The Austin Group made their recommendation about mkdir -p today:
"I suggest replacing the description of -p with:
-p Create any missing intermediate pathname components.
For each dir operand that does not name an existing directory,
before performing the actions described in the DESCRIPTION above
the mkdir utility shall create any pathname components of the path
prefix of dir that do not name an existing directory by performing
actions equivalent to first calling the mkdir() function with the
following arguments:
1. A pathname naming the missing pathname component, ending
with a trailing <slash> character, as the path argument.
2. The value zero as the mode argument.
and then calling the chmod() function with the following
arguments:
1. The same path argument as in the mkdir() call.
2. The value (S_IWUSR|S_IXUSR|~filemask) as the mode argument,
where filemask is the file mode creation mask of the process
(see [xref to umask()]).
Each dir operand that names an existing directory shall be ignored
without error."
http://austingroupbugs.net/view.php?id=161
The significant part about this is that it requires, for 'mkdir -p a/b', that
implementations call mkdir("a/",mode), not mkdir("a",mode). This makes a
difference, given that the POSIX requirements for mkdir(2) will follow through
a dangling symlink "a" to create the underlying directory, where without the
trailing slash it would fail with EEXIST.
Should I go ahead and start modifying the mkdir-p module to preserve the
trailing slash to the underlying mkdir call? It won't immediately change
behavior on GNU/Linux (without either a kernel change or an rpl_mkdir wrapper),
because GNU's behavior for mkdir("dangling/") is contrary to POSIX. But it
will make a difference on Solaris.
This also means that build-aux/mkinstalldirs and build-aux/install-sh are also
impacted. And there, since we can't guarantee whether the system mkdir(1)
gracefully handles trailing slashes, it could get rather difficult to guarantee
creating a directory tree through a dangling symlink using portable shell code.
--
Eric Blake
