Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-02-24 Thread Pádraig Brady 

On 01/02/2024 00:36, Paul Eggert wrote:

On 1/31/24 06:06, Pádraig Brady wrote:

To my mind the most protective option takes precedence.


That's not how POSIX works with mv -i and mv -f. The last flag wins. I
assume this is so that people can have aliases or shell scripts that
make -i the default, but you can override by specifying -f on the
command line. E.g., in mymv:

 #!/bin/sh
 mv -i "$@"

then "mymv -f a b" works as expected.

Wouldn't a similar argument apply to cp's --update options?

Or perhaps we should play it safe, and reject any combination of
--update etc. options that are incompatible. We can always change our
mind later and say that later options override earlier ones, or do
something else that's less conservative.


OK I err'd on the side of changing as little as possible wrt precedence.
-n still has precedence over any -u,--update option.
That's simplest to understand (while not changing existing precedence),
and shouldn't cause any practical issues.

I plan to push the 2 attached patches for this tomorrow.

thanks,
PádraigFrom 51cf6f3ff272467bc9cde75c48d0250446be9b9c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?P=C3=A1draig=20Brady?= 
Date: Sat, 24 Feb 2024 19:51:56 +
Subject: [PATCH 1/2] cp,mv: reinstate that -n exits with success if files
 skipped

* src/cp.c (main): Adjust so that -n will exit success if skipped files.
* src/mv.c (main): Likewise.
* doc/coreutils.texi (cp invocation): Adjust the description of -n.
* src/system.h (emit_update_parameters_note): Adjust --update=none
comparison.
* tests/cp/cp-i.sh: Adjust -n exit status checks.
* tests/mv/mv-n.sh: Likewise.
* NEWS: Mention the change in behavior.
Fixes https://bugs.gnu.org/62572
---
 NEWS   |  4 
 doc/coreutils.texi | 17 +
 src/cp.c   | 14 --
 src/mv.c   | 10 ++
 src/system.h   |  4 ++--
 tests/cp/cp-i.sh   | 11 +--
 tests/mv/mv-n.sh   | 11 +--
 7 files changed, 39 insertions(+), 32 deletions(-)

diff --git a/NEWS b/NEWS
index 36b7fd1fe..a52b4cf66 100644
--- a/NEWS
+++ b/NEWS
@@ -43,6 +43,10 @@ GNU coreutils NEWS-*- outline -*-
   basenc --base16 -d now supports lower case hexadecimal characters.
   Previously an error was given for lower case hex digits.
 
+  cp --no-clobber, and mv -n no longer exit with failure status if
+  existing files are encountered in the destination.  Instead they revert
+  to the behavior from before v9.2, silently skipping existing files.
+
   ls --dired now implies long format output without hyperlinks enabled,
   and will take precedence over previously specified formats or hyperlink mode.
 
diff --git a/doc/coreutils.texi b/doc/coreutils.texi
index c42126955..911e15b46 100644
--- a/doc/coreutils.texi
+++ b/doc/coreutils.texi
@@ -9059,12 +9059,13 @@ a regular file in the destination tree.
 @itemx --no-clobber
 @opindex -n
 @opindex --no-clobber
-Do not overwrite an existing file; silently fail instead.
-This option overrides a previous
-@option{-i} option.  This option is mutually exclusive with @option{-b} or
-@option{--backup} option.
-See also the @option{--update=none} option which will
-skip existing files but not fail.
+Do not overwrite an existing file; silently skip instead.
+This option overrides a previous @option{-i} option.
+This option is mutually exclusive with @option{-b} or @option{--backup} option.
+This option is deprecated due to having a different exit status from
+other platforms.  See also the @option{--update} option which will
+give more control over how to deal with existing files in the destination,
+and over the exit status in particular.
 
 @item -P
 @itemx --no-dereference
@@ -9335,8 +9336,8 @@ This is the default operation when an @option{--update} option is not specified,
 and results in all existing files in the destination being replaced.
 
 @item none
-This is similar to the @option{--no-clobber} option, in that no files in the
-destination are replaced, but also skipping a file does not induce a failure.
+This is like the deprecated @option{--no-clobber} option, where no files in the
+destination are replaced, and also skipping a file does not induce a failure.
 
