(card_of_complement): Use cleaner `sizeof in_set'
rather than `N_CHARS * sizeof(in_set[0])'. Using HPUX's /bin/cc
(aC++/ANSI C B3910B A.05.55 [Dec 04 2003]) on an ia64-hp-hpux11.22
system, those two expressions are not the same (256 vs. 1024).
The effect of this problem was that `tr -c x y' would fail:
tr: when not truncating set1, string2 must be non-empty
directory and with two or more command line arguments including an
absolute-named directory followed by a relative-named directory.
(struct cwd_state): Define.
(AD_pop_and_chdir): Redesign interface so that a restore_cwd failure
can be detected by the caller. Instead of returning a malloc'd
directory name, communicate it to caller via a new parameter, and
return an indication of whether restore_cwd failed. Update caller.
Eliminate an unnecessary call to AC_stack_top.
(remove_dir): Change type of cwd_state parameter to `struct cwd_state'
so we can now communicate to caller whether/how functions like
restore_cwd have failed. Update caller.
(rm_1): Fail if we've failed to restore the working directory
and the name of the next file to remove is `.'-relative.
(rm): Fail if the require_restore_cwd flag is true and we've
failed to restore the working directory.
".". But continue to omit +, =, %, @, #, as they're either
shell metacharacters (for some shells) or are not in some
character sets, or (in the case of '%') must be a
metacharacter somewhere.
symlink matched the desired owner/group. Reported by David Malone.
Also reported in 1999 as http://bugs.debian.org/39642.
(change_file_owner): When --dereference has
been specified, and when processing a symlink, stat it to get the
owner and group of the referent.
fdatasync fails with errno==EINVAL, it means this implementation
does not support synchronized I/O for this file. Do not report
this as an error, as (for example) AIX 5.2 fdatasync reports it
for raw disk devices. Problem reported by Albert Chin in
<http://mail.gnu.org/archive/html/bug-gnu-utils/2004-05/msg00028.html>.
Check for write errors, though: the old code ignored them.
Improve error checking in a few other cases, too (e.g., close of a
directory).
Also, change several 'int' values to 'bool', so that the error
checking is a bit clearer. Similarly, change unsigned values
to size_t where appropriate.
* src/shred.c: Include "dirname.h".
(datasync) [!HAVE_FDATASYNC]: Remove.
(dosync): New function.
(dopass): Use it. Return 1 on write error, -1 on other error.
All callers changed. Report write error if dosync does.
(do_wipefd, wipefd, wipename, wipefile): Return bool (true/false),
not int (0/-1). All callers changed. Return false if there's a
write error.
(incname): Return bool (true/false), not int (0/1). Accept
size_t length, not unsigned. All callers changed. Do not
bother checking for non-digits; it can't happen. Replace
recursion with iteration.
(wipename): Use dir_name, base_name, etc. instead of assuming
Unix file names. Use size_t for length, not unsigned.
Report error if unlink or close fails.
(wipename, main): Use bool for booleans.
(names): Use only digits and uppercase letters, for greater
portability.
(all_programs.list): New rule, copied from
man/Makefile.am and tests/Makefile.am, except that we use the
system tr rather than ./tr and we don't use tr -s.
<http://mail.gnu.org/archive/html/bug-coreutils/2004-05/msg00013.html>.
(remove_entry): Check for errno values like ENOENT
that show the file cannot be directory, instead of for errno
values like EPERM that show the file might be a directory. This
is necessary because, when a single unlink() call has multiple
reasons to fail, it can set errno to any of those reasons; it's
only the rare errno value like ENOENT that excludes all the other
possible reasons to fail even when the file is a directory.
(remove_cwd_entries): Don't attempt chdir if the file is known
to not be a directory.
(remove_dir): Use the same method that remove_cwd_entries uses
(for some reason they differed). Don't assert that saved_errno
must be EPERM; it might be just about anything.
Don't add `1' to the buffer size (it was to protect against malloc
implementations that fail to allocate a buffer of size zero).
That is no longer necessary, since we use a malloc wrapper
on such systems.
mounted on the same mount point, prefer the last one, not the first.
Problem reported by Christian Jones in
<http://mail.gnu.org/archive/html/bug-coreutils/2004-04/msg00200.html>.
(show_disk): Remove unused statp arg. Return bool, not int.
(show_point): Rewrite to avoid gotos. Use the same algorithm
for lofs and dummies for each pass through the mount table,
rather than subtly different algorithms (which are probably
inadvertent).
use the chown(2) function, if possible.
(change_file_owner): Don't hard-code the
open/fchown/close kludge here. Use `chown' instead.
The chown function works just fine on conforming systems.
Other systems now go through the new chown wrapper that
resorts to the old kludge.