* src/copy.c (extent_copy): New arg SCAN, replacing
REQUIRE_NORMAL_COPY. All callers changed.
(enum scantype): New type.
(infer_scantype): Rename from is_probably_sparse and return
the new type. Add args FD and SCAN. All callers changed.
Run "make update-copyright" and then...
* gnulib: Update to latest with copyright year adjusted.
* tests/init.sh: Sync with gnulib to pick up copyright year.
* bootstrap: Likewise.
* tests/sample-test: Adjust to use the single most recent year.
Problem reported by Jakub Kulik (Bug#35713).
* NEWS: Mention this.
* configure.ac (DEV_FD_MIGHT_BE_CHR): New macro.
* src/copy.c (DEV_FD_MIGHT_BE_CHR): Default to false.
(follow_fstatat): New function.
(copy_internal): Use it.
* src/copy.h (XSTAT): Remove; no longer used.
Run "make update-copyright" and then...
* gnulib: Update to latest with copyright year adjusted.
* tests/init.sh: Sync with gnulib to pick up copyright year.
* bootstrap: Likewise.
* tests/sample-test: Adjust to use the single most recent year.
Previously, 'ln A B' did 'stat("B"), lstat("A"), link("A","B")'
where the stat and lstat were necessary to avoid hard-linking
directories on systems that can hard-link directories.
Now, in situations that prohibit hard links to directories,
'ln A B' merely does 'link("A","B")'. The new behavior
avoids some races and should be more efficient.
This patch was inspired by Bug#10020, which was about 'ln'.
* bootstrap.conf (gnulib_modules): Add unlinkdir.
* src/force-link.c (force_linkat, force_symlinkat): New arg for
error number of previous try. Return error number, 0, or -1 if
error, success, or success after removal. All callers changed.
* src/ln.c: Include priv-set.h, unlinkdir.h.
(beware_hard_dir_link): New static var.
(errnoize, atomic_link): New functions.
(target_directory_operand): Use errnoize for simplicity.
(do_link): New arg for error number of previous try. All callers
changed. Do each link atomically if possible.
(main): Do -r check earlier. Remove linkdir privileges so we can
use a single linkat/symlinkat instead of a racy substitute for the
common case of 'ln A B' and 'ln -s A B'. Set beware_hard_dir_link
to disable this optimization.
* src/copy.c (copy_internal): Use the lint protected src_mode,
rather than accessing the src_sb again. Also unconditionally
populate src_sb when !x->move_mode and in lint mode.
Reported by Kamil Dudka with coverity and clang analyzer.
This issue was introduced in commit v8.19-145-g24ebca6
* src/copy.c (copy_internal): With --no-preserve=mode,
only reset permissions for newly created files.
(copy_reg): Likewise.
* NEWS: Mention the fix.
* tests/cp/preserve-mode.sh: Add a test case.
Fixes https://bugs.gnu.org/31675
* src/copy.c (copy_internal): Don't fail immediately upon
getting ELOOP when running stat() on the destination,
rather proceeding if -f specified, allowing the link
to be removed. If the loop is not in the final component
of the destination path, we still fail but at the
subsequent unlink() stage.
* doc/coreutils.texi (cp invocation): Adjust wording to say
that --force doesn't work with dangling links, rather than
all links that can't be traversed.
* tests/cp/thru-dangling.sh: Add a test case.
* NEWS: Mention the change in behavior.
Discussed in https://bugs.gnu.org/31335
Ensure this _does_ recreate the symlink
Given "path1" and "path2" are on different devices.
$ touch "path1/file"
$ cd path2/; ln -s path1/file
$ cp -dsf path1/file .
Ensure this does _not_ overwrite file
$ touch file
$ ln -s file l1
$ cp -sf l1 file
* src/copy.c (same_file_ok): Remove device ids from consideration,
instead deferring to future EXDEV with --link or allowing
the first case above to work.
Also ensure that we do not exist this function too early,
when the destination file is not a symlink, which protects
against the second case.
* tests/cp/cross-dev-symlink.sh: Add a test for the first case.
* tests/cp/same-file.sh: Add a test for the second case above.
* NEWS: Mention the bug fixes.
* THANKS.in: Mention the reporters who also analyzed the code.
