cp --link was "remembering" many name,dev,inode triples unnecessarily.
cp was doing the same, even without --link, for every directory in the
source hierarchy, while it can do its job with entries merely for the
command-line arguments. Prompted by a report from Patrick Shoenfeld.
Details <http://thread.gmane.org/gmane.comp.gnu.coreutils.bugs/15081>.
* src/copy.c (copy_internal): Refrain from remembering
name,dev,inode for most files, when invoked via cp --link.
Record an infloop-avoidance triple for each directory specified
on the command line, not for each directory in the source tree.
Don't record a dir-triple when x->hard_link is set.
* NEWS (Buf fixes): Mention it.
* tests/cp/link-heap: New file. Test for cp's lowered memory usage.
* tests/Makefile.am (TESTS): Add link-heap.
Now that a (void) cast no longer suffices to ignore warnings from gcc
about uses of functions marked with the warn_unused_result attribute,
we need an alternative. For the record, here's one of the ignorable
warnings: "copy.c:233: warning: ignoring return value of 'fchown',
declared with attribute warn_unused_result"
* bootstrap.conf (gnulib_modules): Import ignore-value.
* src/copy.c: Include "ignore-value.h".
(set_owner): Use ignore_value in place of "(void)" casts,
to ignore lchown and fchown failures.
* src/cp.c (re_protect): Likewise, to ignore lchown failure.
* src/remove.c (preprocess_dir): Remove unnecessary "(void)" cast.
* src/copy.c (copy_internal): Revert change of 2005-03-01,
4303f04545
Prompted by Michael Stone, who pointed me to an old bug report from
Ian Jackson: http://thread.gmane.org/gmane.comp.gnu.coreutils.bugs/7504
* tests/cp/special-f: New file. Test for the above.
* tests/cp/Makefile.am (TESTS): Add special-f.
* NEWS: mention this bug fix.
While cp --preserve=links must unlink certain destination files,
mv must never do that.
* src/copy.c (copy_internal): Pull the '! x->move_mode' test "up",
so it affects the entire condition, and not just DEREF_NEVER mode.
Reported by James Ralston in <http://bugzilla.redhat.com/438076>.
* tests/mv/atomic2: New file. Test for the above fix.
* tests/mv/Makefile.am (TESTS): Add atomic2.
* NEWS: Mention the bug-fix.
[Bug introduced in 367719ba5f]
* src/mktemp.c: Update copyright to use newer form.
Make first copyright lines consistent. Some did not end in ", Inc.".
Due to that missing ", Inc" at end of line, emacs' copyright-updating
code missed some opportunities. This corrects most of those, and
adds some mistakenly omitted years.
* src/chown-core.c: Adjust.
* src/chown-core.h: Likewise.
* src/copy.c: Likewise.
* src/copy.h: Likewise.
* src/cp-hash.c: Likewise.
* src/cp.c: Likewise.
* src/group-list.h: Likewise.
* src/mktemp.c: Likewise.
* src/stat.c: Likewise.
* tests/chmod/equal-x: Likewise.
* tests/dd/skip-seek: Likewise.
* tests/envvar-check: Likewise.
* tests/mv/into-self-2: Likewise.
* src/copy.h (struct cp_options): New member owner_privileges.
* src/copy.c (USE_ACL): Define to 0 if not defined, for convenience.
(owner_failure_ok): New function.
(set_owner): Avoid a security-related race by doing an extra chmod
first if it looks like there might be trouble right after a chown.
Accept a source struct stat rather than a uid and gid, and
accept a boolean NEW_DST and destination struct stat.
All callers changed.
* src/copy.h (cp_options_default): New function, replacing the
old chown_privileges.
* src/copy.c (cp_options_default): Likewise.
* src/cp.c (cp_option_init): Use it.
* src/install.c (cp_option_init): Likewise.
* src/mv.c (cp_option_init): Likewise.
* NEWS: Mention this new feature.
* src/copy.c (set_owner): Try to preserve just the GID,
when initial fchown/lchown fails.
* src/cp.c (re_protect): Likewise.
* NEWS: Mention this change.
