
  option is in effect, since security context is a file attribute
  rather than a file system attribute.

  stat now outputs the full sub-second resolution for the atime,
  mtime, and ctime values since the Epoch, when using the %X, %Y, and
  %Z directives of the --format option.  This matches the fact that
  %x, %y, and %z were already doing so for the human-readable variant.

  touch's --file option is no longer recognized.  Use --reference=F (-r)
  instead.  --file has not been documented for 15 years, and its use has
  elicited a warning since coreutils-7.1.

  truncate now supports setting file sizes relative to a reference file.
  Also errors are no longer suppressed for unsupported file types, and
  relative sizes are restricted to supported file types.


* Noteworthy changes in release 8.5 (2010-04-23) [stable]

** Bug fixes

  cp and mv once again support preserving extended attributes.
  [bug introduced in coreutils-8.4]

  cp now preserves "capabilities" when also preserving file ownership.

  ls --color once again honors the 'NORMAL' dircolors directive.
  [bug introduced in coreutils-6.11]

  sort -M now handles abbreviated months that are aligned using blanks
  in the locale database.  Also locales with 8 bit characters are
  handled correctly, including multi byte locales with the caveat
  that multi byte characters are matched case sensitively.

  sort again handles obsolescent key formats (+POS -POS) correctly.
  Previously if -POS was specified, 1 field too many was used in the sort.
  [bug introduced in coreutils-7.2]

** New features

  join now accepts the --header option, to treat the first line of each
  file as a header line to be joined and printed unconditionally.

  timeout now accepts the --kill-after option which sends a kill
  signal to the monitored command if it's still running the specified
  duration after the initial signal was sent.

  who: the "+/-" --mesg (-T) indicator of whether a user/tty is accepting
  messages could be incorrectly listed as "+", when in fact, the user was
  not accepting messages (mesg no).  Before, who would examine only the
  permission bits, and not consider the group of the TTY device file.
  Thus, if a login tty's group would change somehow e.g., to "root",
  that would make it unwritable (via write(1)) by normal users, in spite
  of whatever the permission bits might imply.  Now, when configured
  using the --with-tty-group[=NAME] option, who also compares the group
  of the TTY device with NAME (or "tty" if no group name is specified).

** Changes in behavior

  ls --color no longer emits the final 3-byte color-resetting escape
  sequence when it would be a no-op.

  join -t '' no longer emits an error and instead operates on
  each line as a whole (even if they contain NUL characters).


* Noteworthy changes in release 8.4 (2010-01-13) [stable]

** Bug fixes

  nproc --all is now guaranteed to be as large as the count
  of available processors, which may not have been the case
  on GNU/Linux systems with neither /proc nor /sys available.
  [bug introduced in coreutils-8.1]

** Build-related

  Work around a build failure when using buggy <sys/capability.h>.
  Alternatively, configure with --disable-libcap.

  Compilation would fail on systems using glibc-2.7..2.9 due to changes in
  gnulib's wchar.h that tickled a bug in at least those versions of glibc's
  own <wchar.h> header.  Now, gnulib works around the bug in those older
  glibc <wchar.h> headers.

  Building would fail with a link error (cp/copy.o) when XATTR headers
  were installed without the corresponding library.  Now, configure
  detects that and disables xattr support, as one would expect.


* Noteworthy changes in release 8.3 (2010-01-07) [stable]

** Bug fixes

  cp -p, install -p, mv, and touch -c could trigger a spurious error
  message when using new glibc coupled with an old kernel.
  [bug introduced in coreutils-6.12].

  ls -l --color no longer prints "argetm" in front of dangling
  symlinks when the 'LINK target' directive was given to dircolors.
  [bug introduced in fileutils-4.0]

  pr's page header was improperly formatted for long file names.
  [bug introduced in coreutils-7.2]

  rm -r --one-file-system works once again.
  The rewrite to make rm use fts introduced a regression whereby
  a commmand of the above form would fail for all subdirectories.
  [bug introduced in coreutils-8.0]

  stat -f recognizes more file system types: k-afs, fuseblk, gfs/gfs2, ocfs2,
  and rpc_pipefs. Also Minix V3 is displayed correctly as minix3, not minux3.
  [bug introduced in coreutils-8.1]

  tail -f (inotify-enabled) once again works with remote files.
  The use of inotify with remote files meant that any changes to those
  files that was not done from the local system would go unnoticed.
  [bug introduced in coreutils-7.5]

  tail -F (inotify-enabled) would abort when a tailed file is repeatedly
  renamed-aside and then recreated.
  [bug introduced in coreutils-7.5]

  tail -F (inotify-enabled) could fail to follow renamed files.
  E.g., given a "tail -F a b" process, running "mv a b" would
  make tail stop tracking additions to "b".
  [bug introduced in coreutils-7.5]

  touch -a and touch -m could trigger bugs in some file systems, such
  as xfs or ntfs-3g, and fail to update timestamps.
  [bug introduced in coreutils-8.1]

  wc now prints counts atomically so that concurrent
  processes will not intersperse their output.
  [the issue dates back to the initial implementation]


* Noteworthy changes in release 8.2 (2009-12-11) [stable]

** Bug fixes

  id's use of mgetgroups no longer writes beyond the end of a malloc'd buffer
  [bug introduced in coreutils-8.1]

  id no longer crashes on systems without supplementary group support.
  [bug introduced in coreutils-8.1]

  rm once again handles zero-length arguments properly.
  The rewrite to make rm use fts introduced a regression whereby
  a command like "rm a '' b" would fail to remove "a" and "b", due to
  the presence of the empty string argument.
  [bug introduced in coreutils-8.0]

  sort is now immune to the signal handling of its parent.
  Specifically sort now doesn't exit with an error message
  if it uses helper processes for compression and its parent
  ignores CHLD signals. [bug introduced in coreutils-6.9]

  tail without -f no longer access uninitialized memory
  [bug introduced in coreutils-7.6]

  timeout is now immune to the signal handling of its parent.
  Specifically timeout now doesn't exit with an error message
  if its parent ignores CHLD signals. [bug introduced in coreutils-7.6]

  a user running "make distcheck" in the coreutils source directory,
  with TMPDIR unset or set to the name of a world-writable directory,
  and with a malicious user on the same system
  was vulnerable to arbitrary code execution
  [bug introduced in coreutils-5.0]


* Noteworthy changes in release 8.1 (2009-11-18) [stable]

** Bug fixes

  chcon no longer exits immediately just because SELinux is disabled.
  Even then, chcon may still be useful.
  [bug introduced in coreutils-8.0]

  chcon, chgrp, chmod, chown and du now diagnose an ostensible directory cycle
  and arrange to exit nonzero.  Before, they would silently ignore the
  offending directory and all "contents."

  env -u A=B now fails, rather than silently adding A to the
  environment.  Likewise, printenv A=B silently ignores the invalid
  name.  [the bugs date back to the initial implementation]

  ls --color now handles files with capabilities correctly.  Previously
  files with capabilities were often not colored, and also sometimes, files
  without capabilites were colored in error. [bug introduced in coreutils-7.0]

  md5sum now prints checksums atomically so that concurrent
  processes will not intersperse their output.
  This also affected sum, sha1sum, sha224sum, sha384sum and sha512sum.
  [the bug dates back to the initial implementation]

  mktemp no longer leaves a temporary file behind if it was unable to
  output the name of the file to stdout.
  [the bug dates back to the initial implementation]

  nice -n -1 PROGRAM now runs PROGRAM even when its internal setpriority
  call fails with errno == EACCES.
  [the bug dates back to the initial implementation]

  nice, nohup, and su now refuse to execute the subsidiary program if
  they detect write failure in printing an otherwise non-fatal warning
  message to stderr.

  stat -f recognizes more file system types: afs, cifs, anon-inode FS,
  btrfs, cgroupfs, cramfs-wend, debugfs, futexfs, hfs, inotifyfs, minux3,
  nilfs, securityfs, selinux, xenfs

  tail -f (inotify-enabled) now avoids a race condition.
  Before, any data appended in the tiny interval between the initial
  read-to-EOF and the inotify watch initialization would be ignored
  initially (until more data was appended), or forever, if the file
  were first renamed or unlinked or never modified.
  [The race was introduced in coreutils-7.5]

  tail -F (inotify-enabled) now consistently tails a file that has been
  replaced via renaming.  That operation provokes either of two sequences
  of inotify events.  The less common sequence is now handled as well.
  [The bug came with the implementation change in coreutils-7.5]

  timeout now doesn't exit unless the command it is monitoring does,
  for any specified signal. [bug introduced in coreutils-7.0].

** Changes in behavior

  chroot, env, nice, and su fail with status 125, rather than 1, on
  internal error such as failure to parse command line arguments; this
  is for consistency with stdbuf and timeout, and avoids ambiguity
  with the invoked command failing with status 1.  Likewise, nohup
  fails with status 125 instead of 127.

  du (due to a change in gnulib's fts) can now traverse NFSv4 automounted
  directories in which the stat'd device number of the mount point differs
  during a traversal.  Before, it would fail, because such a mismatch would
  usually represent a serious error or a subversion attempt.

  echo and printf now interpret \e as the Escape character (0x1B).

  rm -f /read-only-fs/nonexistent now succeeds and prints no diagnostic
  on systems with an unlinkat syscall that sets errno to EROFS in that case.
  Before, it would fail with a "Read-only file system" diagnostic.
  Also, "rm /read-only-fs/nonexistent" now reports "file not found" rather
  than the less precise "Read-only file system" error.

** New programs

  nproc: Print the number of processing units available to a process.

** New features

  env and printenv now accept the option --null (-0), as a means to
  avoid ambiguity with newlines embedded in the environment.

  md5sum --check now also accepts openssl-style checksums.
  So do sha1sum, sha224sum, sha384sum and sha512sum.

  mktemp now accepts the option --suffix to provide a known suffix
  after the substitution in the template.  Additionally, uses such as
  "mktemp fileXXXXXX.txt" are able to infer an appropriate --suffix.

  touch now accepts the option --no-dereference (-h), as a means to
  change symlink timestamps on platforms with enough support.


* Noteworthy changes in release 8.0 (2009-10-06) [beta]

** Bug fixes

  cp --preserve=xattr and --archive now preserve extended attributes even
  when the source file doesn't have write access.
  [bug introduced in coreutils-7.1]

  touch -t [[CC]YY]MMDDhhmm[.ss] now accepts a timestamp string ending in .60,
  to accommodate leap seconds.
  [the bug dates back to the initial implementation]

  ls --color now reverts to the color of a base file type consistently
  when the color of a more specific type is disabled.
  [bug introduced in coreutils-5.90]

  ls -LR exits with status 2, not 0, when it encounters a cycle

  ls -is is now consistent with ls -lis in ignoring values returned
  from a failed stat/lstat.  For example ls -Lis now prints "?", not "0",
  for the inode number and allocated size of a dereferenced dangling symlink.

  tail --follow --pid now avoids a race condition where data written
  just before the process dies might not have been output by tail.
  Also, tail no longer delays at all when the specified pid is not live.
  [The race was introduced in coreutils-7.5,
   and the unnecessary delay was present since textutils-1.22o]

** Portability

  On Solaris 9, many commands would mistakenly treat file/ the same as
  file.  Now, even on such a system, path resolution obeys the POSIX
  rules that a trailing slash ensures that the preceeding name is a
  directory or a symlink to a directory.

** Changes in behavior

  id no longer prints SELinux " context=..." when the POSIXLY_CORRECT
  environment variable is set.

  readlink -f now ignores a trailing slash when deciding if the
  last component (possibly via a dangling symlink) can be created,
  since mkdir will succeed in that case.

** New features

  ln now accepts the options --logical (-L) and --physical (-P),
  added by POSIX 2008.  The default behavior is -P on systems like
  GNU/Linux where link(2) creates hard links to symlinks, and -L on
  BSD systems where link(2) follows symlinks.

  stat: without -f, a command-line argument of "-" now means standard input.
  With --file-system (-f), an argument of "-" is now rejected.
  If you really must operate on a file named "-", specify it as
  "./-" or use "--" to separate options from arguments.

** Improvements

  rm: rewrite to use gnulib's fts
  This makes rm -rf significantly faster (400-500%) in some pathological
  cases, and slightly slower (20%) in at least one pathological case.

  rm -r deletes deep hierarchies more efficiently.  Before, execution time
  was quadratic in the depth of the hierarchy, now it is merely linear.
  However, this improvement is not as pronounced as might be expected for
  very deep trees, because prior to this change, for any relative name
  length longer than 8KiB, rm -r would sacrifice official conformance to
  avoid the disproportionate quadratic performance penalty.  Leading to
  another improvement:

  rm -r is now slightly more standards-conformant when operating on
  write-protected files with relative names longer than 8KiB.


* Noteworthy changes in release 7.6 (2009-09-11) [stable]

** Bug fixes

  cp, mv now ignore failure to preserve a symlink time stamp, when it is
  due to their running on a kernel older than what was implied by headers
  and libraries tested at configure time.
  [bug introduced in coreutils-7.5]

  cp --reflink --preserve now preserves attributes when cloning a file.
  [bug introduced in coreutils-7.5]

  cp --preserve=xattr no longer leaks resources on each preservation failure.
  [bug introduced in coreutils-7.1]

  dd now exits with non-zero status when it encounters a write error while
  printing a summary to stderr.
  [bug introduced in coreutils-6.11]

  dd cbs=N conv=unblock would fail to print a final newline when the size
  of the input was not a multiple of N bytes.
  [the non-conforming behavior dates back to the initial implementation]

  df no longer requires that each command-line argument be readable
  [bug introduced in coreutils-7.3]

  ls -i now prints consistent inode numbers also for mount points.
  This makes ls -i DIR less efficient on systems with dysfunctional readdir,
  because ls must stat every file in order to obtain a guaranteed-valid
  inode number.  [bug introduced in coreutils-6.0]

  tail -f (inotify-enabled) now flushes any initial output before blocking.
  Before, this would print nothing and wait: stdbuf -o 4K tail -f /etc/passwd
  Note that this bug affects tail -f only when its standard output is buffered,
  which is relatively unusual.
  [bug introduced in coreutils-7.5]

  tail -f once again works with standard input.  inotify-enabled tail -f
  would fail when operating on a nameless stdin.  I.e., tail -f < /etc/passwd
  would say "tail: cannot watch `-': No such file or directory", yet the
  relatively baroque tail -f /dev/stdin < /etc/passwd would work.  Now, the
  offending usage causes tail to revert to its conventional sleep-based
  (i.e., not inotify-based) implementation.
  [bug introduced in coreutils-7.5]

** Portability

  ln, link: link f z/ would mistakenly succeed on Solaris 10, given an
  existing file, f, and nothing named "z".  ln -T f z/ has the same problem.
  Each would mistakenly create "z" as a link to "f".  Now, even on such a
  system, each command reports the error, e.g.,
  link: cannot create link `z/' to `f': Not a directory

** New features

  cp --reflink accepts a new "auto" parameter which falls back to
  a standard copy if creating a copy-on-write clone is not possible.

