Hello everyboddy,
   I have a (probably) stupid question... Please forgive me if she really is.

I "discovered" today a strange thing:
$ mkdir foo

$ cp -R foo foo
I was waiting something like "cannot copy a directory into himself" or

something like that but in fact this command create a (infinite ? I don't know

because I stop the action before the end) tree, like this:
/foo/foo/foo/foo/foo/foo/foo...
Look like a bug for me. Isn't it ?
--

Rimi Bougard

This looks like fun ... 8-)  And this is open source, so let's follow

the code and learn something as we go along ...
But first, I guess it IS following your instructions ...
You asked it to copy what's in directory foo, recursively.  And you

are changing what's in foo at the same time ...
1. What's in foo? foo

2. So copy foo to foo - giving foo/foo.

3. What's in foo?  foo/foo

4. So copy foo/foo to foo/foo/foo.

5. Repeat.
Until it goes boom.
# mkdir foo

# cp foo foo

cp: foo is a directory (not copied).
(Sorry for the long lines below - haven't quite figured out how to

stop that.)
# mkdir foo

# cp -R foo foo

cp: foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo: name too long (not copied)

cp: foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/fo...
[ message continues ]

unfortunately, using strcmp fails in many cases where the directory

may be accessed using different pathnames.  also, this doesn't address

the case of more than source argument.

I am afraid solving the general case is harder.

You would have to do a full tree search down into the directories

to copy and see if you find the destination anywhere. During the

tree scan you would have to use all rules about following

links that the actual copy recursion would use.
And that would not be enough... I made a little test and if the

target contains directories that have the same name(s) as in

the source, they are retained (their inode number does not change).
This should mean that the target may contain an arbitrarily deep

directory structure that on any point can hard link to somewhere

in the source directory structure, causing a cycle that is very

expensive to find, even if you do the cycle detection

during the copy recursion.
The path length limit actually works as a crude, cheep and effective

cycle detection.
While Windows Finder solves the simple case of copying into

yourself, I do not think it has solved the general case.

Anyone curious to try?
-- 
/ Raimo Niskanen, Erlang/OTP, Ericsson AB

I might try a Linux install (got some Slackware 12.0 CDs) and look at

what it does (and its source code) - sure someone will beat me to it.
Looks like OpenBSD, FreeBSD, and Mac OS X BSD bits have the same

sort of outcome.
I got this from Tom Van Looy (thanks, Tom):
Copy foo to foo only once and quit, I think that's the correct

behaviour. I even think that posix more or less describes that.
$ mkdir foo

$ cp -R foo foo

$ ls -R foo

foo/:

foo
foo/foo:
That's what it should be doing I guess. But it's tricky, if you start

doing symlinks etc, you'll end up looping sooner or later. What if

the directoy's are not named the same (eg: hard links)?
(please reply cc to the list)
And another email from Tom Van Looy:
Oh and yes, it doesn't go boom, the copy just stops when the max

filename length is reached (logically). And "<blablabla> name too

long (not copied)" is an error, you can just redirect it to dev null,

cp -R foo foo 2>/dev/null, if you don't like it on your screen.
Try fixing it so that it doesn't copy a directory into itself when it

is allready copied (so allow it only once). GNU cp does something

like that:
$ mkdir foo

$ cp -R foo foo

cp: cannot copy a directory, `foo', into itself, `foo/foo'

$ ls -R foo

foo:

foo
foo/foo:
Only, I think the warning is not needed (and I didn't further test

it, so it's possibly doing something else after all).

correction: hard links are not allowed on directory's, ...

that being said, comparing inodes seems the best solution
only, don't give an error but copy once

maybe if I have time this weekend I'll try code that behaviour
Anyway, it has worked like that since years, and I guess nobody has had

a problem with it before. I don't think it should be changed just

because some bored guy playing with it noticed strange output ;-p
Regards,
Tom

Sure, but "bored guy" can translate to "new ideas" and testing

somehow-still-untested code paths. It's worth a shot at fixing.
-Nick

For what its worth, all I can offer is moral support since I don't code

C.  
Conceptually, though, why can't cp look at the source directory and take a

snapshot, a to-do-list, of everything it has to copy, then do it?  That

way, any recursion would be completed before the target directory

appeared in the source directory.  With only an -R (no -H -L or -P), it

should copy links as links which should avoid loops.
How do scp and rsync do it?
Doug.