Fixes https://bugs.gnu.org/31364
This issue was introduced in commit v8.19-145-g24ebca6
* src/copy.c (copy_internal): When setting default permissions
to use with --no-preserve=mode, only set executable bits for
directories or sockets.
* NEWS: Mention the fix.
* tests/cp/preserve-mode.sh: Add a test case.
Fixes https://bugs.gnu.org/30534
This builds on a previous patch for mv atomicity (Bug#29961).
It merely improves performance; it does not fix bugs.
* src/copy.h (struct cp_options): New members last_file, rename_errno.
* src/copy.c (copy_internal): Support new rename_errno member
for the copy options. Avoid calling stat when new members
suggest it’s not needed.
(cp_options_default): Initialize new members.
* src/mv.c: Include renameat2.h.
(main): With two arguments, first call ‘renamat2 (AT_FDCWD, "a",
AT_FDCWD, "b", RENAME_NOREPLACE)’. Use its results to skip
remaining processing if possible; for example, if it succeeds
there is no need to stat either "a" or "b". Also, set
x.last_file when it is the last file to rename.
Problem reported by Kamil Dudka (Bug#29961).
* NEWS: Mention this.
* src/copy.c: Include renameat2.h.
(copy_internal): If mv, try renameat2 first thing, with
RENAME_NOREPLACE. If this works, skip most of the remaining code.
Also, fail quickly if it fails with EEXIST, and we are using -n.
Run "make update-copyright" and then...
* gnulib: Update to latest with copyright year adjusted.
* tests/init.sh: Sync with gnulib to pick up copyright year.
* bootstrap: Likewise.
* tests/sample-test: Adjust to use the single most recent year.
Suggested by Kamil Dudka in:
http://lists.gnu.org/archive/html/coreutils/2017-07/msg00072.html
* NEWS: Document the changed nature of the fix.
* doc/coreutils.texi, tests/cp/backup-is-src.sh:
* tests/mv/backup-is-src.sh: Revert previous change.
* src/copy.c (source_is_dst_backup): New function.
(copy_internal): Use it. Fail instead of falling back on numbered
backups when it looks like the backup will overwrite the source.
Although this reintroduces a race, it's more compatible with
previous behavior.
* NEWS, doc/coreutils.texi (Backup options): Document the change.
* bootstrap.conf (gnulib_modules): Add backup-rename.
* src/copy.c (copy_internal): Silently switch to numbered backups
if a simple backup might lose data. Use backup_file_rename
to avoid races with numbered backups.
* tests/cp/backup-is-src.sh, tests/mv/backup-is-src.sh:
Adjust to match new behavior.
* src/copy.c (copy_internal): Honor the x->require_preserve flag
for symlinks as we do for ordinary files, so we don't exit with
failure upon failure to chown a symbolic link.
* NEWS: Mention the bug fix.
* src/copy.c (set_process_security_ctx, set_file_security_ctx):
Export for use in cp.c.
* src/copy.h: Likewise.
* src/cp.c (make_dir_parents_private): Call the exported functions
to set the security context for new and updated directories.
* tests/cp/cp-a-selinux.sh: Add a test case.
Fixes http://bugs.gnu.org/25378
If the file B already exists, commands like 'ln -f A B' and
'cp -fl A B' no longer remove B before creating the new link.
Instead, they arrange for the new link to replace B atomically.
This should fix a race condition reported by Mike Crowe (Bug#25680).
* NEWS, doc/coreutils.texi (cp invocation, ln invocation):
Document this.
* bootstrap.conf (gnulib_modules): Add symlinkat.
* src/copy.c, src/ln.c: Include force-link.h.
* src/copy.c (same_file_ok): It's also OK to remove a destination
symlink when creating symbolic links, or when the source and
destination are on the same file system and when creating hard links.
* src/copy.c (create_hard_link, copy_internal):
* src/ln.c (do_link):
Rewrite using force_linkat and force_symlinkat, to close a window
where the destination temporarily does not exist.
* src/cp.c (main): Do not set x.unlink_dest_before_opening
merely because we are in link-creation mode.
* src/force-link.c, src/force-link.h: New files.
* src/local.mk (copy_sources, src_ln_SOURCES): Add them.