** Changes in behavior

  tail -f now ignores "-" when stdin is a pipe or FIFO.
  tail-with-no-args now ignores -f unconditionally when stdin is a pipe or FIFO.
  Before, it would ignore -f only when no file argument was specified,
  and then only when POSIXLY_CORRECT was set.  Now, :|tail -f - terminates
  immediately.  Before, it would block indefinitely.


* Noteworthy changes in release 7.5 (2009-08-20) [stable]

** Bug fixes

  dd's oflag=direct option now works even when the size of the input
  is not a multiple of e.g., 512 bytes.

  dd now handles signals consistently even when they're received
  before data copying has started.

  install runs faster again with SELinux enabled
  [introduced in coreutils-7.0]

  ls -1U (with two or more arguments, at least one a nonempty directory)
  would print entry names *before* the name of the containing directory.
  Also fixed incorrect output of ls -1RU and ls -1sU.
  [introduced in coreutils-7.0]

  sort now correctly ignores fields whose ending position is specified
  before the start position. Previously in numeric mode the remaining
  part of the line after the start position was used as the sort key.
  [This bug appears to have been present in "the beginning".]

  truncate -s failed to skip all whitespace in the option argument in
  some locales.

** New programs

  stdbuf: A new program to run a command with modified stdio buffering
  for its standard streams.

** Changes in behavior

  ls --color: files with multiple hard links are no longer colored differently
  by default. That can be enabled by changing the LS_COLORS environment
  variable. You can control that using the MULTIHARDLINK dircolors input
  variable which corresponds to the 'mh' LS_COLORS item. Note these variables
  were renamed from 'HARDLINK' and 'hl' which were available since
  coreutils-7.1 when this feature was introduced.

** Deprecated options

  nl --page-increment: deprecated in favor of --line-increment, the new option
  maintains the previous semantics and the same short option, -i.

** New features

  chroot now accepts the options --userspec and --groups.

  cp accepts a new option, --reflink: create a lightweight copy
  using copy-on-write (COW).  This is currently only supported within
  a btrfs file system.

  cp now preserves time stamps on symbolic links, when possible

  sort accepts a new option, --human-numeric-sort (-h): sort numbers
  while honoring human readable suffixes like KiB and MB etc.

  tail --follow now uses inotify when possible, to be more responsive
  to file changes and more efficient when monitoring many files.


* Noteworthy changes in release 7.4 (2009-05-07) [stable]

** Bug fixes

  date -d 'next mon', when run on a Monday, now prints the date
  7 days in the future rather than the current day.  Same for any other
  day-of-the-week name, when run on that same day of the week.
  [This bug appears to have been present in "the beginning". ]

  date -d tuesday, when run on a Tuesday -- using date built from the 7.3
  release tarball, not from git -- would print the date 7 days in the future.
  Now, it works properly and prints the current date.  That was due to
  human error (including not-committed changes in a release tarball)
  and the fact that there is no check to detect when the gnulib/ git
  submodule is dirty.

** Build-related

  make check: two tests have been corrected

** Portability

  There have been some ACL-related portability fixes for *BSD,
  inherited from gnulib.


* Noteworthy changes in release 7.3 (2009-05-01) [stable]

** Bug fixes

  cp now diagnoses failure to preserve selinux/xattr attributes when
  --preserve=context,xattr is specified in combination with -a.
  Also, cp no longer suppresses attribute-preservation diagnostics
  when preserving SELinux context was explicitly requested.

  ls now aligns output correctly in the presence of abbreviated month
  names from the locale database that have differing widths.

  ls -v and sort -V now order names like "#.b#" properly

  mv: do not print diagnostics when failing to preserve xattr's on file
  systems without xattr support.

  sort -m no longer segfaults when its output file is also an input file.
  E.g., with this, touch 1; sort -m -o 1 1, sort would segfault.
  [introduced in coreutils-7.2]

** Changes in behavior

  shred, sort, shuf: now use an internal pseudorandom generator by default.
  This is mainly noticable in shred where the 3 random passes it does by
  default should proceed at the speed of the disk.  Previously /dev/urandom
  was used if available, which is relatively slow on GNU/Linux systems.

** Improved robustness

  cp would exit successfully after copying less than the full contents
  of a file larger than ~4000 bytes from a linux-/proc file system to a
  destination file system with a fundamental block size of 4KiB or greater.
  Reading into a 4KiB-or-larger buffer, cp's "read" syscall would return
  a value smaller than 4096, and cp would interpret that as EOF (POSIX
  allows this).  This optimization, now removed, saved 50% of cp's read
  syscalls when copying small files.  Affected linux kernels: at least
  2.6.9 through 2.6.29.
  [the optimization was introduced in coreutils-6.0]

** Portability

  df now pre-mounts automountable directories even with automounters for
  which stat-like syscalls no longer provoke mounting.  Now, df uses open.

  `id -G $USER` now works correctly even on Darwin and NetBSD. Previously it
  would either truncate the group list to 10, or go into an infinite loop,
  due to their non-standard getgrouplist implementations.
  [truncation introduced in coreutils-6.11]
  [infinite loop introduced in coreutils-7.1]


* Noteworthy changes in release 7.2 (2009-03-31) [stable]

** New features

  pwd now accepts the options --logical (-L) and --physical (-P).  For
  compatibility with existing scripts, -P is the default behavior
  unless POSIXLY_CORRECT is requested.

** Bug fixes

  cat once again immediately outputs data it has processed.
  Previously it would have been buffered and only output if enough
  data was read, or on process exit.
  [bug introduced in coreutils-6.0]

  comm's new --check-order option would fail to detect disorder on any pair
  of lines where one was a prefix of the other.  For example, this would
  fail to report the disorder: printf 'Xb\nX\n'>k; comm --check-order k k
  [bug introduced in coreutils-7.0]

  cp once again diagnoses the invalid "cp -rl dir dir" right away,
  rather than after creating a very deep dir/dir/dir/... hierarchy.
  The bug strikes only with both --recursive (-r, -R) and --link (-l).
  [bug introduced in coreutils-7.1]

  ls --sort=version (-v) sorted names beginning with "." inconsistently.
  Now, names that start with "." are always listed before those that don't.

  pr: fix the bug whereby --indent=N (-o) did not indent header lines
  [bug introduced in coreutils-6.9.90]

  sort now handles specified key ends correctly.
  Previously -k1,1b would have caused leading space from field 2 to be
  included in the sort while -k2,3.0 would have not included field 3.

** Changes in behavior

  cat,cp,install,mv,split: these programs now read and write a minimum
  of 32KiB at a time.  This was seen to double throughput when reading
  cached files on GNU/Linux-based systems.

  cp -a now tries to preserve extended attributes (xattr), but does not
  diagnose xattr-preservation failure.  However, cp --preserve=all still does.

  ls --color: hard link highlighting can be now disabled by changing the
  LS_COLORS environment variable. To disable it you can add something like
  this to your profile: eval `dircolors | sed s/hl=[^:]*:/hl=:/`


* Noteworthy changes in release 7.1 (2009-02-21) [stable]

** New features

  Add extended attribute support available on certain filesystems like ext2
  and XFS.
    cp: Tries to copy xattrs when --preserve=xattr or --preserve=all specified
    mv: Always tries to copy xattrs
    install: Never copies xattrs

  cp and mv accept a new option, --no-clobber (-n): silently refrain
  from overwriting any existing destination file

  dd accepts iflag=cio and oflag=cio to open the file in CIO (concurrent I/O)
  mode where this feature is available.

  install accepts a new option, --compare (-C): compare each pair of source
  and destination files, and if the destination has identical content and
  any specified owner, group, permissions, and possibly SELinux context, then
  do not modify the destination at all.

  ls --color now highlights hard linked files, too

  stat -f recognizes the Lustre file system type

** Bug fixes

  chgrp, chmod, chown --silent (--quiet, -f) no longer print some diagnostics
  [bug introduced in coreutils-5.1]

  cp uses much less memory in some situations

  cp -a now correctly tries to preserve SELinux context (announced in 6.9.90),
  doesn't inform about failure, unlike with --preserve=all

  du --files0-from=FILE no longer reads all of FILE into RAM before
  processing the first file name

  seq 9223372036854775807 9223372036854775808 now prints only two numbers
  on systems with extended long double support and good library support.
  Even with this patch, on some systems, it still produces invalid output,
  from 3 to at least 1026 lines long. [bug introduced in coreutils-6.11]

  seq -w now accounts for a decimal point added to the last number
  to correctly print all numbers to the same width.

  wc --files0-from=FILE no longer reads all of FILE into RAM, before
  processing the first file name, unless the list of names is known
  to be small enough.

** Changes in behavior

  cp and mv: the --reply={yes,no,query} option has been removed.
  Using it has elicited a warning for the last three years.

  dd: user specified offsets that are too big are handled better.
  Previously, erroneous parameters to skip and seek could result
  in redundant reading of the file with no warnings or errors.

  du: -H (initially equivalent to --si) is now equivalent to
  --dereference-args, and thus works as POSIX requires

  shred: now does 3 overwrite passes by default rather than 25.

  ls -l now marks SELinux-only files with the less obtrusive '.',
  rather than '+'.  A file with any other combination of MAC and ACL
  is still marked with a '+'.


* Noteworthy changes in release 7.0 (2008-10-05) [beta]

** New programs

  timeout: Run a command with bounded time.
  truncate: Set the size of a file to a specified size.

** New features

  chgrp, chmod, chown, chcon, du, rm: now all display linear performance,
  even when operating on million-entry directories on ext3 and ext4 file
  systems.  Before, they would exhibit O(N^2) performance, due to linear
  per-entry seek time cost when operating on entries in readdir order.
  Rm was improved directly, while the others inherit the improvement
  from the newer version of fts in gnulib.

  comm now verifies that the inputs are in sorted order.  This check can
  be turned off with the --nocheck-order option.

  comm accepts new option, --output-delimiter=STR, that allows specification
  of an output delimiter other than the default single TAB.

  cp and mv: the deprecated --reply=X option is now also undocumented.

  dd accepts iflag=fullblock to make it accumulate full input blocks.
  With this new option, after a short read, dd repeatedly calls read,
  until it fills the incomplete block, reaches EOF, or encounters an error.

  df accepts a new option --total, which produces a grand total of all
  arguments after all arguments have been processed.

  If the GNU MP library is available at configure time, factor and
  expr support arbitrarily large numbers.  Pollard's rho algorithm is
  used to factor large numbers.

  install accepts a new option --strip-program to specify the program used to
  strip binaries.

  ls now colorizes files with capabilities if libcap is available

  ls -v now uses filevercmp function as sort predicate (instead of strverscmp)

  md5sum now accepts the new option, --quiet, to suppress the printing of
  'OK' messages.  sha1sum, sha224sum, sha384sum, and sha512sum accept it, too.

  sort accepts a new option, --files0-from=F, that specifies a file
  containing a null-separated list of files to sort.  This list is used
  instead of filenames passed on the command-line to avoid problems with
  maximum command-line (argv) length.

  sort accepts a new option --batch-size=NMERGE, where NMERGE
  represents the maximum number of inputs that will be merged at once.
  When processing more than NMERGE inputs, sort uses temporary files.

  sort accepts a new option --version-sort (-V, --sort=version),
  specifying that ordering is to be based on filevercmp.

** Bug fixes

  chcon --verbose now prints a newline after each message

  od no longer suffers from platform bugs in printf(3).  This is
  probably most noticeable when using 'od -tfL' to print long doubles.

  seq -0.1 0.1 2 now prints 2,0 when locale's decimal point is ",".
  Before, it would mistakenly omit the final number in that example.

  shuf honors the --zero-terminated (-z) option, even with --input-range=LO-HI

  shuf --head-count is now correctly documented.  The documentation
  previously claimed it was called --head-lines.

** Improvements

  Improved support for access control lists (ACLs): On MacOS X, Solaris 7..10,
  HP-UX 11, Tru64, AIX, IRIX 6.5, and Cygwin, "ls -l" now displays the presence
  of an ACL on a file via a '+' sign after the mode, and "cp -p" copies ACLs.

  join has significantly better performance due to better memory management

  ls now uses constant memory when not sorting and using one_per_line format,
  no matter how many files are in a given directory

  od now aligns fields across lines when printing multiple -t
  specifiers, and no longer prints fields that resulted entirely from
  padding the input out to the least common multiple width.

** Changes in behavior

  stat's --context (-Z) option has always been a no-op.
  Now it evokes a warning that it is obsolete and will be removed.


* Noteworthy changes in release 6.12 (2008-05-31) [stable]

** Bug fixes

  chcon, runcon: --help output now includes the bug-reporting address

  cp -p copies permissions more portably.  For example, on MacOS X 10.5,
  "cp -p some-fifo some-file" no longer fails while trying to copy the
  permissions from the some-fifo argument.

  id with no options now prints the SELinux context only when invoked
  with no USERNAME argument.

  id and groups once again print the AFS-specific nameless group-ID (PAG).
  Printing of such large-numbered, kernel-only (not in /etc/group) group-IDs
  was suppressed in 6.11 due to ignorance that they are useful.

  uniq: avoid subtle field-skipping malfunction due to isblank misuse.
  In some locales on some systems, isblank(240) (aka &nbsp) is nonzero.
  On such systems, uniq --skip-fields=N would fail to skip the proper
  number of fields for some inputs.

  tac: avoid segfault with --regex (-r) and multiple files, e.g.,
  "echo > x; tac -r x x".  [bug present at least in textutils-1.8b, from 1992]

** Changes in behavior

  install once again sets SELinux context, when possible
  [it was deliberately disabled in 6.9.90]


* Noteworthy changes in release 6.11 (2008-04-19) [stable]

** Bug fixes

  configure --enable-no-install-program=groups now works.