On Fri, Oct 19, 2007 at 01:52:03PM -0400, Douglas A. Tutty wrote:

| Conceptually, though, why can't cp look at the source directory and take a

| snapshot, a to-do-list, of everything it has to copy, then do it?  That

| way, any recursion would be completed before the target directory

| appeared in the source directory.  With only an -R (no -H -L or -P), it

| should copy links as links which should avoid loops.
What will you do if the underlying directory structure has tons and

tons of files and subdirectories ? First traverse this entire tree,

keeping it all in memory ? Sounds pretty expensive.
| How do scp and rsync do it?
Why not try it ? scp local to remote doesn't make sense in this case,

since src and dst are not really the same. scp local to local has the

exact same effect (you even get an error message from cp that the name

is too long).
rsync does exactly what you just propose. It scans the entire src

directory structure, keeping it all in memory, and then copies that to

dst. If there's too many files in your hierarchy, rsync will fail. cp

will not (unless of course you run out of inodes).
Cheers,
Paul 'WEiRD' de Weerd
+++++++++++>-]<.>++[<------------>-]<+.--------------.[-]

                 http://www.weirdnet.nl/

I believe that it is only necessary to do one directory read ahead in

order to avoid the recursive loop. (I just took a quick look at the cp

source in OpenBSD, so, correct me if I am wrong.) 
    $ cp <source1> ... <sourceN> <directory>
We can read the contents of <source1> if source1 is a directory, then

create a directory in <directory> and copy the contents of source1 into

it, and then repeat. This should not cause an endless loop.
--

((name "Aaron Hsu")

 (email/xmpp "arcfide@sacrificumdeo.net")

 (phone "703-597-7656")

 (site "http://www.aaronhsu.com"))
[demime 1.01d removed an attachment of type application/pgp-signature]

Could you scan the tree only for recursions? You wouldn't have to keep

it all in memory, only the problem points, maybe?
None of this feels like the Right Way though.
-Nick

the other thing you can do is not copy directories into themselves.

it's very easy.  i'm not copying a directory into itself right now.  i

can even not do this while sleeping.

At linux 2.6.22 (Ubuntu 7.10)

$ mkdir foo

$ scp -r foo localhost:/home/spyro/foo/
foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo: name too long
Using OpenSSH_4.6p1 Debian-5build1, OpenSSL 0.9.8e 23 Feb 2007

Well, given that I feel targeted by this sentence ;) I think I will look at

the cp's source code and try to "fix" that. Could be fun after all.
--

Remi Bougard

ps: it was a ;-p

Ill try this on a solaris box and a linix box tomorrow at work :P
--

Best Regards
Edd
---------------------------------------------------

http://students.dec.bournemouth.ac.uk/ebarrett

I beat you to trying it on Linux (Ubuntu "Gutsy Gibbon" 7.10):
ropers@tranquility:~$ uname --all

Linux tranquility 2.6.22-14-386 #1 Sun Oct 14 22:36:54 GMT 2007 i686 GNU/Linux

ropers@tranquility:~$ echo $SHELL

/bin/bash

ropers@tranquility:~$ mkdir foo

ropers@tranquility:~$ cp -R foo foo

cp: cannot copy a directory, `foo', into itself, `foo/foo'

ropers@tranquility:~$
IMHO cp behaving like this is somewhat nicer than its current

behaviour on apparently most or all BSD OSes. Then again, I STILL

can't code, so I've no right to complain. ;o)

No I didn't. Others beat me and you to it. Apologies for the unnecessary noise.
I'm surprised now.
I just thought that what I wrote above was stupid, because I thought

that the behaviour of cp was a function of the shell built-in command

cp, not of the OS.

