just like it were fixed in 1998-01-02).
A user (FIXME, who?) reported that using `install -d -g foo 1/2`
only sets the group on the intermediate directory, not the final
component. From Michael Stone.
target cannot be `stat'ed, use the errno from the failed mkdir call,
not the one from the stat call. Before this change, running
`mkdir -p /no-dir/no-dir' as an unprivileged user would wrongly
elicit `No such file or directory' instead of `Permission denied'.
Otherwise, we'd get the wrong definition of PARAMS from libintl.h.
(The method of defining PARAMS in libintl.h doesn't check PROTOTYPES,
which is necessary on Irix4 since cc doesn't define __STDC__.)
From Kaveh Ghazi.
(make_path): Use make_dir rather than open-coding it twice.
This effectively reverses the order of the latter pair of stat/mkdir
calls and fixes a race condition bug whereby one of two concurrent
`mkdir -p' processes could fail with EEXIST.
directory. Fix latent bug (s/&&/||/ in two places -- also, note that
it could not be exercised via install or mkdir) whereby chown would
not be invoked when only one of owner/group is not -1.
calls so that mkdir is called first. Before make_path would first
`stat' a directory, then call mkdir if it didn't exist. But if
some other process created the directory between the stat & mkdir,
the mkdir would fail with EEXIST. Diagnosis and suggestion from
Greg McGary.
Before, it processed O(n^2) directory name components via stat and mkdir.
Now it's O(n). This makes mkdir -p a lot more efficient when creating
directories with very many components. On a Linux 2.0.30 ext2fs filesystem
this command: mkdir -p `perl -e 'print "a/" x 500'` now runs in 0.77 seconds
(user+sys). Contrast that with the 9.5(!) seconds it took before.