  "cp -fR fifo E" now succeeds with an existing E.  Before this fix, using
  -fR to copy a fifo or "special" file onto an existing file would fail
  with EEXIST.  Now, it once again unlinks the destination before trying
  to create the destination file.  [bug introduced in coreutils-5.90]

  dd once again works with unnecessary options like if=/dev/stdin and
  of=/dev/stdout.  [bug introduced in fileutils-4.0h]

  id now uses getgrouplist, when possible.  This results in
  much better performance when there are many users and/or groups.

  ls no longer segfaults on files in /proc when linked with an older version
  of libselinux.  E.g., ls -l /proc/sys would dereference a NULL pointer.

  md5sum would segfault for invalid BSD-style input, e.g.,
  echo 'MD5 (' | md5sum -c -  Now, md5sum ignores that line.
  sha1sum, sha224sum, sha384sum, and sha512sum are affected, too.
  [bug introduced in coreutils-5.1.0]

  md5sum -c would accept a NUL-containing checksum string like "abcd\0..."
  and would unnecessarily read and compute the checksum of the named file,
  and then compare that checksum to the invalid one: guaranteed to fail.
  Now, it recognizes that the line is not valid and skips it.
  sha1sum, sha224sum, sha384sum, and sha512sum are affected, too.
  [bug present in the original version, in coreutils-4.5.1, 1995]

  "mkdir -Z x dir" no longer segfaults when diagnosing invalid context "x"
  mkfifo and mknod would fail similarly.  Now they're fixed.

  mv would mistakenly unlink a destination file before calling rename,
  when the destination had two or more hard links.  It no longer does that.
  [bug introduced in coreutils-5.3.0]

  "paste -d'\' file" no longer overruns memory (heap since coreutils-5.1.2,
  stack before then) [bug present in the original version, in 1992]

  "pr -e" with a mix of backspaces and TABs no longer corrupts the heap
  [bug present in the original version, in 1992]

  "ptx -F'\' long-file-name" would overrun a malloc'd buffer and corrupt
  the heap.  That was triggered by a lone backslash (or odd number of them)
  at the end of the option argument to --flag-truncation=STRING (-F),
  --word-regexp=REGEXP (-W), or --sentence-regexp=REGEXP (-S).

  "rm -r DIR" would mistakenly declare to be "write protected" -- and
  prompt about -- full DIR-relative names longer than MIN (PATH_MAX, 8192).

  "rmdir --ignore-fail-on-non-empty" detects and ignores the failure
  in more cases when a directory is empty.

  "seq -f % 1" would issue the erroneous diagnostic "seq: memory exhausted"
  rather than reporting the invalid string format.
  [bug introduced in coreutils-6.0]

** New features

  join now verifies that the inputs are in sorted order.  This check can
  be turned off with the --nocheck-order option.

  sort accepts the new option --sort=WORD, where WORD can be one of
  general-numeric, month, numeric or random.  These are equivalent to the
  options --general-numeric-sort/-g, --month-sort/-M, --numeric-sort/-n
  and --random-sort/-R, resp.

** Improvements

  id and groups work around an AFS-related bug whereby those programs
  would print an invalid group number, when given no user-name argument.

  ls --color no longer outputs unnecessary escape sequences

  seq gives better diagnostics for invalid formats.

** Portability

  rm now works properly even on systems like BeOS and Haiku,
  which have negative errno values.

** Consistency

  install, mkdir, rmdir and split now write --verbose output to stdout,
  not to stderr.


* Noteworthy changes in release 6.10 (2008-01-22) [stable]

** Bug fixes

  Fix a non-portable use of sed in configure.ac.
  [bug introduced in coreutils-6.9.92]


* Noteworthy changes in release 6.9.92 (2008-01-12) [beta]

** Bug fixes

  cp --parents no longer uses uninitialized memory when restoring the
  permissions of a just-created destination directory.
  [bug introduced in coreutils-6.9.90]

  tr's case conversion would fail in a locale with differing numbers
  of lower case and upper case characters.  E.g., this would fail:
  env LC_CTYPE=en_US.ISO-8859-1 tr '[:upper:]' '[:lower:]'
  [bug introduced in coreutils-6.9.90]

** Improvements

  "touch -d now writable-but-owned-by-someone-else" now succeeds
  whenever that same command would succeed without "-d now".
  Before, it would work fine with no -d option, yet it would
  fail with the ostensibly-equivalent "-d now".


* Noteworthy changes in release 6.9.91 (2007-12-15) [beta]

** Bug fixes

  "ls -l" would not output "+" on SELinux hosts unless -Z was also given.

  "rm" would fail to unlink a non-directory when run in an environment
  in which the user running rm is capable of unlinking a directory.
  [bug introduced in coreutils-6.9]


* Noteworthy changes in release 6.9.90 (2007-12-01) [beta]

** New programs

  arch: equivalent to uname -m, not installed by default
  But don't install this program on Solaris systems.

  chcon: change the SELinux security context of a file

  mktemp: create a temporary file or directory (or names)

  runcon: run a program in a different SELinux security context

** Programs no longer installed by default

  hostname, su

** Changes in behavior

  cp, by default, refuses to copy through a dangling destination symlink
  Set POSIXLY_CORRECT if you require the old, risk-prone behavior.

  pr -F no longer suppresses the footer or the first two blank lines in
  the header.  This is for compatibility with BSD and POSIX.

  tr now warns about an unescaped backslash at end of string.
  The tr from coreutils-5.2.1 and earlier would fail for such usage,
  and Solaris' tr ignores that final byte.

** New features

  Add SELinux support, based on the patch from Fedora:
  * cp accepts new --preserve=context option.
  * "cp -a" works with SELinux:
  Now, cp -a attempts to preserve context, but failure to do so does
  not change cp's exit status.  However "cp --preserve=context" is
  similar, but failure *does* cause cp to exit with nonzero status.
  * install accepts new "-Z, --context=C" option.
  * id accepts new "-Z" option.
  * stat honors the new %C format directive: SELinux security context string
  * ls accepts a slightly modified -Z option.
  * ls: contrary to Fedora version, does not accept --lcontext and --scontext

  The following commands and options now support the standard size
  suffixes kB, M, MB, G, GB, and so on for T, P, Y, Z, and Y:
  head -c, head -n, od -j, od -N, od -S, split -b, split -C,
  tail -c, tail -n.

  cp -p tries to preserve the GID of a file even if preserving the UID
  is not possible.

  uniq accepts a new option: --zero-terminated (-z).  As with the sort
  option of the same name, this makes uniq consume and produce
  NUL-terminated lines rather than newline-terminated lines.

  wc no longer warns about character decoding errors in multibyte locales.
  This means for example that "wc /bin/sh" now produces normal output
  (though the word count will have no real meaning) rather than many
  error messages.

** New build options

  By default, "make install" no longer attempts to install (or even build) su.
  To change that, use ./configure --enable-install-program=su.
  If you also want to install the new "arch" program, do this:
  ./configure --enable-install-program=arch,su.

  You can inhibit the compilation and installation of selected programs
  at configure time.  For example, to avoid installing "hostname" and
  "uptime", use ./configure --enable-no-install-program=hostname,uptime
  Note: currently, "make check" passes, even when arch and su are not
  built (that's the new default).  However, if you inhibit the building
  and installation of other programs, don't be surprised if some parts
  of "make check" fail.

** Remove deprecated options

  df no longer accepts the --kilobytes option.
  du no longer accepts the --kilobytes or --megabytes options.
  ls no longer accepts the --kilobytes option.
  ptx longer accepts the --copyright option.
  who no longer accepts -i or --idle.

** Improved robustness

  ln -f can no longer silently clobber a just-created hard link.
  In some cases, ln could be seen as being responsible for data loss.
  For example, given directories a, b, c, and files a/f and b/f, we
  should be able to do this safely: ln -f a/f b/f c && rm -f a/f b/f
  However, before this change, ln would succeed, and thus cause the
  loss of the contents of a/f.

  stty no longer silently accepts certain invalid hex values
  in its 35-colon command-line argument

** Bug fixes

  chmod no longer ignores a dangling symlink.  Now, chmod fails
  with a diagnostic saying that it cannot operate on such a file.
  [bug introduced in coreutils-5.1.0]

  cp attempts to read a regular file, even if stat says it is empty.
  Before, "cp /proc/cpuinfo c" would create an empty file when the kernel
  reports stat.st_size == 0, while "cat /proc/cpuinfo > c" would "work",
  and create a nonempty one. [bug introduced in coreutils-6.0]

  cp --parents no longer mishandles symlinks to directories in file
  name components in the source, e.g., "cp --parents symlink/a/b d"
  no longer fails.  Also, 'cp' no longer considers a destination
  symlink to be the same as the referenced file when copying links
  or making backups.  For example, if SYM is a symlink to FILE,
  "cp -l FILE SYM" now reports an error instead of silently doing
  nothing.  The behavior of 'cp' is now better documented when the
  destination is a symlink.

  "cp -i --update older newer" no longer prompts; same for mv

  "cp -i" now detects read errors on standard input, and no longer consumes
  too much seekable input; same for ln, install, mv, and rm.

  cut now diagnoses a range starting with zero (e.g., -f 0-2) as invalid;
  before, it would treat it as if it started with 1 (-f 1-2).

  "cut -f 2-0" now fails; before, it was equivalent to "cut -f 2-"

  cut now diagnoses the '-' in "cut -f -" as an invalid range, rather
  than interpreting it as the unlimited range, "1-".

  date -d now accepts strings of the form e.g., 'YYYYMMDD +N days',
  in addition to the usual 'YYYYMMDD N days'.

  du -s now includes the size of any stat'able-but-inaccessible directory
  in the total size.

  du (without -s) prints whatever it knows of the size of an inaccessible
  directory.  Before, du would print nothing for such a directory.

  ls -x DIR would sometimes output the wrong string in place of the
  first entry.  [introduced in coreutils-6.8]

  ls --color would mistakenly color a dangling symlink as if it were
  a regular symlink.  This would happen only when the dangling symlink
  was not a command-line argument and in a directory with d_type support.
  [introduced in coreutils-6.0]

  ls --color, (with a custom LS_COLORS envvar value including the
  ln=target attribute) would mistakenly output the string "target"
  before the name of each symlink.  [introduced in coreutils-6.0]

  od's --skip (-j) option now works even when the kernel says that a
  nonempty regular file has stat.st_size = 0.  This happens at least
  with files in /proc and linux-2.6.22.

  "od -j L FILE" had a bug: when the number of bytes to skip, L, is exactly
  the same as the length of FILE, od would skip *no* bytes.  When the number
  of bytes to skip is exactly the sum of the lengths of the first N files,
  od would skip only the first N-1 files. [introduced in textutils-2.0.9]

  ./printf %.10000000f 1 could get an internal ENOMEM error and generate
  no output, yet erroneously exit with status 0.  Now it diagnoses the error
  and exits with nonzero status.  [present in initial implementation]

  seq no longer mishandles obvious cases like "seq 0 0.000001 0.000003",
  so workarounds like "seq 0 0.000001 0.0000031" are no longer needed.

  seq would mistakenly reject some valid format strings containing %%,
  and would mistakenly accept some invalid ones. e.g., %g%% and %%g, resp.

  "seq .1 .1" would mistakenly generate no output on some systems

  Obsolete sort usage with an invalid ordering-option character, e.g.,
  "env _POSIX2_VERSION=199209 sort +1x" no longer makes sort free an
  invalid pointer [introduced in coreutils-6.5]

  sorting very long lines (relative to the amount of available memory)
  no longer provokes unaligned memory access

  split --line-bytes=N (-C N) no longer creates an empty file
  [this bug is present at least as far back as textutils-1.22 (Jan, 1997)]

  tr -c no longer aborts when translating with Set2 larger than the
  complement of Set1.  [present in the original version, in 1992]

  tr no longer rejects an unmatched [:lower:] or [:upper:] in SET1.
  [present in the original version]


* Noteworthy changes in release 6.9 (2007-03-22) [stable]

** Bug fixes

  cp -x (--one-file-system) would fail to set mount point permissions

  The default block size and output format for df -P are now unaffected by
  the DF_BLOCK_SIZE, BLOCK_SIZE, and BLOCKSIZE environment variables.  It
  is still affected by POSIXLY_CORRECT, though.

  Using pr -m -s (i.e. merging files, with TAB as the output separator)
  no longer inserts extraneous spaces between output columns.

* Noteworthy changes in release 6.8 (2007-02-24) [not-unstable]

** Bug fixes

  chgrp, chmod, and chown now honor the --preserve-root option.
  Before, they would warn, yet continuing traversing and operating on /.

  chmod no longer fails in an environment (e.g., a chroot) with openat
  support but with insufficient /proc support.

  "cp --parents F/G D" no longer creates a directory D/F when F is not
  a directory (and F/G is therefore invalid).

  "cp --preserve=mode" would create directories that briefly had
  too-generous permissions in some cases.  For example, when copying a
  directory with permissions 777 the destination directory might
  temporarily be setgid on some file systems, which would allow other
  users to create subfiles with the same group as the directory.  Fix
  similar problems with 'install' and 'mv'.

  cut no longer dumps core for usage like "cut -f2- f1 f2" with two or
  more file arguments.  This was due to a double-free bug, introduced
  in coreutils-5.3.0.

  dd bs= operands now silently override any later ibs= and obs=
  operands, as POSIX and tradition require.

  "ls -FRL" always follows symbolic links on Linux.  Introduced in
  coreutils-6.0.

  A cross-partition "mv /etc/passwd ~" (by non-root) now prints
  a reasonable diagnostic.  Before, it would print this:
  "mv: cannot remove `/etc/passwd': Not a directory".

  pwd and "readlink -e ." no longer fail unnecessarily when a parent
  directory is unreadable.

  rm (without -f) could prompt when it shouldn't, or fail to prompt
  when it should, when operating on a full name longer than 511 bytes
  and getting an ENOMEM error while trying to form the long name.

  rm could mistakenly traverse into the wrong directory under unusual
  conditions: when a full name longer than 511 bytes specifies a search-only
  directory, and when forming that name fails with ENOMEM, rm would attempt
  to open a truncated-to-511-byte name with the first five bytes replaced
  with "[...]".  If such a directory were to actually exist, rm would attempt
  to remove it.