 @item older
 This is the default operation when @option{--update} is specified, and results
diff --git a/src/cp.c b/src/cp.c
index 0355ed97f..36ae4fb66 100644
--- a/src/cp.c
+++ b/src/cp.c
@@ -195,8 +195,8 @@ Copy SOURCE to DEST, or multiple SOURCE(s) to DIRECTORY.\n\
   -L, --dereferencealways follow symbolic links in SOURCE\n\
 "), stdout);
   fputs (_("\
-  -n, --no-clobber ensure no existing files overwritten, and fail\n\
- silently instead. See also --update\n\
+  -n, --no-clobber silently skip existing files.\n\
+ See also --update\n\
 "), stdout);
   fputs (_("\
   -P, --no-dereference never follow symbolic links in SOURCE\n\
@@ -984,6 +984,7

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-31 Thread Paul Eggert 

On 1/31/24 06:06, Pádraig Brady wrote:

To my mind the most protective option takes precedence.


That's not how POSIX works with mv -i and mv -f. The last flag wins. I 
assume this is so that people can have aliases or shell scripts that 
make -i the default, but you can override by specifying -f on the 
command line. E.g., in mymv:


   #!/bin/sh
   mv -i "$@"

then "mymv -f a b" works as expected.

Wouldn't a similar argument apply to cp's --update options?

Or perhaps we should play it safe, and reject any combination of 
--update etc. options that are incompatible. We can always change our 
mind later and say that later options override earlier ones, or do 
something else that's less conservative.

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-31 Thread Pádraig Brady 

On 30/01/2024 18:31, Paul Eggert wrote:

On 2024-01-30 03:18, Pádraig Brady wrote:

So we now have the proposed change as:

    - revert -n to old silent success behavior
    - document -n as deprecated
    - Leave --update=none as is (will be synonymous with -n)
    - Provide --update=none-fail to diagnose and exit failure


Thanks, that's a better proposal, but I still see several opportunities
for confusion.

If I understand things correctly, cp --update=none is not synonymous
with the proposed (i.e., old-behavior) cp -n, because -n overrides
previous -i options but --update=none does not. Also, -n overrides
either previous or following --update=UPDATE options, but --update=none
overrides only previous --update=UPDATE options. (For what it's worth,
FreeBSD -n overrides


Good point.
Well -n is a protection mechanism really, so should override.
Since --update now incorporates a protection mode, it should also override I 
think.


Some of this complication seems to be for consistency with how mv
behaves with -f, -i, -n, and --update, and similarly with how rm behaves
with -f, -i, -I, and --interactive. To be honest I don't quite
understand the reason for all this complexity, which suggests it should
be documented somewhere (the manual?) if it isn't already.


To my mind the most protective option takes precedence.
So from least to most that would be -f -n --update=none --update=none-fail -i
So the main consideration there is that -n should not override 
--update=none-fail


This raises more questions:

* If we deprecate cp -n, what about mv -n? FreeBSD mv -n behaves like
Coreutils mv -n: it silently does nothing and exits successfully. So
there's no compatibility argument for changing mv -n's behavior.
However, whatever --update option we add to cp (to output a diagnostic
and exit with failure) should surely also be added to mv, to aid
consistency.


Yes I agree.


* Should cp --update=none be changed so that it really behaves like the
old cp -n, in that it overrides other options in ways that differ from
how the other --update=UPDATE options behave? I'm leaning toward "no" as
this adds complexity that I don't see the use for.

* If we don't change cp --update=none's overriding behavior, is it still
OK to tell users to substitute --update=none for -n even though the two
options are not exactly equivalent? I'm leaning towards "yes" but would
like other opinions.


Yes I think we should still give that advice to deprecate -n,
if we ensure that -n does not override --update=none=fail.

thanks,
Pádraig

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-30 Thread Paul Eggert 

On 2024-01-30 03:18, Pádraig Brady wrote:

So we now have the proposed change as:

   - revert -n to old silent success behavior
   - document -n as deprecated
   - Leave --update=none as is (will be synonymous with -n)
   - Provide --update=none-fail to diagnose and exit failure


Thanks, that's a better proposal, but I still see several opportunities 
for confusion.


If I understand things correctly, cp --update=none is not synonymous 
with the proposed (i.e., old-behavior) cp -n, because -n overrides 
previous -i options but --update=none does not. Also, -n overrides 
either previous or following --update=UPDATE options, but --update=none 
overrides only previous --update=UPDATE options. (For what it's worth, 
FreeBSD -n overrides


Some of this complication seems to be for consistency with how mv 
behaves with -f, -i, -n, and --update, and similarly with how rm behaves 
with -f, -i, -I, and --interactive. To be honest I don't quite 
understand the reason for all this complexity, which suggests it should 
be documented somewhere (the manual?) if it isn't already.