* tests/cp/same-file.sh: Adjust test case to match fixed behavior.
* src/copy.c (punch_hole): Work around an empty definition
of HAVE_FALLOCATE which leads to a build error of:
"error: #if with no expression"
That was triggered by the inclusion of <linux/fs.h> in
commit v8.25-68-g89e1fef with kernel-headers-2.6.18.
Reported by Nelson H. F. Beebe
Run "make update-copyright" and then...
* gnulib: Update to latest with copyright year adjusted.
* tests/init.sh: Sync with gnulib to pick up copyright year.
* bootstrap: Likewise.
* tests/sample-test: Adjust to use the single most recent year.
* src/copy.h (cp_options): Add a new flag for install(1).
* src/copy.c (copy_internal): For cp, also output remediation
advice which also indicates why directories aren't copied by default.
The message is unchanged for install(1).
* src/cp.c (cp_option_init): Init install_mode to false.
* src/mv.c (cp_option_init): Likewise.
* src/install.c (cp_option_init): Init install_mode to true.
* tests/install/basic-1.sh: Add a test case.
* tests/cp/link-deref.sh: Adjust test case.
Fixes http://bugs.gnu.org/24958
die() has the advantage of being apparent to the compiler
that it doesn't return, which will avoid warnings in some cases,
and possibly generate better code.
* cfg.mk (sc_die_EXIT_FAILURE): A new syntax check rule to
catch any new uses of error (CONSTANT, ...);
This doesn't affect the generated code on my system; it's merely
a cleaner way to use the recently-introduced Linux-specific API.
* m4/jm-macros.m4 (gl_CHECK_ALL_HEADERS): Check for linux/fs.h.
* src/copy.c: Include <linux.fs.h> if available.
(FICLONE) [__linux__]: Define if not already defined.
(clone_file): Use FICLONE instead of the older BTRFS_IOC_CLONE,
as this ioctl is no longer btrfs-specific.
* src/copy.c (copy internal): Remember directories irrespective
of their link count, because on some file systems like BTRFS,
directories always have a link count of 1.
commit v8.23-31-g90aa291 failed to consider this case,
where the previous rename has failed, thus causing the
following to remove the specified directory:
mv dir dir dir
* src/copy.c (copy_internal): Assume this rename attempt has
succeeded, as a previous failure will already have been handled,
and we don't want to remove the source directory in this case.
* tests/cp/duplicate-sources.sh: Consolidate this test file to...
* tests/mv/dup-source.sh: ...here. Add test cases for same
source and dest.
* tests/local.mk: Remove the consolidated test.
* NEWS: Mention the bug fix.
Reported at https://bugzilla.redhat.com/1297464
Run "make update-copyright" and then...
* gnulib: Update to latest with copyright year adjusted.
* tests/init.sh: Sync with gnulib to pick up copyright year.
* bootstrap: Likewise.
* tests/sample-test: Adjust to use the single most recent year.
fallocate can allocate extents beyond EOF via FALLOC_FL_KEEP_SIZE.
Where there is a gap (hole) between the extents, and EOF is within
that gap, the final hole wasn't reproduced, resulting in silent
data corruption in the copied file (size too small).
* src/copy.c (extent_copy): Ensure we don't process extents
beyond the apparent file size, since processing and allocating
those is not currently supported.
* tests/cp/fiemap-extents.sh: Renamed from tests/cp/fiemap-empty.sh
and re-enable parts checking the extents at and beyond EOF.
* tests/local.mk: Reference the renamed test.
* NEWS: Mention the bug fix.
Fixes http://bugs.gnu.org/21790
Quote file names using the "shell-escape" or "shell-escape-always"
methods, which quote as appropriate for most shells,
and better support copy and paste of presented names.
The "always" variant is used when the file name is
embedded in an error message with surrounding spaces.
* cfg.mk (sc_error_shell_quotes): A new syntax check rule
to suggest quotef() where appropriate.
(sc_error_shell_always_quotes): Likewise for quoteaf().
* src/system.h (quotef): A new define to apply shell quoting
when needed. I.E. when shell character or ':' is present.
(quoteaf): Likewise, but always quote.
* src/*.c: Use quotef() and quoteaf() rather than quote()
where appropriate.