  "rm -rf /etc/passwd" (run by non-root) now prints a diagnostic.
  Before it would print nothing.

  "rm --interactive=never F" no longer prompts for an unwritable F

  "rm -rf D" would emit an misleading diagnostic when failing to
  remove a symbolic link within the unwritable directory, D.
  Introduced in coreutils-6.0.  Similarly, when a cross-partition
  "mv" fails because the source directory is unwritable, it now gives
  a reasonable diagnostic.  Before, this would print
    $ mkdir /tmp/x; touch /tmp/x/y; chmod -w /tmp/x;
    $ test $(stat -c %d /tmp/x) -ne $(stat -c %d .) && mv /tmp/x/y .
    mv: cannot remove `/tmp/x/y': Not a directory
  Now it prints this:
    mv: cannot remove `/tmp/x/y': Permission denied.

** New features

  sort's new --compress-program=PROG option specifies a compression
  program to use when writing and reading temporary files.
  This can help save both time and disk space when sorting large inputs.

  sort accepts the new option -C, which acts like -c except no diagnostic
  is printed.  Its --check option now accepts an optional argument, and
  --check=quiet and --check=silent are now aliases for -C, while
  --check=diagnose-first is an alias for -c or plain --check.


* Noteworthy changes in release 6.7 (2006-12-08) [stable]

** Bug fixes

  When cp -p copied a file with special mode bits set, the same bits
  were set on the copy even when ownership could not be preserved.
  This could result in files that were setuid to the wrong user.
  To fix this, special mode bits are now set in the copy only if its
  ownership is successfully preserved.  Similar problems were fixed
  with mv when copying across file system boundaries.  This problem
  affects all versions of coreutils through 6.6.

  cp --preserve=ownership would create output files that temporarily
  had too-generous permissions in some cases.  For example, when
  copying a file with group A and mode 644 into a group-B sticky
  directory, the output file was briefly readable by group B.
  Fix similar problems with cp options like -p that imply
  --preserve=ownership, with install -d when combined with either -o
  or -g, and with mv when copying across file system boundaries.
  This bug affects all versions of coreutils through 6.6.

  du --one-file-system (-x) would skip subdirectories of any directory
  listed as second or subsequent command line argument.  This bug affects
  coreutils-6.4, 6.5 and 6.6.


* Noteworthy changes in release 6.6 (2006-11-22) [stable]

** Bug fixes

  ls would segfault (dereference a NULL pointer) for a file with a
  nameless group or owner.  This bug was introduced in coreutils-6.5.

  A bug in the latest official m4/gettext.m4 (from gettext-0.15)
  made configure fail to detect gettext support, due to the unusual
  way in which coreutils uses AM_GNU_GETTEXT.

** Improved robustness

  Now, du (and the other fts clients: chmod, chgrp, chown) honor a
  trailing slash in the name of a symlink-to-directory even on
  Solaris 9, by working around its buggy fstatat implementation.


* Major changes in release 6.5 (2006-11-19) [stable]

** Bug fixes

  du (and the other fts clients: chmod, chgrp, chown) would exit early
  when encountering an inaccessible directory on a system with native
  openat support (i.e., linux-2.6.16 or newer along with glibc-2.4
  or newer).  This bug was introduced with the switch to gnulib's
  openat-based variant of fts, for coreutils-6.0.

  "ln --backup f f" now produces a sensible diagnostic

** New features

  rm accepts a new option: --one-file-system


* Major changes in release 6.4 (2006-10-22) [stable]

** Bug fixes

  chgrp and chown would malfunction when invoked with both -R and -H and
  with one or more of the following: --preserve-root, --verbose, --changes,
  --from=o:g (chown only).  This bug was introduced with the switch to
  gnulib's openat-based variant of fts, for coreutils-6.0.

  cp --backup dir1 dir2, would rename an existing dir2/dir1 to dir2/dir1~.
  This bug was introduced in coreutils-6.0.

  With --force (-f), rm no longer fails for ENOTDIR.
  For example, "rm -f existing-non-directory/anything" now exits
  successfully, ignoring the error about a nonexistent file.


* Major changes in release 6.3 (2006-09-30) [stable]

** Improved robustness

  pinky no longer segfaults on Darwin 7.9.0 (MacOS X 10.3.9) due to a
  buggy native getaddrinfo function.

  rm works around a bug in Darwin 7.9.0 (MacOS X 10.3.9) that would
  sometimes keep it from removing all entries in a directory on an HFS+
  or NFS-mounted partition.

  sort would fail to handle very large input (around 40GB) on systems with a
  mkstemp function that returns a file descriptor limited to 32-bit offsets.

** Bug fixes

  chmod would fail unnecessarily in an unusual case: when an initially-
  inaccessible argument is rendered accessible by chmod's action on a
  preceding command line argument.  This bug also affects chgrp, but
  it is harder to demonstrate.  It does not affect chown.  The bug was
  introduced with the switch from explicit recursion to the use of fts
  in coreutils-5.1.0 (2003-10-15).

  cp -i and mv -i occasionally neglected to prompt when the copy or move
  action was bound to fail.  This bug dates back to before fileutils-4.0.

  With --verbose (-v), cp and mv would sometimes generate no output,
  or neglect to report file removal.

  For the "groups" command:

    "groups" no longer prefixes the output with "user :" unless more
    than one user is specified; this is for compatibility with BSD.

    "groups user" now exits nonzero when it gets a write error.

    "groups" now processes options like --help more compatibly.

  shuf would infloop, given 8KB or more of piped input

** Portability

  Versions of chmod, chown, chgrp, du, and rm (tools that use openat etc.)
  compiled for Solaris 8 now also work when run on Solaris 10.


* Major changes in release 6.2 (2006-09-18) [stable candidate]

** Changes in behavior

  mkdir -p and install -d (or -D) now use a method that forks a child
  process if the working directory is unreadable and a later argument
  uses a relative file name.  This avoids some race conditions, but it
  means you may need to kill two processes to stop these programs.

  rm now rejects attempts to remove the root directory, e.g., `rm -fr /'
  now fails without removing anything.  Likewise for any file name with
  a final `./' or `../' component.

  tail now ignores the -f option if POSIXLY_CORRECT is set, no file
  operand is given, and standard input is any FIFO; formerly it did
  this only for pipes.

** Infrastructure changes

  Coreutils now uses gnulib via the gnulib-tool script.
  If you check the source out from CVS, then follow the instructions
  in README-cvs.  Although this represents a large change to the
  infrastructure, it should cause no change in how the tools work.

** Bug fixes

  cp --backup no longer fails when the last component of a source file
  name is "." or "..".

  "ls --color" would highlight other-writable and sticky directories
  no differently than regular directories on a file system with
  dirent.d_type support.

  "mv -T --verbose --backup=t A B" now prints the " (backup: B.~1~)"
  suffix when A and B are directories as well as when they are not.

  mv and "cp -r" no longer fail when invoked with two arguments
  where the first one names a directory and the second name ends in
  a slash and doesn't exist.  E.g., "mv dir B/", for nonexistent B,
  now succeeds, once more.  This bug was introduced in coreutils-5.3.0.


* Major changes in release 6.1 (2006-08-19) [unstable]

** Changes in behavior

  df now considers BSD "kernfs" file systems to be dummies

** New features

  printf now supports the 'I' flag on hosts whose underlying printf
  implementations support 'I', e.g., "printf %Id 2".

** Bug fixes

  cp --sparse preserves sparseness at the end of a file, even when
  the file's apparent size is not a multiple of its block size.
  [introduced with the original design, in fileutils-4.0r, 2000-04-29]

  df (with a command line argument) once again prints its header
  [introduced in coreutils-6.0]

  ls -CF would misalign columns in some cases involving non-stat'able files
  [introduced in coreutils-6.0]

* Major changes in release 6.0 (2006-08-15) [unstable]

** Improved robustness

  df: if the file system claims to have more available than total blocks,
  report the number of used blocks as being "total - available"
  (a negative number) rather than as garbage.

  dircolors: a new autoconf run-test for AIX's buggy strndup function
  prevents malfunction on that system;  may also affect cut, expand,
  and unexpand.

  fts no longer changes the current working directory, so its clients
  (chmod, chown, chgrp, du) no longer malfunction under extreme conditions.

  pwd and other programs using lib/getcwd.c work even on file systems
  where dirent.d_ino values are inconsistent with those from stat.st_ino.

  rm's core is now reentrant: rm --recursive (-r) now processes
  hierarchies without changing the working directory at all.

** Changes in behavior

  basename and dirname now treat // as different from / on platforms
  where the two are distinct.

  chmod, install, and mkdir now preserve a directory's set-user-ID and
  set-group-ID bits unless you explicitly request otherwise.  E.g.,
  `chmod 755 DIR' and `chmod u=rwx,go=rx DIR' now preserve DIR's
  set-user-ID and set-group-ID bits instead of clearing them, and
  similarly for `mkdir -m 755 DIR' and `mkdir -m u=rwx,go=rx DIR'.  To
  clear the bits, mention them explicitly in a symbolic mode, e.g.,
  `mkdir -m u=rwx,go=rx,-s DIR'.  To set them, mention them explicitly
  in either a symbolic or a numeric mode, e.g., `mkdir -m 2755 DIR',
  `mkdir -m u=rwx,go=rx,g+s' DIR.  This change is for convenience on
  systems where these bits inherit from parents.  Unfortunately other
  operating systems are not consistent here, and portable scripts
  cannot assume the bits are set, cleared, or preserved, even when the
  bits are explicitly mentioned.  For example, OpenBSD 3.9 `mkdir -m
  777 D' preserves D's setgid bit but `chmod 777 D' clears it.
  Conversely, Solaris 10 `mkdir -m 777 D', `mkdir -m g-s D', and
  `chmod 0777 D' all preserve D's setgid bit, and you must use
  something like `chmod g-s D' to clear it.

  `cp --link --no-dereference' now works also on systems where the
  link system call cannot create a hard link to a symbolic link.
  This change has no effect on systems with a Linux-based kernel.

  csplit and nl now use POSIX syntax for regular expressions, not
  Emacs syntax.  As a result, character classes like [[:print:]] and
  interval expressions like A\{1,9\} now have their usual meaning,
  . no longer matches the null character, and \ must precede the + and
  ? operators.

  date: a command like date -d '2006-04-23 21 days ago' would print
  the wrong date in some time zones.  (see the test for an example)

  df changes:

    df now considers "none" and "proc" file systems to be dummies and
    therefore does not normally display them.  Also, inaccessible file
    systems (which can be caused by shadowed mount points or by
    chrooted bind mounts) are now dummies, too.

    df now fails if it generates no output, so you can inspect the
    exit status of a command like "df -t ext3 -t reiserfs DIR" to test
    whether DIR is on a file system of type "ext3" or "reiserfs".

  expr no longer complains about leading ^ in a regular expression
  (the anchor is ignored), or about regular expressions like A** (the
  second "*" is ignored).  expr now exits with status 2 (not 3) for
  errors it detects in the expression's values; exit status 3 is now
  used only for internal errors (such as integer overflow, which expr
  now checks for).

  install and mkdir now implement the X permission symbol correctly,
  e.g., `mkdir -m a+X dir'; previously the X was ignored.

  install now creates parent directories with mode u=rwx,go=rx (755)
  instead of using the mode specified by the -m option; and it does
  not change the owner or group of parent directories.  This is for
  compatibility with BSD and closes some race conditions.

  ln now uses different (and we hope clearer) diagnostics when it fails.
  ln -v now acts more like FreeBSD, so it generates output only when
  successful and the output is easier to parse.

  ls now defaults to --time-style='locale', not --time-style='posix-long-iso'.
  However, the 'locale' time style now behaves like 'posix-long-iso'
  if your locale settings appear to be messed up.  This change
  attempts to have the default be the best of both worlds.

  mkfifo and mknod no longer set special mode bits (setuid, setgid,
  and sticky) with the -m option.

  nohup's usual diagnostic now more precisely specifies the I/O
  redirections, e.g., "ignoring input and appending output to
  nohup.out".  Also, nohup now redirects stderr to nohup.out (or
  $HOME/nohup.out) if stdout is closed and stderr is a tty; this is in
  response to Open Group XCU ERN 71.

  rm --interactive now takes an optional argument, although the
  default of using no argument still acts like -i.

  rm no longer fails to remove an empty, unreadable directory

  seq changes:

    seq defaults to a minimal fixed point format that does not lose
    information if seq's operands are all fixed point decimal numbers.
    You no longer need the `-f%.f' in `seq -f%.f 1048575 1024 1050623',
    for example, since the default format now has the same effect.

    seq now lets you use %a, %A, %E, %F, and %G formats.

    seq now uses long double internally rather than double.

  sort now reports incompatible options (e.g., -i and -n) rather than
  silently ignoring one of them.

  stat's --format=FMT option now works the way it did before 5.3.0:
  FMT is automatically newline terminated.  The first stable release
  containing this change was 5.92.

  stat accepts the new option --printf=FMT, where FMT is *not*
  automatically newline terminated.

  stat: backslash escapes are interpreted in a format string specified
  via --printf=FMT, but not one specified via --format=FMT.  That includes
  octal (\ooo, at most three octal digits), hexadecimal (\xhh, one or
  two hex digits), and the standard sequences (\a, \b, \f, \n, \r, \t,
  \v, \", \\).

  With no operand, 'tail -f' now silently ignores the '-f' only if
  standard input is a FIFO or pipe and POSIXLY_CORRECT is set.
  Formerly, it ignored the '-f' when standard input was a FIFO, pipe,
  or socket.

** Scheduled for removal

  ptx's --copyright (-C) option is scheduled for removal in 2007, and
  now evokes a warning.  Use --version instead.

  rm's --directory (-d) option is scheduled for removal in 2006.  This
  option has been silently ignored since coreutils 5.0.  On systems
  that support unlinking of directories, you can use the "unlink"
  command to unlink a directory.

  Similarly, we are considering the removal of ln's --directory (-d,
  -F) option in 2006.  Please write to <bug-coreutils@gnu.org> if this
  would cause a problem for you.  On systems that support hard links
  to directories, you can use the "link" command to create one.