This raises more questions:

* If we deprecate cp -n, what about mv -n? FreeBSD mv -n behaves like 
Coreutils mv -n: it silently does nothing and exits successfully. So 
there's no compatibility argument for changing mv -n's behavior. 
However, whatever --update option we add to cp (to output a diagnostic 
and exit with failure) should surely also be added to mv, to aid 
consistency.


* Should cp --update=none be changed so that it really behaves like the 
old cp -n, in that it overrides other options in ways that differ from 
how the other --update=UPDATE options behave? I'm leaning toward "no" as 
this adds complexity that I don't see the use for.


* If we don't change cp --update=none's overriding behavior, is it still 
OK to tell users to substitute --update=none for -n even though the two 
options are not exactly equivalent? I'm leaning towards "yes" but would 
like other opinions.

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-30 Thread Pádraig Brady 

On 29/01/2024 21:44, Paul Eggert wrote:

On 1/29/24 08:11, Pádraig Brady wrote:


Right, that's why I'm still leaning towards my proposal in the last mail.


Well, I won't insist on doing nothing; however, the proposal needs
ironing out and now's a good time to do it before installing changes.



    - revert to previous exit success -n behavior
    - document -n as deprecated
    - provide --update=noclobber to give exit failure functionality


So --update=noclobber would differ in meaning from the deprecated-in-9.5
--no-clobber, but would agree in meaning with 9.4 --no-clobber? That
sounds pretty confusing for future users. (And a nit: why should one
spelling have a hyphen but the other doesn't?)


That's a fair point; just avoid "no-clobber" entirely.
We could choose a new name so. How about --update=none-fail


      - BTW, it probably makes sense to print a diagnostic for each
skipped file here
    as it's exceptional behavior, for which we're exiting with
failure for.


Coreutils 9.4 cp -n already does that, no? So I'm not sure what's being
proposed here.

$ touch a b
$ cp -n a b; echo $?
cp: not replacing 'b'
1


Sorry I got confused between your suggestion of added diagnostics,
and FreeBSD's silent behavior here. Looks like we just keep the
existing behavior of diagnosing 'not replacing' iff exiting with failure.


    - the existing --update=none provides the exit success functionality


It seems to me that this proposal conflates two questions:

* What rules should cp use to decide whether to update a destination?

* When cp decides not to update a destination, what should it do? Exit
with nonzero status? Output a diagnostic? Both? Neither?

Aren't these independent axes? If so, shouldn't they have independent
options? For example, since we have --update=older, shouldn't there be a
way to say "I want to copy A to B only if B is older than A, and I want
the exit status to be zero only if A was copied to B"?


Well they're not entirely independent.
It's more appropriate to output a diagnostic if exiting failure,
and POSIX also advises along the same lines.
We also have --verbose to control diagnostics somewhat
for the non failure case.

So we now have the proposed change as:

  - revert -n to old silent success behavior
  - document -n as deprecated
  - Leave --update=none as is (will be synonymous with -n)
  - Provide --update=none-fail to diagnose and exit failure

thanks,
Pádraig.

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-29 Thread Paul Eggert 

On 1/29/24 08:11, Pádraig Brady wrote:


Right, that's why I'm still leaning towards my proposal in the last mail.


Well, I won't insist on doing nothing; however, the proposal needs 
ironing out and now's a good time to do it before installing changes.




   - revert to previous exit success -n behavior
   - document -n as deprecated
   - provide --update=noclobber to give exit failure functionality


So --update=noclobber would differ in meaning from the deprecated-in-9.5 
--no-clobber, but would agree in meaning with 9.4 --no-clobber? That 
sounds pretty confusing for future users. (And a nit: why should one 
spelling have a hyphen but the other doesn't?)



     - BTW, it probably makes sense to print a diagnostic for each 
skipped file here
   as it's exceptional behavior, for which we're exiting with 
failure for.


Coreutils 9.4 cp -n already does that, no? So I'm not sure what's being 
proposed here.


  $ touch a b
  $ cp -n a b; echo $?
  cp: not replacing 'b'
  1



   - the existing --update=none provides the exit success functionality


It seems to me that this proposal conflates two questions:

* What rules should cp use to decide whether to update a destination?