* doc/coreutils.texi (cp invocation): Describe the new behavior.
* src/copy.c: No longer include "canonicalize.h".
(copy_reg): Upon failure to open a dangling destination symlink, don't
canonicalize the name, but rather fail (default) or, with POSIXLY_CORRECT,
repeat the open call without O_EXCL (potentially dangerous).
* src/copy.h (struct cp_options) [open_dangling_dest_symlink]:
New member. Reorder the others, grouping "bool" and "enum"
members together.
* tests/cp/thru-dangling: Test for changed and new behavior.
* src/cp.c (cp_option_init): Initialize new member.
* src/install.c (cp_option_init): Likewise.
* src/mv.c (cp_option_init): Likewise.
Signed-off-by: Jim Meyering <meyering@redhat.com>
Those programs must not dereference a destination symlink.
* src/copy.c (copy_reg): Don't treat a dangling destination symlink
differently in move mode. In move mode, the only way the added
O_EXCL can cause failure is when some other process has recreated
the file this code unlinked a few instructions before.
* src/copy.c (copy_internal): When same-file detection requires
'stat'ing the destination file, also 'lstat' it and ensure that
it wasn't the destination of a preceding copy operation.
This bug was introduced on 2007-06-18.
* tests/cp/abuse: New test for the above.
* tests/cp/Makefile.am (TESTS): Add abuse.
* bootstrap.conf (gnulib_modules): Add file-set.
* gl/lib/file-set.c (record_file, seen_file): Functions from copy.c.
* gl/lib/file-set.h: Add prototypes.
* gl/lib/hash-triple.c (triple_hash, triple_hash_no_name):
(triple_compare, triple_free): Functions from copy.c.
* gl/lib/hash-triple.h (struct F_triple): Define. From copy.c.
Add prototypes.
* gl/modules/file-set: New module.
* gl/modules/hash-triple: New module.
* src/Makefile.am (copy_sources): New variable.
(ginstall_SOURCES, cp_SOURCES, mv_SOURCES): Use it.
* src/copy.c: Include hash-triple.h.
No longer include hash-pjw.h.
(copy_internal): Don't pass a NULL third argument to record_file,
since that function no longer accepts that.
(record_file): Move this function to file-set.c.
Along the way, remove the code to allow a NULL stat-buffer pointer.
Adjust sole caller.
(seen_file): Move this function to file-set.c.
(struct F_triple): Move declaration to hash-triple.h.
(triple_compare, triple_free, triple_hash, triple_hash_no_name):
Move these functions to hash-triple.c.
Signed-off-by: Jim Meyering <jim@meyering.net>
* NEWS: Document this bug fix.
* src/copy.c (copy_reg): Read from a regular file, even if it
appears (stat.st_size == 0) to be empty. This reverts an
optimization introduced on 2005-11-23 for coreutils-6.0.
Otherwise, "cp /proc/cpuinfo /tmp" creates an empty file,
on e.g., linux-2.6.20.
* tests/cp/proc-zero-len: New file. Test for the above.
* tests/cp/Makefile.am (TESTS): Add proc-zero-len.
Reported by Dan Berrangé.
* src/copy.c (copy_internal): Perform "update" check before the
possible interactive prompt. Reported by zeno_AT_biyg_DOT_org
in <http://bugzilla.redhat.com/248591>
* tests/mv/update: Add tests for the above.
* NEWS: Mention the bug fix.
--parents and symlinks. Close some race conditions possible when
the destination replaces a newly-created file with a symlink.
* NEWS: Document that 'cp --parents' no longer mishandles
symlinks in file name components of source.
* src/copy.c (HAVE_LCHOWN): Default to false.
(lchown) [!defined HAVE_LCHOWN]: Define to chown, for convenience.
* src/cp.c (lchown) [!HAVE_LCHOWN]: Likewise.
* src/install.c (lchown [!HAVE_LCHOWN]: Likewise.
* src/copy.c (set_owner): Use lchown instead of chown, for safety
in case the file got replaced by a symlink in the meantime.
* src/cp.c (re_protect): Likewise.
* src/install.c (change_attributes): Likewise.