** New programs

  base64: base64 encoding and decoding (RFC 3548) functionality.
  sha224sum: print or check a SHA224 (224-bit) checksum
  sha256sum: print or check a SHA256 (256-bit) checksum
  sha384sum: print or check a SHA384 (384-bit) checksum
  sha512sum: print or check a SHA512 (512-bit) checksum
  shuf: Shuffle lines of text.

** New features

  chgrp now supports --preserve-root, --no-preserve-root (default),
  as it was documented to do, and just as chmod, chown, and rm do.

  New dd iflag= and oflag= flags:

    'directory' causes dd to fail unless the file is a directory, on
    hosts that support this (e.g., Linux kernels, version 2.1.126 and
    later).  This has limited utility but is present for completeness.

    'noatime' causes dd to read a file without updating its access
    time, on hosts that support this (e.g., Linux kernels, version
    2.6.8 and later).

    'nolinks' causes dd to fail if the file has multiple hard links,
    on hosts that support this (e.g., Solaris 10 and later).

  ls accepts the new option --group-directories-first, to make it
  list directories before files.

  rm now accepts the -I (--interactive=once) option.  This new option
  prompts once if rm is invoked recursively or if more than three
  files are being deleted, which is less intrusive than -i prompting
  for every file, but provides almost the same level of protection
  against mistakes.

  shred and sort now accept the --random-source option.

  sort now accepts the --random-sort (-R) option and `R' ordering option.

  sort now supports obsolete usages like "sort +1 -2" unless
  POSIXLY_CORRECT is set.  However, when conforming to POSIX
  1003.1-2001 "sort +1" still sorts the file named "+1".

  wc accepts a new option --files0-from=FILE, where FILE contains a
  list of NUL-terminated file names.

** Bug fixes

  cat with any of the options, -A -v -e -E -T, when applied to a
  file in /proc or /sys (linux-specific), would truncate its output,
  usually printing nothing.

  cp -p would fail in a /proc-less chroot, on some systems

  When `cp -RL' encounters the same directory more than once in the
  hierarchy beneath a single command-line argument, it no longer confuses
  them with hard-linked directories.

  fts-using tools (chmod, chown, chgrp, du) no longer fail due to
  a double-free bug -- it could be triggered by making a directory
  inaccessible while e.g., du is traversing the hierarchy under it.

  fts-using tools (chmod, chown, chgrp, du) no longer misinterpret
  a very long symlink chain as a dangling symlink.  Before, such a
  misinterpretation would cause these tools not to diagnose an ELOOP error.

  ls --indicator-style=file-type would sometimes stat a symlink
  unnecessarily.

  ls --file-type worked like --indicator-style=slash (-p),
  rather than like --indicator-style=file-type.

  mv: moving a symlink into the place of an existing non-directory is
  now done atomically;  before, mv would first unlink the destination.

  mv -T DIR EMPTY_DIR no longer fails unconditionally.  Also, mv can
  now remove an empty destination directory: mkdir -p a b/a; mv a b

  rm (on systems with openat) can no longer exit before processing
  all command-line arguments.

  rm is no longer susceptible to a few low-probability memory leaks.

  rm -r no longer fails to remove an inaccessible and empty directory

  rm -r's cycle detection code can no longer be tricked into reporting
  a false positive (introduced in fileutils-4.1.9).

  shred --remove FILE no longer segfaults on Gentoo systems

  sort would fail for large inputs (~50MB) on systems with a buggy
  mkstemp function.  sort and tac now use the replacement mkstemp
  function, and hence are no longer subject to limitations (of 26 or 32,
  on the maximum number of files from a given template) on HP-UX 10.20,
  SunOS 4.1.4, Solaris 2.5.1 and OSF1/Tru64 V4.0F&V5.1.

  tail -f once again works on a file with the append-only
  attribute (affects at least Linux ext2, ext3, xfs file systems)

* Major changes in release 5.97 (2006-06-24) [stable]
* Major changes in release 5.96 (2006-05-22) [stable]
* Major changes in release 5.95 (2006-05-12) [stable]
* Major changes in release 5.94 (2006-02-13) [stable]

[see the b5_9x branch for details]

* Major changes in release 5.93 (2005-11-06) [stable]

** Bug fixes

  dircolors no longer segfaults upon an attempt to use the new
  STICKY_OTHER_WRITABLE (OWT) attribute.

  du no longer overflows a counter when processing a file larger than
  2^31-1 on some 32-bit systems (at least some AIX 5.1 configurations).

  md5sum once again defaults to using the ` ' non-binary marker
  (rather than the `*' binary marker) by default on Unix-like systems.

  mkdir -p and install -d no longer exit nonzero when asked to create
  a directory like `nonexistent/.'

  rm emits a better diagnostic when (without -r) it fails to remove
  a directory on e.g., Solaris 9/10 systems.

  tac now works when stdin is a tty, even on non-Linux systems.

  "tail -c 2 FILE" and "touch 0101000000" now operate as POSIX
  1003.1-2001 requires, even when coreutils is conforming to older
  POSIX standards, as the newly-required behavior is upward-compatible
  with the old.

  The documentation no longer mentions rm's --directory (-d) option.

** Build-related bug fixes

  installing .mo files would fail


* Major changes in release 5.92 (2005-10-22) [stable]

** Bug fixes

  chmod now diagnoses an invalid mode string starting with an octal digit

  dircolors now properly quotes single-quote characters


* Major changes in release 5.91 (2005-10-17) [stable candidate]

** Bug fixes

  "mkdir -p /a/b/c" no longer fails merely because a leading prefix
  directory (e.g., /a or /a/b) exists on a read-only file system.

** Removed options

  tail's --allow-missing option has been removed.  Use --retry instead.

  stat's --link and -l options have been removed.
  Use --dereference (-L) instead.

** Deprecated options

  Using ls, du, or df with the --kilobytes option now evokes a warning
  that the long-named option is deprecated.  Use `-k' instead.

  du's long-named --megabytes option now evokes a warning.
  Use -m instead.


* Major changes in release 5.90 (2005-09-29) [unstable]

** Bring back support for `head -NUM', `tail -NUM', etc. even when
  conforming to POSIX 1003.1-2001.  The following changes apply only
  when conforming to POSIX 1003.1-2001; there is no effect when
  conforming to older POSIX versions.

  The following usages now behave just as when conforming to older POSIX:

    date -I
    expand -TAB1[,TAB2,...]
    fold -WIDTH
    head -NUM
    join -j FIELD
    join -j1 FIELD
    join -j2 FIELD
    join -o FIELD_NAME1 FIELD_NAME2...
    nice -NUM
    od -w
    pr -S
    split -NUM
    tail -[NUM][bcl][f] [FILE]

  The following usages no longer work, due to the above changes:

    date -I TIMESPEC  (use `date -ITIMESPEC' instead)
    od -w WIDTH       (use `od -wWIDTH' instead)
    pr -S STRING      (use `pr -SSTRING' instead)

  A few usages still have behavior that depends on which POSIX standard is
  being conformed to, and portable applications should beware these
  problematic usages.  These include:

    Problematic       Standard-conforming replacement, depending on
       usage            whether you prefer the behavior of:
                      POSIX 1003.2-1992    POSIX 1003.1-2001
    sort +4           sort -k 5            sort ./+4
    tail +4           tail -n +4           tail ./+4
    tail - f          tail f               [see (*) below]
    tail -c 4         tail -c 10 ./4       tail -c4
    touch 12312359 f  touch -t 12312359 f  touch ./12312359 f
    uniq +4           uniq -s 4            uniq ./+4

    (*) "tail - f" does not conform to POSIX 1003.1-2001; to read
    standard input and then "f", use the command "tail -- - f".

  These changes are in response to decisions taken in the January 2005
  Austin Group standardization meeting.  For more details, please see
  "Utility Syntax Guidelines" in the Minutes of the January 2005
  Meeting <http://www.opengroup.org/austin/docs/austin_239.html>.

** Binary input and output are now implemented more consistently.
  These changes affect only platforms like MS-DOS that distinguish
  between binary and text files.

  The following programs now always use text input/output:

    expand unexpand

  The following programs now always use binary input/output to copy data:

    cp install mv shred

  The following programs now always use binary input/output to copy
  data, except for stdin and stdout when it is a terminal.

    head tac tail tee tr
    (cat behaves similarly, unless one of the options -bensAE is used.)

  cat's --binary or -B option has been removed.  It existed only on
  MS-DOS-like platforms, and didn't work as documented there.

  md5sum and sha1sum now obey the -b or --binary option, even if
  standard input is a terminal, and they no longer report files to be
  binary if they actually read them in text mode.

** Changes for better conformance to POSIX

  cp, ln, mv, rm changes:

    Leading white space is now significant in responses to yes-or-no questions.
    For example, if "rm" asks "remove regular file `foo'?" and you respond
    with " y" (i.e., space before "y"), it counts as "no".

  dd changes:

    On a QUIT or PIPE signal, dd now exits without printing statistics.

    On hosts lacking the INFO signal, dd no longer treats the USR1
    signal as if it were INFO when POSIXLY_CORRECT is set.

    If the file F is non-seekable and contains fewer than N blocks,
    then before copying "dd seek=N of=F" now extends F with zeroed
    blocks until F contains N blocks.

  fold changes:

    When POSIXLY_CORRECT is set, "fold file -3" is now equivalent to
    "fold file ./-3", not the obviously-erroneous "fold file ./-w3".

  ls changes:

    -p now marks only directories; it is equivalent to the new option
    --indicator-style=slash.  Use --file-type or
    --indicator-style=file-type to get -p's old behavior.

  nice changes:

    Documentation and diagnostics now refer to "nicenesses" (commonly
    in the range -20...19) rather than "nice values" (commonly 0...39).

  nohup changes:

    nohup now ignores the umask when creating nohup.out.

    nohup now closes stderr if it is a terminal and stdout is closed.

    nohup now exits with status 127 (not 1) when given an invalid option.

  pathchk changes:

    It now rejects the empty name in the normal case.  That is,
    "pathchk -p ''" now fails, and "pathchk ''" fails unless the
    current host (contra POSIX) allows empty file names.

    The new -P option checks whether a file name component has leading "-",
    as suggested in interpretation "Austin-039:XCU:pathchk:pathchk -p"
    <http://www.opengroup.org/austin/interps/doc.tpl?gdid=6232>.
    It also rejects the empty name even if the current host accepts it; see
    <http://www.opengroup.org/austin/interps/doc.tpl?gdid=6233>.

    The --portability option is now equivalent to -p -P.

** Bug fixes

  chmod, mkdir, mkfifo, and mknod formerly mishandled rarely-used symbolic
  permissions like =xX and =u, and did not properly diagnose some invalid
  strings like g+gr, ug,+x, and +1.  These bugs have been fixed.

  csplit could produce corrupt output, given input lines longer than 8KB

  dd now computes statistics using a realtime clock (if available)
  rather than the time-of-day clock, to avoid glitches if the
  time-of-day is changed while dd is running.  Also, it avoids
  using unsafe code in signal handlers; this fixes some core dumps.

  expr and test now correctly compare integers of unlimited magnitude.

  expr now detects integer overflow when converting strings to integers,
  rather than silently wrapping around.

  ls now refuses to generate time stamps containing more than 1000 bytes, to
  foil potential denial-of-service attacks on hosts with very large stacks.

  "mkdir -m =+x dir" no longer ignores the umask when evaluating "+x",
  and similarly for mkfifo and mknod.

  "mkdir -p /tmp/a/b dir" no longer attempts to create the `.'-relative
  directory, dir (in /tmp/a), when, after creating /tmp/a/b, it is unable
  to return to its initial working directory.  Similarly for "install -D
  file /tmp/a/b/file".

  "pr -D FORMAT" now accepts the same formats that "date +FORMAT" does.

  stat now exits nonzero if a file operand does not exist

** Improved robustness

  Date no longer needs to allocate virtual memory to do its job,
  so it can no longer fail due to an out-of-memory condition,
  no matter how large the result.

** Improved portability

  hostid now prints exactly 8 hexadecimal digits, possibly with leading zeros,
  and without any spurious leading "fff..." on 64-bit hosts.

  nice now works on Darwin 7.7.0 in spite of its invalid definition of NZERO.

  `rm -r' can remove all entries in a directory even when it is on a
  file system for which readdir is buggy and that was not checked by
  coreutils' old configure-time run-test.

  sleep no longer fails when resumed after being suspended on linux-2.6.8.1,
  in spite of that kernel's buggy nanosleep implementation.

** New features

  chmod -w now complains if its behavior differs from what chmod a-w
  would do, and similarly for chmod -r, chmod -x, etc.

  cp and mv: the --reply=X option is deprecated

  date accepts the new option --rfc-3339=TIMESPEC.  The old --iso-8601 (-I)
  option is deprecated; it still works, but new applications should avoid it.
  date, du, ls, and pr's time formats now support new %:z, %::z, %:::z
  specifiers for numeric time zone offsets like -07:00, -07:00:00, and -07.

  dd has new iflag= and oflag= flags "binary" and "text", which have an
  effect only on nonstandard platforms that distinguish text from binary I/O.

  dircolors now supports SETUID, SETGID, STICKY_OTHER_WRITABLE,
  OTHER_WRITABLE, and STICKY, with ls providing default colors for these
  categories if not specified by dircolors.

  du accepts new options: --time[=TYPE] and --time-style=STYLE

  join now supports a NUL field separator, e.g., "join -t '\0'".
  join now detects and reports incompatible options, e.g., "join -t x -t y",

  ls no longer outputs an extra space between the mode and the link count
  when none of the listed files has an ACL.

  md5sum --check now accepts multiple input files, and similarly for sha1sum.

  If stdin is a terminal, nohup now redirects it from /dev/null to
  prevent the command from tying up an OpenSSH session after you logout.

  "rm -FOO" now suggests "rm ./-FOO" if the file "-FOO" exists and
  "-FOO" is not a valid option.

  stat -f -c %S outputs the fundamental block size (used for block counts).
  stat -f's default output format has been changed to output this size as well.
  stat -f recognizes file systems of type XFS and JFS

  "touch -" now touches standard output, not a file named "-".

  uname -a no longer generates the -p and -i outputs if they are unknown.

* Major changes in release 5.3.0 (2005-01-08) [unstable]

** Bug fixes

  Several fixes to chgrp and chown for compatibility with POSIX and BSD:

    Do not affect symbolic links by default.
    Now, operate on whatever a symbolic link points to, instead.
    To get the old behavior, use --no-dereference (-h).