* When cp decides not to update a destination, what should it do? Exit 
with nonzero status? Output a diagnostic? Both? Neither?


Aren't these independent axes? If so, shouldn't they have independent 
options? For example, since we have --update=older, shouldn't there be a 
way to say "I want to copy A to B only if B is older than A, and I want 
the exit status to be zero only if A was copied to B"?

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-29 Thread Michael Stone 

On Mon, Jan 29, 2024 at 04:11:05PM +, Pádraig Brady wrote:

You've introduced a silent incompatibility and I'm trying to find some
way to make that clear. If upstream would provide a better solution I
would certainly use it. I have despaired of there being such since your
attitude thus far seems to be entirely dismissive of compatibility
concerns.


That's a bit unfair.  The current upstream -n behavior is with a view
to being _more_ compat across all systems.
Now I agree this may not be worth it in this case,
but it is a laudable goal.


You are saying that again without explicitly acknowledging that "being 
_more_ compat" in this case means "becoming _incompat_ with the vast 
majority of installed systems". IMO it could be reasonably phrased as 
"being more compatible across all systems in the long term when all 
existing legacy systems are gone", but the key here is that I read 
"_more_ compat across all systems" as dismissing the coreutils installed 
base as part of "all systems". I understand that may not be/have been 
the intent, but I also can't help feeling the way that I do when the 
benefits of compatability with freebsd are repeatedly emphasized while 
the costs of incompatibility with the coreutils installed base are 
dismissed with something along the lines of "we'll see what breaks". (If 
the costs of incompatibility are really that low in this case, why would 
compatability be a worthwhile goal in this case?)


I do wish that more users had noticed the change earlier and that we're 
fairly deep into a mess, but it's not always easy to see the impact of 
what seems like a relatively minor patch. I do appreciate that the new 
version printed some diagnostics when the change was triggered, as that 
certainly helped call attention to scripts which were impacted.



With the above in place for the next coreutils release,
then debian could remove its noisy patch.


I would certainly align with that, and the sooner the better to decrease 
the chances that different distributions handle this in different ways 
or we get to the point of having to release in an interim state. If you 
commit a final version I'll apply that patch if the next release isn't 
imminent.

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-29 Thread Pádraig Brady 

On 29/01/2024 14:01, Michael Stone wrote:

On Sun, Jan 28, 2024 at 11:14:14PM -0800, Paul Eggert wrote:

I'm not sure reverting would be best. It would introduce more
confusion, and would make coreutils incompatible with FreeBSD again.


Reverting makes more sense than the current situation. I do not
understand why you seem to value FreeBSD compatibility more than
compatibility with the vast majority of installed coreutils/linux
systems.


Yes, it's not a good place to be. Surely current coreutils is better
than what Debian is doing.


You've introduced a silent incompatibility and I'm trying to find some
way to make that clear. If upstream would provide a better solution I
would certainly use it. I have despaired of there being such since your
attitude thus far seems to be entirely dismissive of compatibility
concerns.


That's a bit unfair.  The current upstream -n behavior is with a view
to being _more_ compat across all systems.
Now I agree this may not be worth it in this case,
but it is a laudable goal.


Another possibility is to add a warning that is emitted only at the
end of 'cp'. The warning would occur only if the exit code differs
because of this cp -n business.


You'd only emit a notification of a change in behavior if some
(potentially uncommon/rarely encountered) situation arises which would
actually trigger breakage? So people can't prepare ahead of time and
change their script to handle the necessary change in logic, they can
only maybe figure out why something broke at 2am when the uncommon event
occurred?


Yes I agree with this point, mostly.
Outputting a diagnostic would help users diagnose what's going on,
and help users to fix forward and avoid their problematic -n usage.
But yes, the crux of the issue with fixing issues as they occur,
is it depends on the state of the destination and so can become an issue
at any time.  Now I previously did an audit with github and debian code search
and noticed very few problematic uses of cp -n, but that does miss
the mountain of private code.


At the end of the day, -n is basically a useless option with unknowable
semantics which should be avoided by everyone. In the past it was an
option which wasn't portable between coreutils/linux and freebsd systems,
and I guess you've "fixed" that (by making it an option everyone should
avoid entirely), but let's be honest about how common that concern was.


