hardlinks and cp -R

[Philip Rowlands]

It's been 9 years, but I still don't feel comfortable with the outcome of:
http://lists.gnu.org/archive/html/bug-coreutils/2004-08/msg00054.html

$ rm -rf 1 2; mkdir -p 1/d 2/d
$ echo foo > 1/d/1; ln 1/d/1 1/d/2; ln 1/d/1 1/d/3
$ echo foo > 2/d/1
$ echo bar > 2/d/2; ln 2/d/2 2/d/3; ln 2/d/2 2/d/4

$ grep . {1,2}/d/*
1/d/1:foo
1/d/2:foo
1/d/3:foo
2/d/1:foo
2/d/2:bar
2/d/3:bar
2/d/4:bar


$ cp -R 2/d 1
$ diff -ur 2 1
diff -ur 2/d/1 1/d/1
--- 2/d/1       2013-08-17 00:18:06.786991409 +0100
+++ 1/d/1       2013-08-17 00:18:31.931010671 +0100
@@ -1 +1 @@
-foo
+bar

The order matters greatly here; whichever of 1/d/{1,2,3} is written to 
last will win, by virtue of being hard-linked. As the source files have 
different contents, we can have different outcomes depending how 
getdents(2) chooses to order the files.

Should I expect to get the same result as "rsync -aA 2/d 1"? Is it my 
own risk for failing to ask for --preserve=links or 
--remove-destination? As we're firmly in POSIX-land, what does POSIX 
say, and could cp do anything other than its current behaviour (nothing 
requires or forbids reordering in 2.f - "The files in the directory 
source_file shall be copied ...")


Cheers,
Phil