    --dereference now works, even when the specified owner
    and/or group match those of an affected symlink.

    Check for incompatible options.  When -R and --dereference are
    both used, then either -H or -L must also be used.  When -R and -h
    are both used, then -P must be in effect.

    -H, -L, and -P have no effect unless -R is also specified.
    If -P and -R are both specified, -h is assumed.

    Do not optimize away the chown() system call when the file's owner
    and group already have the desired value.  This optimization was
    incorrect, as it failed to update the last-changed time and reset
    special permission bits, as POSIX requires.

    "chown : file", "chown '' file", and "chgrp '' file" now succeed
    without changing the uid or gid, instead of reporting an error.

    Do not report an error if the owner or group of a
    recursively-encountered symbolic link cannot be updated because
    the file system does not support it.

  chmod now accepts multiple mode-like options, e.g., "chmod -r -w f".

  chown is no longer subject to a race condition vulnerability, when
  used with --from=O:G and without the (-h) --no-dereference option.

  cut's --output-delimiter=D option works with abutting byte ranges.

  dircolors's documentation now recommends that shell scripts eval
  "`dircolors`" rather than `dircolors`, to avoid shell expansion pitfalls.

  du no longer segfaults when a subdirectory of an operand
  directory is removed while du is traversing that subdirectory.
  Since the bug was in the underlying fts.c module, it also affected
  chown, chmod, and chgrp.

  du's --exclude-from=FILE and --exclude=P options now compare patterns
  against the entire name of each file, rather than against just the
  final component.

  echo now conforms to POSIX better.  It supports the \0ooo syntax for
  octal escapes, and \c now terminates printing immediately.  If
  POSIXLY_CORRECT is set and the first argument is not "-n", echo now
  outputs all option-like arguments instead of treating them as options.

  expand and unexpand now conform to POSIX better.  They check for
  blanks (which can include characters other than space and tab in
  non-POSIX locales) instead of spaces and tabs.  Unexpand now
  preserves some blanks instead of converting them to tabs or spaces.

  "ln x d/" now reports an error if d/x is a directory and x a file,
  instead of incorrectly creating a link to d/x/x.

  ls no longer segfaults on systems for which SIZE_MAX != (size_t) -1.

  md5sum and sha1sum now report an error when given so many input
  lines that their line counter overflows, instead of silently
  reporting incorrect results.

  Fixes for "nice":

    If it fails to lower the niceness due to lack of permissions,
    it goes ahead and runs the command anyway, as POSIX requires.

    It no longer incorrectly reports an error if the current niceness
    happens to be -1.

    It no longer assumes that nicenesses range from -20 through 19.

    It now consistently adjusts out-of-range nicenesses to the
    closest values in range; formerly it sometimes reported an error.

  pathchk no longer accepts trailing options, e.g., "pathchk -p foo -b"
  now treats -b as a file name to check, not as an invalid option.

  `pr --columns=N' was not equivalent to `pr -N' when also using
  either -s or -w.

  pr now supports page numbers up to 2**64 on most hosts, and it
  detects page number overflow instead of silently wrapping around.
  pr now accepts file names that begin with "+" so long as the rest of
  the file name does not look like a page range.

  printf has several changes:

    It now uses 'intmax_t' (not 'long int') to format integers, so it
    can now format 64-bit integers on most modern hosts.

    On modern hosts it now supports the C99-inspired %a, %A, %F conversion
    specs, the "'" and "0" flags, and the ll, j, t, and z length modifiers
    (this is compatible with recent Bash versions).

    The printf command now rejects invalid conversion specifications
    like %#d, instead of relying on undefined behavior in the underlying
    printf function.

  ptx now diagnoses invalid values for its --width=N (-w)
  and --gap-size=N (-g) options.

  mv (when moving between partitions) no longer fails when
  operating on too many command-line-specified nonempty directories.

  "readlink -f" is more compatible with prior implementations

  rm (without -f) no longer hangs when attempting to remove a symlink
  to a file on an off-line NFS-mounted partition.

  rm no longer gets a failed assertion under some unusual conditions.

  rm no longer requires read access to the current directory.

  "rm -r" would mistakenly fail to remove files under a directory
  for some types of errors (e.g., read-only file system, I/O error)
  when first encountering the directory.

  "sort" fixes:

    "sort -o -" now writes to a file named "-" instead of to standard
    output; POSIX requires this.

    An unlikely race condition has been fixed where "sort" could have
    mistakenly removed a temporary file belonging to some other process.

    "sort" no longer has O(N**2) behavior when it creates many temporary files.

  tac can now handle regular, nonseekable files like Linux's
  /proc/modules.  Before, it would produce no output for such a file.

  tac would exit immediately upon I/O or temp-file creation failure.
  Now it continues on, processing any remaining command line arguments.

  "tail -f" no longer mishandles pipes and fifos.  With no operands,
  tail now ignores -f if standard input is a pipe, as POSIX requires.
  When conforming to POSIX 1003.2-1992, tail now supports the SUSv2 b
  modifier (e.g., "tail -10b file") and it handles some obscure cases
  more correctly, e.g., "tail +cl" now reads the file "+cl" rather
  than reporting an error, "tail -c file" no longer reports an error,
  and "tail - file" no longer reads standard input.

  tee now exits when it gets a SIGPIPE signal, as POSIX requires.
  To get tee's old behavior, use the shell command "(trap '' PIPE; tee)".
  Also, "tee -" now writes to standard output instead of to a file named "-".

  "touch -- MMDDhhmm[yy] file" is now equivalent to
  "touch MMDDhhmm[yy] file" even when conforming to pre-2001 POSIX.

  tr no longer mishandles a second operand with leading "-".

  who now prints user names in full instead of truncating them after 8 bytes.

  The following commands now reject unknown options instead of
  accepting them as operands, so that users are properly warned that
  options may be added later.  Formerly they accepted unknown options
  as operands; e.g., "basename -a a" acted like "basename -- -a a".

    basename dirname factor hostname link nohup sync unlink yes

** New features

  For efficiency, `sort -m' no longer copies input to a temporary file
  merely because the input happens to come from a pipe.  As a result,
  some relatively-contrived examples like `cat F | sort -m -o F - G'
  are no longer safe, as `sort' might start writing F before `cat' is
  done reading it.  This problem cannot occur unless `-m' is used.

  When outside the default POSIX locale, the 'who' and 'pinky'
  commands now output time stamps like "2004-06-21 13:09" instead of
  the traditional "Jun 21 13:09".

  pwd now works even when run from a working directory whose name
  is longer than PATH_MAX.

  cp, install, ln, and mv have a new --no-target-directory (-T) option,
  and -t is now a short name for their --target-directory option.

  cp -pu and mv -u (when copying) now don't bother to update the
  destination if the resulting time stamp would be no newer than the
  preexisting time stamp.  This saves work in the common case when
  copying or moving multiple times to the same destination in a file
  system with a coarse time stamp resolution.

  cut accepts a new option, --complement, to complement the set of
  selected bytes, characters, or fields.

  dd now also prints the number of bytes transferred, the time, and the
  transfer rate.  The new "status=noxfer" operand suppresses this change.

  dd has new conversions for the conv= option:

    nocreat   do not create the output file
    excl      fail if the output file already exists
    fdatasync physically write output file data before finishing
    fsync     likewise, but also write metadata

  dd has new iflag= and oflag= options with the following flags:

    append    append mode (makes sense for output file only)
    direct    use direct I/O for data
    dsync     use synchronized I/O for data
    sync      likewise, but also for metadata
    nonblock  use non-blocking I/O
    nofollow  do not follow symlinks
    noctty    do not assign controlling terminal from file

  stty now provides support (iutf8) for setting UTF-8 input mode.

  With stat, a specified format is no longer automatically newline terminated.
  If you want a newline at the end of your output, append `\n' to the format
  string.

  'df', 'du', and 'ls' now take the default block size from the
  BLOCKSIZE environment variable if the BLOCK_SIZE, DF_BLOCK_SIZE,
  DU_BLOCK_SIZE, and LS_BLOCK_SIZE environment variables are not set.
  Unlike the other variables, though, BLOCKSIZE does not affect
  values like 'ls -l' sizes that are normally displayed as bytes.
  This new behavior is for compatibility with BSD.

  du accepts a new option --files0-from=FILE, where FILE contains a
  list of NUL-terminated file names.

  Date syntax as used by date -d, date -f, and touch -d has been
  changed as follows:

    Dates like `January 32' with out-of-range components are now rejected.

    Dates can have fractional time stamps like 2004-02-27 14:19:13.489392193.

    Dates can be entered via integer counts of seconds since 1970 when
    prefixed by `@'.  For example, `@321' represents 1970-01-01 00:05:21 UTC.

    Time zone corrections can now separate hours and minutes with a colon,
    and can follow standard abbreviations like "UTC".  For example,
    "UTC +0530" and "+05:30" are supported, and are both equivalent to "+0530".

    Date values can now have leading TZ="..." assignments that override
    the environment only while that date is being processed.  For example,
    the following shell command converts from Paris to New York time:

      TZ="America/New_York" date --date='TZ="Europe/Paris" 2004-10-31 06:30'

  `date' has a new option --iso-8601=ns that outputs
  nanosecond-resolution time stamps.

  echo -e '\xHH' now outputs a byte whose hexadecimal value is HH,
  for compatibility with bash.

  ls now exits with status 1 on minor problems, 2 if serious trouble.

  ls has a new --hide=PATTERN option that behaves like
  --ignore=PATTERN, except that it is overridden by -a or -A.
  This can be useful for aliases, e.g., if lh is an alias for
  "ls --hide='*~'", then "lh -A" lists the file "README~".

  In the following cases POSIX allows the default GNU behavior,
  so when POSIXLY_CORRECT is set:

    false, printf, true, unlink, and yes all support --help and --option.
    ls supports TABSIZE.
    pr no longer depends on LC_TIME for the date format in non-POSIX locales.
    printf supports \u, \U, \x.
    tail supports two or more files when using the obsolete option syntax.

  The usual `--' operand is now supported by chroot, hostid, hostname,
  pwd, sync, and yes.

  `od' now conforms to POSIX better, and is more compatible with BSD:

    The older syntax "od [-abcdfilosx]... [FILE] [[+]OFFSET[.][b]]" now works
    even without --traditional.  This is a change in behavior if there
    are one or two operands and the last one begins with +, or if
    there are two operands and the latter one begins with a digit.
    For example, "od foo 10" and "od +10" now treat the last operand as
    an offset, not as a file name.

    -h is no longer documented, and may be withdrawn in future versions.
    Use -x or -t x2 instead.

    -i is now equivalent to -t dI (not -t d2), and
    -l is now equivalent to -t dL (not -t d4).

    -s is now equivalent to -t d2.  The old "-s[NUM]" or "-s NUM"
    option has been renamed to "-S NUM".

    The default output format is now -t oS, not -t o2, i.e., short int
    rather than two-byte int.  This makes a difference only on hosts like
    Cray systems where the C short int type requires more than two bytes.

  readlink accepts new options: --canonicalize-existing (-e)
  and --canonicalize-missing (-m).

  The stat option --filesystem has been renamed to --file-system, for
  consistency with POSIX "file system" and with cp and du --one-file-system.

** Removed features

  md5sum and sha1sum's undocumented --string option has been removed.

  tail's undocumented --max-consecutive-size-changes option has been removed.

* Major changes in release 5.2.1 (2004-03-12) [stable]

** Bug fixes

  mv could mistakenly fail to preserve hard links when moving two
  or more arguments between partitions.

  `cp --sparse=always F /dev/hdx' no longer tries to use lseek to create
  holes in the destination.

  nohup now sets the close-on-exec flag for its copy of the stderr file
  descriptor.  This avoids some nohup-induced hangs.  For example, before
  this change, if you ran `ssh localhost', then `nohup sleep 600 </dev/null &',
  and then exited that remote shell, the ssh session would hang until the
  10-minute sleep terminated.  With the fixed nohup, the ssh session
  terminates immediately.

  `expr' now conforms to POSIX better:

    Integers like -0 and 00 are now treated as zero.

    The `|' operator now returns 0, not its first argument, if both
    arguments are null or zero.  E.g., `expr "" \| ""' now returns 0,
    not the empty string.

    The `|' and `&' operators now use short-circuit evaluation, e.g.,
    `expr 1 \| 1 / 0' no longer reports a division by zero.

** New features

  `chown user.group file' now has its traditional meaning even when
  conforming to POSIX 1003.1-2001, so long as no user has a name
  containing `.' that happens to equal `user.group'.


* Major changes in release 5.2.0 (2004-02-19) [stable]

** Bug fixes

  none


* Major changes in release 5.1.3 (2004-02-08): candidate to become stable 5.2.0

** Bug fixes

  `cp -d' now works as required even on systems like OSF V5.1 that
  declare stat and lstat as `static inline' functions.

  time stamps output by stat now include actual fractional seconds,
  when available -- or .0000000 for files without that information.

  seq no longer infloops when printing 2^31 or more numbers.
  For reference, seq `echo 2^31|bc` > /dev/null takes about one hour
  on a 1.6 GHz Athlon 2000 XP.  Now it can output 2^53-1 numbers before
  misbehaving.

* Major changes in release 5.1.2 (2004-01-25):

** Bug fixes

  rmdir -p exits with status 1 on error; formerly it sometimes exited
  with status 0 when given more than one argument.

  nohup now always exits with status 127 when it finds an error,
  as POSIX requires; formerly it sometimes exited with status 1.

  Several programs (including cut, date, dd, env, hostname, nl, pr,
  stty, and tr) now always exit with status 1 when they find an error;
  formerly they sometimes exited with status 2.

  factor no longer reports a usage error if stdin has the wrong format.

  paste no longer infloops on ppc systems (bug introduced in 5.1.1)


* Major changes in release 5.1.1 (2004-01-17):

** Configuration option

  You can select the default level of POSIX conformance at configure-time,
  e.g., by ./configure DEFAULT_POSIX2_VERSION=199209

** Bug fixes

  fold -s works once again on systems with differing sizes for int
  and size_t (bug introduced in 5.1.0)

** New features

  touch -r now specifies the origin for any relative times in the -d
  operand, if both options are given.  For example, "touch -r FOO -d
  '-5 seconds' BAR" sets BAR's modification time to be five seconds
  before FOO's.

  join: The obsolete options "-j1 FIELD", "-j2 FIELD", and
  "-o LIST1 LIST2..." are no longer supported on POSIX 1003.1-2001 systems.
  Portable scripts should use "-1 FIELD", "-2 FIELD", and
  "-o LIST1,LIST2..." respectively.  If join was compiled on a
  POSIX 1003.1-2001 system, you may enable the old behavior
  by setting _POSIX2_VERSION=199209 in your environment.
  [This change was reverted in coreutils 5.3.1.]