Right, that's why I'm still leaning towards my proposal in the last mail.
  - revert to previous exit success -n behavior
  - document -n as deprecated
  - provide --update=noclobber to give exit failure functionality
- BTW, it probably makes sense to print a diagnostic for each skipped file 
here
  as it's exceptional behavior, for which we're exiting with failure for.
  - the existing --update=none provides the exit success functionality

With the above in place for the next coreutils release,
then debian could remove its noisy patch.

thanks,
Pádraig

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-29 Thread Michael Stone 

On Sun, Jan 28, 2024 at 11:14:14PM -0800, Paul Eggert wrote:
I'm not sure reverting would be best. It would introduce more 
confusion, and would make coreutils incompatible with FreeBSD again.


Reverting makes more sense than the current situation. I do not 
understand why you seem to value FreeBSD compatibility more than 
compatibility with the vast majority of installed coreutils/linux

systems.

Yes, it's not a good place to be. Surely current coreutils is better 
than what Debian is doing.


You've introduced a silent incompatibility and I'm trying to find some 
way to make that clear. If upstream would provide a better solution I 
would certainly use it. I have despaired of there being such since your 
attitude thus far seems to be entirely dismissive of compatibility 
concerns.


Another possibility is to add a warning that is emitted only at the 
end of 'cp'. The warning would occur only if the exit code differs 
because of this cp -n business.


You'd only emit a notification of a change in behavior if some 
(potentially uncommon/rarely encountered) situation arises which would 
actually trigger breakage? So people can't prepare ahead of time and 
change their script to handle the necessary change in logic, they can 
only maybe figure out why something broke at 2am when the uncommon event 
occurred?


At the end of the day, -n is basically a useless option with unknowable 
semantics which should be avoided by everyone. In the past it was an 
option which wasn't portable between coreutils/linux and freebsd systems, 
and I guess you've "fixed" that (by making it an option everyone should 
avoid entirely), but let's be honest about how common that concern was.

Bug#1058752: bug#62572: Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2024-01-28 Thread Paul Eggert 

On 2024-01-28 05:22, Pádraig Brady wrote:

At this stage it seems best for us go back to the original Linux 
behiavor (use case 3),
and to silently deprecate -n in docs to document the portability issues 
with it.


I'm not sure reverting would be best. It would introduce more confusion, 
and would make coreutils incompatible with FreeBSD again.


The recent Debian change indicates that their intent is to move to the 
FreeBSD behavior too. This would improve cross-platform portability and 
I don't think we should discourage that.




  $ cp -n /bin/true tmp
  cp: warning: behavior of -n is non-portable and may change in future; use 
--update=none instead

This is problematic as:

  - It's noisy


Yes that's a problem, and I doubt whether we should mimic Debian.


  - There is no way to get the behavior of indicating failure if existing files 
present


Yes, it's not a good place to be. Surely current coreutils is better 
than what Debian is doing.



  - The --update=none advice is only portable to newer coreutils


True, but that's not a deal-killer. No advice that we give can be 100% 
portable to all platforms.



We should also provide --update=noclobber for use case 1.
Having the control on the --update option, allows use to more clearly 
deprecate -n.


Adding an --update=noclobber sounds like a good thing to do.

Another possibility is to add a warning that is emitted only at the end 
of 'cp'. The warning would occur only if the exit code differs because 
of this cp -n business. We could stretch things a bit and have a 
configure-time option --enable-compat-warnings that builders like Debian 
could use if they want such warnings.

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-17 Thread Michael Stone 

On Sun, Dec 17, 2023 at 12:34:11AM -0800, Paul Eggert wrote:

On 2023-12-16 13:46, Bernhard Voelker wrote:

Whether the implementation is race-prone or not is an internal thing.


I wasn't referring to the internal implementation. I was referring to 
cp users. With the newer Coreutils (FreeBSD) behavior, you can 
reliably write a script to do something if cp -n didn't copy the file 
because the destination already existed. With the older Coreutils 
behavior you cannot do that reliably; there will always be a race 
condition.


You can now reliably write a script using the new long option. Changing 
the behavior of the short option helped nobody.

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-17 Thread Pádraig Brady 

On 16/12/2023 21:46, Bernhard Voelker wrote:

On 12/15/23 21:13, Michael Stone wrote:

On Fri, Dec 15, 2023 at 11:21:06AM -0800, Paul Eggert wrote:

Stlll, Pádraig gave a reasonable summary of why the change was made,


To clarify my summary a little, there I said that -n now _immediately_ fails.
I should have said _silently_ fails.  I.e. the complete copy operation
proceeds as before, and only the exit status is at issue here.


despite its incompatibility with previous behavior. (One thing I'd add
is that the FreeBSD behavior is inherently less race-prone.)