To confirm this, I installed the OpenBSD default shell pdksh on

Ubuntu. However, pdksh on Ubuntu gives the same result as bash on

Ubuntu. So is this a function of the OS after all?
| ropers@tranquility:~$ uname -a

| Linux tranquility 2.6.22-14-386 #1 Sun Oct 14 22:36:54 GMT 2007 i686 GNU/Linux

| ropers@tranquility:~$ echo $SHELL

| /bin/bash
We're on Linux and we're using bash.
| ropers@tranquility:~$ mkdir foo

| ropers@tranquility:~$ cp -r foo foo

| cp: cannot copy a directory, `foo', into itself, `foo/foo'
Bash behaves as expected.
| ropers@tranquility:~$ sudo apt-get install pdksh

| Reading package lists... Done

| Building dependency tree

| Reading state information... Done

| The following NEW packages will be installed:

|   pdksh

| 0 upgraded, 1 newly installed, 0 to remove and 0 not upgraded.

| Need to get 0B/261kB of archives.

| After unpacking 442kB of additional disk space will be used.

| Selecting previously deselected package pdksh.

| (Reading database ... 167230 files and directories currently installed.)

| Unpacking pdksh (from .../pdksh_5.2.14-20build1_i386.deb) ...

| Setting up pdksh (5.2.14-20build1) ...
Ok, now pdksh is installed.
| ropers@tranquility:~$ pdksh

| $ ps | grep sh

|  6567 pts/0    00:00:00 bash

|  6816 pts/0    00:00:00 pdksh

|  6818 pts/0    00:00:00 pdksh
Now we're running pdksh (echo $SHELL isn't changed when launching

another shell interactively, hence the use of ps to confirm).
| $ rm -rf foo
Need to rm foo to start from scratch.
| $ mkdir foo

| $ cp -r foo foo

| cp: cannot copy a directory, `foo', into itself, `foo/foo'
Strange.
pdksh on Linux behaves just like bash on Linux, and unlike pdksh on OpenBSD.

I didn't expect that. So does that error messa...
[ message continues ]

It's official: I am a fucking moron.
cp is not a SHELL BUILTIN COMMAND. It's /bin/cp. And of course the

same /bin/cp is run regardless which shell is being used. That

explains the identical error message.
Sorry for the noise.

Hello,
cp is part of the libc.
--

Remi Bougard

the behavior of the cp program depends on which cp program you are
or not.

On 2007 Oct 18, at 4:40 PM, Edd Barrett wrote:
 > On 18/10/2007, Richard Toohey <richardtoohey@paradise.net.nz> wrote:

 >

 >>> $ mkdir foo

 >>> $ cp -R foo foo

 >

 > Ill try this on a solaris box and a linix box tomorrow at work :P
Mac OS X 10.4 behaves exactly the same way as OpenBSD does.
Cheers,
b&
[demime 1.01d removed an attachment of type application/pkcs7-signature which had a name of smime.p7s]

For what's it's worth, this is what OS X decides:

Axkbk:~ kousu$ mkdir test

Axkbk:~ kousu$ cd test/

Axkbk:~/test kousu$ mkdir foo

Axkbk:~/test kousu$ cp -r foo foo

cp:

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo:

name too long (not copied)

cp:

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/fo

o/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo

/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/

foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f

oo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/foo/f...
[ message continues ]

looks like a feature to me.  ;)
-B

Agreed, although it does not seem to exists on GNU/Linux since GNU's cp

is different from BSD's.  The feature is present on {Net,Open,Free}BSD.
It's not that a big deal, is it?  Eventually, the question could be: what

should be limiting cp there?  a max_path value, or... himself? I think

the former's the best.
Just my 2 cents. :)
--

B+ A nation is a society united by a delusion about it's ancestry and by

  common hatred of its neighbours. B;            -- Dean William R. Inge

penguin's behaviour:
elachistos| cp -R foo foo

On Debian, you also end up with a directory structure consisting of

one new 'foo' directory within the original 'foo' directory, which is

contradicting the message about not being able to copy foo into

itself...
$ mkdir foo

$ touch foo/bar

$ cp -R foo foo

cp: cannot copy a directory, `foo', into itself, `foo/foo'

$ ls -lR foo

foo:

total 4

-rw-r--r-- 1 ak ak    0 2007-10-19 11:14 bar

drwxr-xr-x 2 ak ak 4096 2007-10-19 11:14 foo
foo/foo:

total 0

-rw-r--r-- 1 ak ak 0 2007-10-19 11:14 bar
According to SUSv3, the cp utility *may* issue a diagnostic message

when the source and target arguments are the same. IMHO we're doing

the right thing with regards to that part. I'm not sure about

recursively creating a very deep directory structure, but it's not a

problem really.
Andreas
--

Andreas Kahari

Somewhere in the general Cambridge area, UK