* Major changes in release 5.1.0 (2003-12-21):

** New features

  chgrp, chmod, and chown can now process (with -R) hierarchies of virtually
  unlimited depth.  Before, they would fail to operate on any file they
  encountered with a relative name of length PATH_MAX (often 4096) or longer.

  chgrp, chmod, chown, and rm accept the new options:
  --preserve-root, --no-preserve-root (default)

  chgrp and chown now accept POSIX-mandated -L, -H, and -P options

  du can now process hierarchies of virtually unlimited depth.
  Before, du was limited by the user's stack size and it would get a
  stack overflow error (often a segmentation fault) when applied to
  a hierarchy of depth around 30,000 or larger.

  du works even when run from an inaccessible directory

  du -D now dereferences all symlinks specified on the command line,
  not just the ones that reference directories

  du now accepts -P (--no-dereference), for compatibility with du
  of NetBSD and for consistency with e.g., chown and chgrp

  du's -H option will soon have the meaning required by POSIX
  (--dereference-args, aka -D) rather then the current meaning of --si.
  Now, using -H elicits a warning to that effect.

  When given -l and similar options, ls now adjusts the output column
  widths to fit the data, so that output lines are shorter and have
  columns that line up better.  This may adversely affect shell
  scripts that expect fixed-width columns, but such shell scripts were
  not portable anyway, even with old GNU ls where the columns became
  ragged when a datum was too wide.

  du accepts a new option, -0/--null, to make it produce NUL-terminated
  output lines

** Bug fixes

  printf, seq, tail, and sleep now parse floating-point operands
  and options in the C locale.  POSIX requires this for printf.

  od -c -w9999999 no longer segfaults

  csplit no longer reads from freed memory (dumping core on some systems)

  csplit would mistakenly exhaust virtual memory in some cases

  ls --width=N (for very large N) is no longer subject to an address
  arithmetic bug that could result in bounds violations.

  ls --width=N (with -x or -C) no longer allocates more space
  (potentially much more) than necessary for a given directory.

  dd `unblock' and `sync' may now be combined (e.g., dd conv=unblock,sync)

* Major changes in release 5.0.91 (2003-09-08):

** New features

  date accepts a new option --rfc-2822, an alias for --rfc-822.

  split accepts a new option -d or --numeric-suffixes.

  cp, install, mv, and touch now preserve microsecond resolution on
  file timestamps, on platforms that have the 'utimes' system call.
  Unfortunately there is no system call yet to preserve file
  timestamps to their full nanosecond resolution; microsecond
  resolution is the best we can do right now.

  sort now supports the zero byte (NUL) as a field separator; use -t '\0'.
  The -t '' option, which formerly had no effect, is now an error.

  sort option order no longer matters for the options -S, -d, -i, -o, and -t.
  Stronger options override weaker, and incompatible options are diagnosed.

  `sha1sum --check' now accepts the BSD format for SHA1 message digests
  in addition to the BSD format for MD5 ones.

  who -l now means `who --login', not `who --lookup', per POSIX.
  who's -l option has been eliciting an unconditional warning about
  this impending change since sh-utils-2.0.12 (April 2002).

** Bug fixes

  Mistakenly renaming a file onto itself, e.g., via `mv B b' when `B' is
  the same directory entry as `b' no longer destroys the directory entry
  referenced by both `b' and `B'.  Note that this would happen only on
  file systems like VFAT where two different names may refer to the same
  directory entry, usually due to lower->upper case mapping of file names.
  Now, the above can happen only on file systems that perform name mapping and
  that support hard links (stat.st_nlink > 1).  This mitigates the problem
  in two ways: few file systems appear to be affected (hpfs and ntfs are),
  when the bug is triggered, mv no longer removes the last hard link to a file.
  *** ATTENTION ***: if you know how to distinguish the following two cases
  without writing to the file system in question, please let me know:
  1) B and b refer to the same directory entry on a file system like NTFS
       (B may well have a link count larger than 1)
  2) B and b are hard links to the same file

  stat no longer overruns a buffer for format strings ending in `%'

  fold -s -wN would infloop for N < 8 with TABs in the input.
  E.g., this would not terminate: printf 'a\t' | fold -w2 -s

  `split -a0', although of questionable utility, is accepted once again.

  `df DIR' used to hang under some conditions on OSF/1 5.1.  Now it doesn't.

  seq's --width (-w) option now works properly even when the endpoint
  requiring the larger width is negative and smaller than the other endpoint.

  seq's default step is 1, even if LAST < FIRST.

  paste no longer mistakenly outputs 0xFF bytes for a nonempty input file
  without a trailing newline.

  `tail -n0 -f FILE' and `tail -c0 -f FILE' no longer perform what amounted
  to a busy wait, rather than sleeping between iterations.

  tail's long-undocumented --allow-missing option now elicits a warning


* Major changes in release 5.0.90 (2003-07-29):

** New features

  sort is now up to 30% more CPU-efficient in some cases

  `test' is now more compatible with Bash and POSIX:

    `test -t', `test --help', and `test --version' now silently exit
    with status 0.  To test whether standard output is a terminal, use
    `test -t 1'.  To get help and version info for `test', use
    `[ --help' and `[ --version'.

    `test' now exits with status 2 (not 1) if there is an error.

  wc count field widths now are heuristically adjusted depending on the input
  size, if known.  If only one count is printed, it is guaranteed to
  be printed without leading spaces.

  Previously, wc did not align the count fields if POSIXLY_CORRECT was set,
  but POSIX did not actually require this undesirable behavior, so it
  has been removed.

** Bug fixes

  kill no longer tries to operate on argv[0] (introduced in 5.0.1)
  Why wasn't this noticed?  Although many tests use kill, none of
  them made an effort to avoid using the shell's built-in kill.

  `[' invoked with no arguments no longer evokes a segfault

  rm without --recursive (aka -r or -R) no longer prompts regarding
  unwritable directories, as required by POSIX.

  uniq -c now uses a SPACE, not a TAB between the count and the
  corresponding line, as required by POSIX.

  expr now exits with status 2 if the expression is syntactically valid,
  and with status 3 if an error occurred.  POSIX requires this.

  expr now reports trouble if string comparison fails due to a collation error.

  split now generates suffixes properly on EBCDIC hosts.

  split -a0 now works, as POSIX requires.

  `sort --version' and `sort --help' fail, as they should
  when their output is redirected to /dev/full.

  `su --version > /dev/full' now fails, as it should.

** Fewer arbitrary limitations

  cut requires 97% less memory when very large field numbers or
  byte offsets are specified.


* Major changes in release 5.0.1 (2003-07-15):

** New programs
- new program: `[' (much like `test')

** New features
- head now accepts --lines=-N (--bytes=-N) to print all but the
  N lines (bytes) at the end of the file
- md5sum --check now accepts the output of the BSD md5sum program, e.g.,
  MD5 (f) = d41d8cd98f00b204e9800998ecf8427e
- date -d DATE can now parse a DATE string like May-23-2003
- chown: `.' is no longer recognized as a separator in the OWNER:GROUP
  specifier on POSIX 1003.1-2001 systems.  If chown *was not* compiled
  on such a system, then it still accepts `.', by default.  If chown
  was compiled on a POSIX 1003.1-2001 system, then you may enable the
  old behavior by setting _POSIX2_VERSION=199209 in your environment.
- chown no longer tries to preserve set-user-ID and set-group-ID bits;
  on some systems, the chown syscall resets those bits, and previous
  versions of the chown command would call chmod to restore the original,
  pre-chown(2) settings, but that behavior is problematic.
  1) There was a window whereby a malicious user, M, could subvert a
  chown command run by some other user and operating on files in a
  directory where M has write access.
  2) Before (and even now, on systems with chown(2) that doesn't reset
  those bits), an unwary admin. could use chown unwittingly to create e.g.,
  a set-user-ID root copy of /bin/sh.

** Bug fixes
- chown --dereference no longer leaks a file descriptor per symlink processed
- `du /' once again prints the `/' on the last line
- split's --verbose option works once again [broken in 4.5.10 and 5.0]
- tail -f is no longer subject to a race condition that could make it
  delay displaying the last part of a file that had stopped growing.  That
  bug could also make tail -f give an unwarranted `file truncated' warning.
- du no longer runs out of file descriptors unnecessarily
- df and `readlink --canonicalize' no longer corrupt the heap on
  non-glibc, non-solaris systems
- `env -u UNSET_VARIABLE' no longer dumps core on non-glibc systems
- readlink's --canonicalize option now works on systems like Solaris that
  lack the canonicalize_file_name function but do have resolvepath.
- mv now removes `a' in this example on all systems: touch a; ln a b; mv a b
  This behavior is contrary to POSIX (which requires that the mv command do
  nothing and exit successfully), but I suspect POSIX will change.
- date's %r format directive now honors locale settings
- date's `-' (no-pad) format flag now affects the space-padded-by-default
  conversion specifiers, %e, %k, %l
- fmt now diagnoses invalid obsolescent width specifications like `-72x'
- fmt now exits nonzero when unable to open an input file
- tsort now fails when given an odd number of input tokens,
  as required by POSIX.  Before, it would act as if the final token
  appeared one additional time.

** Fewer arbitrary limitations
- tail's byte and line counts are no longer limited to OFF_T_MAX.
  Now the limit is UINTMAX_MAX (usually 2^64).
- split can now handle --bytes=N and --lines=N with N=2^31 or more.

** Portability
- `kill -t' now prints signal descriptions (rather than `?') on systems
  like Tru64 with __sys_siglist but no strsignal function.
- stat.c now compiles on Ultrix systems
- sleep now works on AIX systems that lack support for clock_gettime
- rm now works around Darwin6.5's broken readdir function
  Before `rm -rf DIR' would fail to remove all files in DIR
  if there were more than 338.

* Major changes in release 5.0 (2003-04-02):
- false --help now exits nonzero

[4.5.12]
* printf no longer treats \x specially when POSIXLY_CORRECT is set
* printf avoids buffer overrun with format ending in a backslash and
* printf avoids buffer overrun with incomplete conversion specifier
* printf accepts multiple flags in a single conversion specifier

[4.5.11]
* seq no longer requires that a field width be specified
* seq no longer fails when given a field width of `0'
* seq now accepts ` ' and `'' as valid format flag characters
* df now shows a HOSTNAME: prefix for each remote-mounted file system on AIX 5.1
* portability tweaks for HP-UX, AIX 5.1, DJGPP

[4.5.10]
* printf no longer segfaults for a negative field width or precision
* shred now always enables --exact for non-regular files
* du no longer lists hard-linked files more than once
* du no longer dumps core on some systems due to `infinite' recursion
  via nftw's use of the buggy replacement function in getcwd.c
* portability patches for a few vendor compilers and 64-bit systems
* du -S *really* now works like it did before the change in 4.5.5

[4.5.9]
* du no longer truncates file sizes or sums to fit in 32-bit size_t
* work around Linux kernel bug in getcwd (fixed in 2.4.21-pre4), so that pwd
  now fails if the name of the working directory is so long that getcwd
  truncates it.  Before it would print the truncated name and exit successfully.
* `df /some/mount-point' no longer hangs on a GNU libc system when another
  hard-mounted NFS file system (preceding /some/mount-point in /proc/mounts)
  is inaccessible.
* rm -rf now gives an accurate diagnostic when failing to remove a file
  under certain unusual conditions
* mv and `cp --preserve=links' now preserve multiple hard links even under
  certain unusual conditions where they used to fail

[4.5.8]
* du -S once again works like it did before the change in 4.5.5
* stat accepts a new file format, %B, for the size of each block reported by %b
* du accepts new option: --apparent-size
* du --bytes (-b) works the same way it did in fileutils-3.16 and before
* du reports proper sizes for directories (not zero) (broken in 4.5.6 or 4.5.7)
* df now always displays under `Filesystem', the device file name
  corresponding to the listed mount point.  Before, for a block- or character-
  special file command line argument, df would display that argument.  E.g.,
  `df /dev/hda' would list `/dev/hda' as the `Filesystem', rather than say
  /dev/hda3 (the device on which `/' is mounted), as it does now.
* test now works properly when invoked from a set user ID or set group ID
  context and when testing access to files subject to alternate protection
  mechanisms.  For example, without this change, a set-UID program that invoked
  `test -w F' (to see if F is writable) could mistakenly report that it *was*
  writable, even though F was on a read-only file system, or F had an ACL
  prohibiting write access, or F was marked as immutable.

[4.5.7]
* du would fail with more than one DIR argument when any but the last did not
  contain a slash (due to a bug in ftw.c)

[4.5.6]
* du no longer segfaults on Solaris systems (fixed heap-corrupting bug in ftw.c)
* du --exclude=FILE works once again (this was broken by the rewrite for 4.5.5)
* du no longer gets a failed assertion for certain hierarchy lay-outs
  involving hard-linked directories
* `who -r' no longer segfaults when using non-C-locale messages
* df now displays a mount point (usually `/') for non-mounted
  character-special and block files

[4.5.5]
* ls --dired produces correct byte offset for file names containing
  nonprintable characters in a multibyte locale
* du has been rewritten to use a variant of GNU libc's ftw.c
* du now counts the space associated with a directory's directory entry,
  even if it cannot list or chdir into that subdirectory.
* du -S now includes the st_size of each entry corresponding to a subdirectory
* rm on FreeBSD can once again remove directories from NFS-mounted file systems
* ls has a new option --dereference-command-line-symlink-to-dir, which
  corresponds to the new default behavior when none of -d, -l -F, -H, -L
  has been specified.
* ls dangling-symlink now prints `dangling-symlink'.
  Before, it would fail with `no such file or directory'.
* ls -s symlink-to-non-dir and ls -i symlink-to-non-dir now print
  attributes of `symlink', rather than attributes of their referents.
* Fix a bug introduced in 4.5.4 that made it so that ls --color would no
  longer highlight the names of files with the execute bit set when not
  specified on the command line.
* shred's --zero (-z) option no longer gobbles up any following argument.
  Before, `shred --zero file' would produce `shred: missing file argument',
  and worse, `shred --zero f1 f2 ...' would appear to work, but would leave
  the first file untouched.
* readlink: new program
* cut: new feature: when used to select ranges of byte offsets (as opposed
  to ranges of fields) and when --output-delimiter=STRING is specified,
  output STRING between ranges of selected bytes.
* rm -r can no longer be tricked into mistakenly reporting a cycle.
* when rm detects a directory cycle, it no longer aborts the entire command,
  but rather merely stops processing the affected command line argument.