Whether the implementation is race-prone or not is an internal thing.
I think we're currently discussing more on a user-perspective level.

IIUC then the question is whether `cp -n` should continue to behave like
the (new) `cp --update=none` which returns EXIT_SUCCESS.

Regardless what other implementations do, when reading the -n description
from a user's point of view:

-n, --no-clobber do not overwrite an existing file (overrides a
   -u or previous -i option). See also --update

then I'd expect the tool to just skip existing files like `rsync 
--ignore-existing`
does.  In that regard I would be surprised if skipping files would result in an 
error.
Well, I would understand if there'd be a '--no-clobber=fail' option.


Agreed we should improve the docs a bit for this option.
I'll apply this at least:

diff --git a/doc/coreutils.texi b/doc/coreutils.texi
index 1f8b356d1..bf0f424d3 100644
--- a/doc/coreutils.texi
+++ b/doc/coreutils.texi
@@ -9057,6 +9057,8 @@ Do not overwrite an existing file; silently fail instead.
 This option overrides a previous
 @option{-i} option.  This option is mutually exclusive with @option{-b} or
 @option{--backup} option.
+See also the @option{--update=none} option which will
+skip existing files but not fail.

 @item -P
 @itemx --no-dereference
diff --git a/src/cp.c b/src/cp.c
index 04a5cbee3..3ccc4c4e6 100644
--- a/src/cp.c
+++ b/src/cp.c
@@ -192,8 +192,8 @@ Copy SOURCE to DEST, or multiple SOURCE(s) to DIRECTORY.\n\
   -L, --dereferencealways follow symbolic links in SOURCE\n\
 "), stdout);
   fputs (_("\
-  -n, --no-clobber do not overwrite an existing file (overrides 
a\n\
- -u or previous -i option). See also 
--update\n\
+  -n, --no-clobber ensure no existing files overwritten, and 
fail\n\
+ silently instead. See also --update\n\
 "), stdout);
   fputs (_("\
   -P, --no-dereference never follow symbolic links in SOURCE\n\



As Kamil added the option in 2009, I'd assume that the same patch was already
active in RHEL versions for quite some longer time.
Now changing the exit code feels kind of rough.


Well RHEL 6 came out a bit after (2010), and had the --no-clobber change,
while RHEL 5 before that did not.

Taking about distros, it's worth noting that the change is Fedora 39
which has been released for a month now.
We'll keep a close eye on issues, but haven't heard much as
of yet at least.


Therefore, from a pure user's perspective and regarding many years of 
precedence,
I am 80:20 for reverting the exit code change.


Thanks for your thoughts,
appreciated as always.

cheers,
Pádraig

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-17 Thread Dominique Martinet 
Paul Eggert wrote on Fri, Dec 15, 2023 at 11:21:06AM -0800:
> The cat is to some extent out of the bag. Unless one insists on (FreeBSD |
> coreutils 9.2-9.4), or insist on coreutils 7.1-9.1, one should not rely on
> cp -n failing or silently succeeding when the destination already exists.
> This will remain true regardless of whether coreutils reverts to its 7.1-9.1
> behavior.

This. Scripts that want to be portable already can't assume cp -n will
do what they want, so at this point it doesn't really matter what
coreutils does in the grand scheme of things.

For distros like debian since even -testing hasn't seen coreutils 9.2,
there's still value in reverting locally (with a warning that it's not
reliable perhaps?), but in general coreutils 9.2 has been out for 9
months (2023 March 20), so many systems can already be considered
affected; but it's a disservice to users to just try to hide the problem
under the rug.


(To give a data point, this did bite us as well, and I was annoyed
enough that I went to look for the old bug report back in September, but
at that point 9.3 had already been out and I had given up without
reporting anything as nothing would change the fact that my scripts
would need updating. For the gory details I also need compatibility
with busybox cp (where -n silently ignores existing files), so
--update=none is not an option, but I for this particular usage I
settled for '-u' (--update=older, that busybox also support as short
option only...), and I since hurried to forget about it)

-- 
Dominique Martinet | Asmadeus

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-17 Thread Paul Eggert 

On 2023-12-16 13:46, Bernhard Voelker wrote:

Whether the implementation is race-prone or not is an internal thing.