* tests/: Adjust accordingly.
* src/factor.c: Move LIKELY() definition to...
* src/system.h: ...here.
(is_nul): Reimplement with a version that doesn't
require a sentinel after the buffer, and which calls
down to (the system optimized) memcmp.
Performance analyzed at http://rusty.ozlabs.org/?p=560
* src/dd.c (alloc_obuf): Simplify the is_nul() call by
not needing to write the sentinel.
* src/copy.c (sparse_copy): Likewise.
(copy_reg): Simplify the buffer allocation by avoiding
consideration of the sentinel in the buffer size calculation.
* src/copy.c (CAN_HARDLINK_SYMLINKS): Don't enable use of linkat()
on Darwin 14, as the gnulib fallback emulation there doesn't
preserve ownership and timestamps etc. This fixes a test failure
in tests/cp/link-symlink.sh
* tests/cp/link-deref.sh: Adjust accordingly.
Without this change, very recent gcc (e.g., version 6.0.0 20150509)
would print the following when configured with --enable-gcc-warnings:
src/copy.c:165:30: error: logical 'or' of equal expressions \
[-Werror=logical-op]
&& (errno == EOPNOTSUPP || errno == ENOTSUP || errno == ENOSYS))
^
* src/system.h (is_ENOTSUP): New function.
* src/copy.c (punch_hole): Use it.
* src/ls.c (errno_unsupported, gobble_file): Use it.
Run "make update-copyright" and then...
* tests/sample-test: Adjust to use the single most recent year.
* tests/du/bind-mount-dir-cycle-v2.sh: Fix case in copyright message,
so that year is updated automatically in future.
* NEWS: Update the recent entry to also mention the avoidance
of incorrectly unlinking a multi-hardlinked "source" file when
presented with source and dest that only differ in case.
* src/copy.c (same_file_ok): Mention the case issue with same_name().
* tests/mv/hardlink-case.sh: Test the issue on HFS+.
* tests/local.mk: Reference the new test case.
* tests/mv/vfat: Remove an old related but unused test case.
We may run into a race condition if we treat hard links to the same file
as distinct files. If we do 'mv a b' and 'mv b a' in parallel, both a
and b can disappear from the file system. The reason is that in this
case the unlink on src is called and the system calls can end up being
run in the order where unlink(a) and unlink(b) are the last two system
calls. Therefore exit with an error code so that we avoid the potential
data loss.
* src/copy.c (same_file_ok): Don't set unlink_src that was used by mv,
and return false for two hardlinks to a file in move_mode.
*src/copy.c (copy_internal): No longer honor the unlink_src option,
used only by mv.
NEWS: Mention the change in behavior.
* tests/cp/same-file.sh: Augment to cover the `cp -a hlsl1 sl1` case.
* tests/mv/hard-verbose.sh: Remove no longer needed test.
* tests/local.mk: Remove the reference to hard-verbose.sh.
* tests/mv/hard-4.sh: Adjust so we fail in this case.
* tests/mv/i-4.sh: Likewise.
* tests/mv/symlink-onto-hardlink-to-self.sh: Likewise.
With --sparse=always use fallocate(...PUNCH_HOLE...) to
avoid any permanent allocation due to speculative
preallocation employed by file systems such as XFS.
* m4/jm-macros.m4: Check for <linux/falloc.h> and fallocate().
* src/copy.c (punch_hole): A new function to try and punch
a hole at the specified offset if supported.
(create_hole): Call punch_hole() after requesting a hole.
(extent_copy): Likewise.
* NEWS: Mention the improvement.
Previously cp would not detect runs of NULs that were
smaller than the buffer size used for I/O (currently 128KiB).
* src/copy.c (copy_reg): Use an independent hole_size, set to
st_blksize, to increase the chances of detecting a representable hole,
in a run of NULs read from the input.
(create_hole): A new function refactored from sparse_copy() and
extent_copy() so we have a single place to handle holes.
(sparse_copy): Adjust to loop over the larger input buffer
in chunks of the passed hole size. Also adjust to only call
lseek once per hole, rather than at least once per input buffer.
* tests/cp/sparse.sh: Add test cases for various sparse chunk sizes.
* NEWS: Mention the improvement.