[4.5.4]
* cp no longer fails to parse options like this: --preserve=mode,ownership
* `ls --color -F symlink-to-dir' works properly
* ls is much more efficient on directories with valid dirent.d_type.
* stty supports all baud rates defined in linux-2.4.19.
* `du symlink-to-dir/' would improperly remove the trailing slash
* `du ""' would evoke a bounds violation.
* In the unlikely event that running `du /' resulted in `stat ("/", ...)'
  failing, du would give a diagnostic about `' (empty string) rather than `/'.
* printf: a hexadecimal escape sequence has at most two hex. digits, not three.
* The following features have been added to the --block-size option
  and similar environment variables of df, du, and ls.
  - A leading "'" generates numbers with thousands separators.
    For example:
      $ ls -l --block-size="'1" file
      -rw-rw-r--    1 eggert   src      47,483,707 Sep 24 23:40 file
  - A size suffix without a leading integer generates a suffix in the output.
    For example:
      $ ls -l --block-size="K"
      -rw-rw-r--    1 eggert   src          46371K Sep 24 23:40 file
* ls's --block-size option now affects file sizes in all cases, not
  just for --block-size=human-readable and --block-size=si.  Fractional
  sizes are now always rounded up, for consistency with df and du.
* df now displays the block size using powers of 1000 if the requested
  block size seems to be a multiple of a power of 1000.
* nl no longer gets a segfault when run like this `yes|nl -s%n'

[4.5.3]
* du --dereference-args (-D) no longer fails in certain cases
* `ln --target-dir=DIR' no longer fails when given a single argument

[4.5.2]
* `rm -i dir' (without --recursive (-r)) no longer recurses into dir
* `tail -c N FILE' now works with files of size >= 4GB
* `mkdir -p' can now create very deep (e.g. 40,000-component) directories
* rmdir -p dir-with-trailing-slash/ no longer fails
* printf now honors the `--' command line delimiter
* od's 8-byte formats x8, o8, and u8 now work
* tail now accepts fractional seconds for its --sleep-interval=S (-s) option

[4.5.1]
* du and ls now report sizes of symbolic links (before they'd always report 0)
* uniq now obeys the LC_COLLATE locale, as per POSIX 1003.1-2001 TC1.

========================================================================
Here are the NEWS entries made from fileutils-4.1 until the
point at which the packages merged to form the coreutils:

[4.1.11]
* `rm symlink-to-unwritable' doesn't prompt [introduced in 4.1.10]
[4.1.10]
* rm once again gives a reasonable diagnostic when failing to remove a file
    owned by someone else in a sticky directory [introduced in 4.1.9]
* df now rounds all quantities up, as per POSIX.
* New ls time style: long-iso, which generates YYYY-MM-DD HH:MM.
* Any time style can be preceded by "posix-"; this causes "ls" to
  use traditional timestamp format when in the POSIX locale.
* The default time style is now posix-long-iso instead of posix-iso.
  Set TIME_STYLE="posix-iso" to revert to the behavior of 4.1.1 thru 4.1.9.
* `rm dangling-symlink' doesn't prompt [introduced in 4.1.9]
* stat: remove support for --secure/-s option and related %S and %C format specs
* stat: rename --link/-l to --dereference/-L.
    The old options will continue to work for a while.
[4.1.9]
* rm can now remove very deep hierarchies, in spite of any limit on stack size
* new programs: link, unlink, and stat
* New ls option: --author (for the Hurd).
* `touch -c no-such-file' no longer fails, per POSIX
[4.1.8]
* mv no longer mistakenly creates links to preexisting destination files
    that aren't moved
[4.1.7]
* rm: close a hole that would allow a running rm process to be subverted
[4.1.6]
* New cp option: --copy-contents.
* cp -r is now equivalent to cp -R.  Use cp -R -L --copy-contents to get the
  traditional (and rarely desirable) cp -r behavior.
* ls now accepts --time-style=+FORMAT, where +FORMAT works like date's format
* The obsolete usage `touch [-acm] MMDDhhmm[YY] FILE...' is no longer
  supported on systems conforming to POSIX 1003.1-2001.  Use touch -t instead.
* cp and inter-partition mv no longer give a misleading diagnostic in some
    unusual cases
[4.1.5]
* cp -r no longer preserves symlinks
* The block size notation is now compatible with SI and with IEC 60027-2.
  For example, --block-size=1MB now means --block-size=1000000,
  whereas --block-size=1MiB now means --block-size=1048576.
  A missing `B' (e.g. `1M') has the same meaning as before.
  A trailing `B' now means decimal, not binary; this is a silent change.
  The nonstandard `D' suffix (e.g. `1MD') is now obsolescent.
* -H or --si now outputs the trailing 'B', for consistency with the above.
* Programs now output trailing 'K' (not 'k') to mean 1024, as per IEC 60027-2.
* New df, du short option -B is short for --block-size.
* You can omit an integer `1' before a block size suffix,
  e.g. `df -BG' is equivalent to `df -B 1G' and to `df --block-size=1G'.
* The following options are now obsolescent, as their names are
  incompatible with IEC 60027-2:
   df, du: -m or --megabytes (use -BM or --block-size=1M)
   df, du, ls: --kilobytes (use --block-size=1K)
[4.1.4]
* df --local no longer lists smbfs file systems whose name starts with //
* dd now detects the Linux/tape/lseek bug at run time and warns about it.
[4.1.3]
* ls -R once again outputs a blank line between per-directory groups of files.
    This was broken by the cycle-detection change in 4.1.1.
* dd once again uses `lseek' on character devices like /dev/mem and /dev/kmem.
    On systems with the linux kernel (at least up to 2.4.16), dd must still
    resort to emulating `skip=N' behavior using reads on tape devices, because
    lseek has no effect, yet appears to succeed.  This may be a kernel bug.
[4.1.2]
* cp no longer fails when two or more source files are the same;
    now it just gives a warning and doesn't copy the file the second time.
    E.g., cp a a d/ produces this:
    cp: warning: source file `a' specified more than once
* chmod would set the wrong bit when given symbolic mode strings like
    these: g=o, o=g, o=u.  E.g., `chmod a=,o=w,ug=o f' would give a mode
    of --w-r---w- rather than --w--w--w-.
[4.1.1]
* mv (likewise for cp), now fails rather than silently clobbering one of
    the source files in the following example:
    rm -rf a b c; mkdir a b c; touch a/f b/f; mv a/f b/f c
* ls -R detects directory cycles, per POSIX.  It warns and doesn't infloop.
* cp's -P option now means the same as --no-dereference, per POSIX.
    Use --parents to get the old meaning.
* When copying with the -H and -L options, cp can preserve logical
    links between source files with --preserve=links
* cp accepts new options:
    --preserve[={mode,ownership,timestamps,links,all}]
    --no-preserve={mode,ownership,timestamps,links,all}
* cp's -p and --preserve options remain unchanged and are equivalent
    to `--preserve=mode,ownership,timestamps'
* mv and cp accept a new option: --reply={yes,no,query};  provides a consistent
    mechanism to control whether one is prompted about certain existing
    destination files.  Note that cp's and mv's -f options don't have the
    same meaning: cp's -f option no longer merely turns off `-i'.
* remove portability limitations (e.g., PATH_MAX on the Hurd, fixes for
    64-bit systems)
* mv now prompts before overwriting an existing, unwritable destination file
    when stdin is a tty, unless --force (-f) is specified, as per POSIX.
* mv: fix the bug whereby `mv -uf source dest' would delete source,
    even though it's older than dest.
* chown's --from=CURRENT_OWNER:CURRENT_GROUP option now works
* cp now ensures that the set-user-ID and set-group-ID bits are cleared for
    the destination file when when copying and not preserving permissions.
* `ln -f --backup k k' gives a clearer diagnostic
* ls no longer truncates user names or group names that are longer
    than 8 characters.
* ls's new --dereference-command-line option causes it to dereference
  symbolic links on the command-line only.  It is the default unless
  one of the -d, -F, or -l options are given.
* ls -H now means the same as ls --dereference-command-line, as per POSIX.
* ls -g now acts like ls -l, except it does not display owner, as per POSIX.
* ls -n now implies -l, as per POSIX.
* ls can now display dates and times in one of four time styles:

  - The `full-iso' time style gives full ISO-style time stamps like
    `2001-05-14 23:45:56.477817180 -0700'.
  - The 'iso' time style gives ISO-style time stamps like '2001-05-14 '
    and '05-14 23:45'.
  - The 'locale' time style gives locale-dependent time stamps like
    'touko  14  2001' and 'touko  14 23:45' (in a Finnish locale).
  - The 'posix-iso' time style gives traditional POSIX-locale
    time stamps like 'May 14  2001' and 'May 14 23:45' unless the user
    specifies a non-POSIX locale, in which case it uses ISO-style dates.
    This is the default.

  You can specify a time style with an option like --time-style='iso'
  or with an environment variable like TIME_STYLE='iso'.  GNU Emacs 21
  and later can parse ISO dates, but older Emacs versions cannot, so
  if you are using an older version of Emacs outside the default POSIX
  locale, you may need to set TIME_STYLE="locale".

* --full-time is now an alias for "-l --time-style=full-iso".


========================================================================
Here are the NEWS entries made from sh-utils-2.0 until the
point at which the packages merged to form the coreutils:

 [2.0.15]
* date no longer accepts e.g., September 31 in the MMDDhhmm syntax
* fix a bug in this package's .m4 files and in configure.ac
 [2.0.14]
* nohup's behavior is changed as follows, to conform to POSIX 1003.1-2001:
  - nohup no longer adjusts scheduling priority; use "nice" for that.
  - nohup now redirects stderr to stdout, if stderr is not a terminal.
  - nohup exit status is now 126 if command was found but not invoked,
    127 if nohup failed or if command was not found.
 [2.0.13]
* uname and uptime work better on *BSD systems
* pathchk now exits nonzero for a path with a directory component
    that specifies a non-directory
 [2.0.12]
* kill: new program
* who accepts new options: --all (-a), --boot (-b), --dead (-d), --login,
   --process (-p), --runlevel (-r), --short (-s), --time (-t), --users (-u).
   The -u option now produces POSIX-specified results and is the same as
   the long option `--users'.  --idle is no longer the same as -u.
* The following changes apply on systems conforming to POSIX 1003.1-2001:
   - `date -I' is no longer supported.  Instead, use `date --iso-8601'.
   - `nice -NUM' is no longer supported.  Instead, use `nice -n NUM'.
  [This change was reverted in coreutils 5.3.1.]
* New 'uname' options -i or --hardware-platform, and -o or --operating-system.
   'uname -a' now outputs -i and -o information at the end.
   New uname option --kernel-version is an alias for -v.
   Uname option --release has been renamed to --kernel-release,
   and --sysname has been renamed to --kernel-name;
   the old options will work for a while, but are no longer documented.
* 'expr' now uses the LC_COLLATE locale for string comparison, as per POSIX.
* 'expr' now requires '+' rather than 'quote' to quote tokens;
    this removes an incompatibility with POSIX.
* date -d 'last friday' would print a date/time that was one hour off
    (e.g., 23:00 on *thursday* rather than 00:00 of the preceding friday)
    when run such that the current time and the target date/time fall on
    opposite sides of a daylight savings time transition.
    This problem arose only with relative date strings like `last monday'.
    It was not a problem with strings that include absolute dates.
* factor is twice as fast, for large numbers
 [2.0.11]
* setting the date now works properly, even when using -u
* `date -f - < /dev/null' no longer dumps core
* some DOS/Windows portability changes
 [2.0j]
* `date -d DATE' now parses certain relative DATEs correctly
 [2.0i]
* fixed a bug introduced in 2.0h that made many programs fail with a
 `write error' when invoked with the --version option
 [2.0h]
* all programs fail when printing --help or --version output to a full device
* printf exits nonzero upon write failure
* yes now detects and terminates upon write failure
* date --rfc-822 now always emits day and month names from the `C' locale
* portability tweaks for Solaris8, Ultrix, and DOS
 [2.0g]
* date now handles two-digit years with leading zeros correctly.
* printf interprets unicode, \uNNNN \UNNNNNNNN, on systems with the
  required support;  from Bruno Haible.
* stty's rprnt attribute now works on HPUX 10.20
* seq's --equal-width option works more portably
 [2.0f]
* fix build problems with ut_name vs. ut_user
 [2.0e]
* stty: fix long-standing bug that caused test failures on at least HPUX
  systems when COLUMNS was set to zero
* still more portability fixes
* unified lib/: now that directory and most of the configuration framework
  is common between fileutils, textutils, and sh-utils
 [2.0d]
* fix portability problem with sleep vs lib/strtod.c's requirement for -lm
 [2.0c]
* fix portability problems with nanosleep.c and with the new code in sleep.c
 [2.0b]
* Regenerate lib/Makefile.in so that nanosleep.c is distributed.
 [2.0a]
* sleep accepts floating point arguments on command line
* sleep's clock continues counting down when sleep is suspended
* when a suspended sleep process is resumed, it continues sleeping if
  there is any time remaining
* who once again prints whatever host information it has, even without --lookup

========================================================================
For older NEWS entries for the fileutils, textutils, and sh-utils
packages, see ./old/*/NEWS.

  This package began as the union of the following:
  textutils-2.1, fileutils-4.1.11, sh-utils-2.0.15.

========================================================================

Copyright (C) 2001-2011 Free Software Foundation, Inc.

Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.3 or
any later version published by the Free Software Foundation; with no
Invariant Sections, with no Front-Cover Texts, and with no Back-Cover
Texts.  A copy of the license is included in the ``GNU Free
Documentation License'' file as part of this distribution.