Why does OS X `install` errors on redirected input when the same version of `install` on linux works fine?

Question

I would like to use install command in order to create a new executable file with pre-populated content (e.g. with single pwd command in it).

So I've extended this example which creates a new empty executable file:

install -b -m 755 /dev/null newfile 

into this one:

install -m755 <(echo pwd) newfile 

or:

echo pwd | install -m755 /dev/stdin newfile 

Where I expect to create a new newfile executable file to be created with content pwd inside.

It works on Linux, however on OS X it fails with the following error:

• BSD install (/usr/bin/install)

  install: /dev/fd/63: Inappropriate file type or format

• GNU install (/usr/local/opt/coreutils/libexec/gnubin/install)

  install: skipping file /dev/fd/63, as it was replaced while being copied

Why this doesn't work on Unix, but it works on Linux? I'm missing anything? Is there any way to bypass the above warning by using different syntax (without creating a file in separate commands and using chmod after that)?

---

On both environments (Linux & OS X) I've the same version of install:

$ install --version install (GNU coreutils) 8.23 

Answer 1

The BSD install found on OpenBSD systems has this piece of code in it (from src/usr.bin/xinstall/xinstall.c):

if (!S_ISREG(to_sb.st_mode)) errc(1, EFTYPE, "%s", to_name); 

This emits the error

install: /dev/fd/4: Inappropriate file type or format 

when it's discovered that /dev/df/4 is not a regular file. (There's a separate earlier check for /dev/null)

That was fairly straight forward.

GNU install has this code (src/install.c in coreutils):

 /* Allow installing from non-regular files like /dev/null. Charles Karney reported that some Sun version of install allows that and that sendmail's installation process relies on the behavior. However, since !x->recursive, the call to "copy" will fail if FROM is a directory. */ return copy (from, to, false, x, &copy_into_self, NULL); 

The code emitting the error comes from src/copy.c:

 source_desc = open (src_name, (O_RDONLY | O_BINARY | (x->dereference == DEREF_NEVER ? O_NOFOLLOW : 0))); 

(a few lines omitted)

 if (fstat (source_desc, &src_open_sb) != 0) { error (0, errno, _("cannot fstat %s"), quoteaf (src_name)); return_val = false; goto close_src_desc; } /* Compare the source dev/ino from the open file to the incoming, saved ones obtained via a previous call to stat. */ if (! SAME_INODE (*src_sb, src_open_sb)) { error (0, 0, _("skipping file %s, as it was replaced while being copied"), quoteaf (src_name)); 

This is in copy_reg() which copies a regular file. The SAME_INODE macro evaluates to false because the inodes differ in the two stat structs *src_sb and src_open_sb. The *src_sb comes from a stat() or lstat() call on the source file name and src_open_sb from fstat() as seen above, on a newly open descriptor.

I can kinda see why opening a new file descriptor and comparing its inode to that of the file descriptor given by the shell (/dev/fd/4 in my case) will fail, but I can't put it into definite words unfortunately.