* src/copy.c (copy_internal): Use ordinary C rather than an #if.
* src/cp.c (lchown) [!HAVE_LCHOWN]: Define to chown, for convenience.
(struct dir_attr): Cache the entire struct stat of the directory,
rather than just its mode, so that we needn't stat the directory
twice (which can lead to races).
(re_protect): Don't use XSTAT as that's not appropriate in
this context (symlinks should be followed here). Instead, use
the cached stat value.
(make_dir_parents_private): Save dir's entire struct stat, not
just its mode.
* tests/cp/cp-parents: Add test to check against bug with
cp --parents and symlinks.
* NEWS: "cp" no longer considers a destination symlink to be the
same as the referenced file when copying links or making backups.
* src/copy.c (copy_reg): When following a symlink, use the
followed name in later chown etc. requests, so that the created
file is affected, rather than the symlink. Use O_NOFOLLOW on
source when not dereferencing symlinks; this avoids a race.
Preserve errno correctly when doing multiple open attempts on the
destination.
(copy_internal): Follow destination symlinks only when copying a
regular file and only when we don't intend to remove or rename the
destination first, regardless of whether following source
symlinks; this is because since POSIX and tradition (e.g.,
FreeBSD) say we should ordinarily follow destination symlinks if
the system calls would ordinarily do so.
* src/copy.h (struct cp_options): Add comment that 'dereference'
is only for source files.
* src/cp.c (usage): Note that --derereference etc. are only for
source files.
(make_dir_parents_private): Follow symlinks, regardless of whether
--dereference is specified, because these are destination symlinks.
* tests/cp/same-file: Adjust tests to match revised behavior.
Filter out perror output since it might vary from host to host.
Use sed alone instead of also using echo.
* doc/coreutils.texi (cp invocation): Document the behavior better when
the destination is a symlink. Clarify source versus destination
symlinks. Describe the new behavior for destination symlinks.
2007-06-15 Jim Meyering <jim@meyering.net>
* src/copy.c: Include "canonicalize.h".
(copy_reg): Use canonicalize_filename_mode to follow the symlink,
so that we can always open with O_EXCL and avoid a race.
* NEWS: Mention the bug fix.
* src/copy.c (copy_reg): When open fails with EEXIST, the destination
is lstat'able, and a symlink, call open again, but now without O_EXCL.
* tests/cp/thru-dangling: New file, to test for the above fix.
* tests/cp/Makefile.am (TESTS): Add thru-dangling.
* THANKS: Add Michael McLagan.
Bug report from Michael McLagan in <http://bugzilla.redhat.com/243588>.
* tests/cp/Makefile.am (TESTS): Add cp-a-selinux.
* tests/selinux: New file.
* tests/Makefile.am (EXTRA_DIST): Add selinux.
* tests/misc/selinux: Source the new script, rather than open coding it.
Change how "cp -a" and "cp --preserve=context" work 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.
* src/copy.h (struct cp_options) [require_preserve_context]: New member.
* src/copy.c (copy_reg, copy_internal): Implement the above.
* src/mv.c (cp_option_init): Initialize the new member.
* src/install.c (cp_option_init): Likewise.
* src/cp.c (cp_option_init): Likewise.
(decode_preserve_arg): Set it or reset it.
FIXME: add an on-writable-NFS-only test
patch, mv and cp do not provide the "-Z context" option.
* src/copy.c: Include <selinux/selinux.h>.
(restore_default_fscreatecon): New function.
(copy_reg): Make cp --preserve=context work for existing destination.
(copy_internal): Likewise for new destinations.
* src/copy.h (cp_options) [preserve_security_context]: New member.
* src/cp.c: Include <selinux/selinux.h>.
(selinux_enabled): New global.
(usage): Mention new --preserve=context option.
(PRESERVE_CONTEXT): Define/use.
(decode_preserve_arg): Handle PRESERVE_CONTEXT.
(main): Remove an obsolete comment.
If --preserve=context is specified on a system without SELinux
enabled, give a diagnostic and fail.
* src/mv.c: Include <selinux/selinux.h>.
Set x->preserve_security_context if SELinux is enabled.
* src/install.c: Accept new "-Z, --context=C" option.
Accept --preserve-context option (but not -P option).
Accept alternate spelling: --preserve_context, for now.
Include <selinux/selinux.h> and "quotearg.h".
(selinux_enabled, use_default_selinux_context): New globals.
(PRESERVE_CONTEXT_OPTION): Define.
(cp_option_init): Default: do not preserve security context.
(setdefaultfilecon): New function.
(main): Honor new options.
* src/Makefile.am (mv_LDADD, cp_LDADD, ginstall_LDADD):
Add $(LIB_SELINUX).
with 'rm -r dir' (without -f), if we are root, or if we are removing
a directory tree that is full of symbolic links.
* bootstrap.conf (gnulib_modules): Add write-any-file.
* src/copy.c: Include write-any-file.h.
(UNWRITABLE): Remove macro, replacing with....
(writable_destination): New function, which uses can_write_any_file
to avoid the need for euidaccess when we are privileged.
(overwrite_prompt, abandon_move): Use it.
* src/remove.c: Include write-any-file.h.
(D_TYPE): New macro.
(DT_UNKNOWN, DT_DIR, DT_LNK) [!HAVE_STRUCT_DIRENT_D_TYPE]: New macros.
(write_protected_non_symlink): Don't bother to stat if we can write
any file.
(prompt): New arg PDIRENT_TYPE. All callers changed.
Use readdir dirent type to avoid the need for 'lstat' on each directory
entry in cases like 'rm -r dir', if we are root, or if the tree is
full of symbolic links.
(DT_IS_KNOWN, DT_MUST_BE): Remove.
(remove_entry): New arg DIRENT_TYPE_ARG. All callers changed.
(overwrite_prompt): Say "try to overwrite", not "overwrite", to
make it clearer that the attempt may fail. Problem reported by
Dan Jacobson in:
http://lists.gnu.org/archive/html/bug-coreutils/2007-03/msg00130.html
Output symbolic mode as well as numeric.
* tests/mv/i-2 (fail): Adjust to new prompt format.
* NEWS: mention cp -x bug fix
* src/copy.c (copy_internal): Don't return immediately after
copying a mount point that we do not intend to recurse under.
Based on a patch by Andrew Church.
* src/copy.c (copy_reg): Undo previous change. Instead, make
it clearer that we're using a single-byte sentinel, and
[lint]: Initialize uintptr_t-1 bytes after the sentinel.
Reported by Andreas Schwab.
* src/copy.c (copy_reg): Place the sentinel by setting a
full word's worth of bits, not just a single byte's worth.
This avoids a harmless (but distracting) case of memory being
used-uninitialized.
* src/copy.c (copy_internal): Omit the group- or other-writeable
permissions when creating a directory, to avoid a race condition
if the special mode bits aren't right just after the directory is
created.
* src/cp.c (make_dir_parents_private): Likewise.
* tests/cp/parent-perm-race: Test for the "cp --preserve=mode"
race fix in copy.c.
* NEWS: Document the cp -p fix for special bits.
* src/copy.c (set_owner): Now returns a three-way result, so
that the caller can clear the special bits. All callers changed.
(copy_reg): Don't set the special bits if chown failed.
(copy_internal): Likewise.
* tests/cp/special-bits: Test this fix.
Signed-off-by: Jim Meyering <jim@meyering.net>
* m4/jm-macros.m4 (coreutils_MACROS): Check for fchmod.
* src/copy.c (fchmod_or_lchmod): New function.
(copy_reg): New arg OMITTED_PERMISSIONS. All uses changed.
Omit confusing and unused ", dst_mode" arg to 'open' without O_CREAT.
When creating a file, use O_EXCL, so we're more likely to detect
funny business by other processes. At the end, if permissions
were omitted, chmod them back in.
(copy_internal): If the ownership might change, omit some permissions
at first, then restore them after chowning the file.
* src/cp.c (make_dir_parents_private): Likewise.
* src/copy.c (cached_umask): New function.
* src/copy.h (cached_umask): New decl.