I wasn't referring to the internal implementation. I was referring to cp 
users. With the newer Coreutils (FreeBSD) behavior, you can reliably 
write a script to do something if cp -n didn't copy the file because the 
destination already existed. With the older Coreutils behavior you 
cannot do that reliably; there will always be a race condition.

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-15 Thread Michael Stone 

On Fri, Dec 15, 2023 at 11:21:06AM -0800, Paul Eggert wrote:
Stlll, Pádraig gave a reasonable summary of why the change was made, 
despite its incompatibility with previous behavior. (One thing I'd add 
is that the FreeBSD behavior is inherently less race-prone.) It seemed 
like a good idea at the time all things considered, and to my mind 
still does.


I think you underestimate the value of maintaining compatibity with 
deployed versions. In the abstract it may have been a nice cleanup, but 
there are a lot of dumb things in the posix utilities that have been 
dumb for so long it's not worth the pain of changing them. Since this 
change hasn't yet hit mainstream debian, ubuntu, rhel, or suse users, I 
strongly suspect that this is a case where the absence of complaints is 
simply a sign that most of the people who'd be impacted haven't 
experienced the change yet.


Even if we tell people not to use -n at all, that doesn't mean we 
should revert to the coreutils 9.1 behavior.


It does, IMO, as it would be less likely to break scripts written by 
existing coreutils users.


The cat is to some extent out of the bag. Unless one insists on 
(FreeBSD | coreutils 9.2-9.4), or insist on coreutils 7.1-9.1, one 
should not rely on cp -n failing or silently succeeding when the 
destination already exists. This will remain true regardless of 
whether coreutils reverts to its 7.1-9.1 behavior.


Or you use a distribution that has to patch to maintain compatibility 
between versions. Ideally upstream would revert the behavior for now, 
deprecate as the long term fix, and all distributions would work the 
same. The other option is that each distribution decides whether to be 
compatible with upstream coreutils or their own previous release.

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-15 Thread Paul Eggert 

On 2023-12-15 10:49, Michael Stone wrote:

There's no compelling reason to force this change


Well, certainly nobody compelled us at gunpoint

Stlll, Pádraig gave a reasonable summary of why the change was made, 
despite its incompatibility with previous behavior. (One thing I'd add 
is that the FreeBSD behavior is inherently less race-prone.) It seemed 
like a good idea at the time all things considered, and to my mind still 
does.




Essentially the current situation is that -n shouldn't be used if you expect a 
certain behavior for this case and you are writing a script for linux systems. 
Maybe in 10 years you'll be able to assume the new behavior. Better to just 
tell people to not use it at all, and leave the historic behavior alone until 
everyone has stopped using -n entirely.


Even if we tell people not to use -n at all, that doesn't mean we should 
revert to the coreutils 9.1 behavior.


The cat is to some extent out of the bag. Unless one insists on (FreeBSD 
| coreutils 9.2-9.4), or insist on coreutils 7.1-9.1, one should not 
rely on cp -n failing or silently succeeding when the destination 
already exists. This will remain true regardless of whether coreutils 
reverts to its 7.1-9.1 behavior.

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-15 Thread Michael Stone 

On Fri, Dec 15, 2023 at 06:33:00PM +, Pádraig Brady wrote:

Advantages of leaving as is:
We get consistency of "noclobber" behavior across systems / shells.


You don't, unless you ignore the coreutils/linux installed base 
entirely. Essentially the current situation is that -n shouldn't be used 
if you expect a certain behavior for this case and you are writing a 
script for linux systems. Maybe in 10 years you'll be able to assume 
the new behavior. Better to just tell people to not use it at all, and 
leave the historic behavior alone until everyone has stopped using -n 
entirely.



There is no potential for data loss etc.


There may not be, strictly speaking, if you look only at cp without 
context, but we have absolutely no idea what the impact is on the 
unknown number of existing scripts that depend on the historic behavior. 
This is causing breakages in practice.



so it just comes
down to how disruptive it is, or how often -n was used
with the "skip behavior" assumption.


IMO, it should come down to trying to avoid breaking changes in core 
system utilities. There's no compelling reason to force this change, so 
why break anything that depended on the historic behavior--especially 
without any notice or transition period--regardless of arguments over 
whether the historic behavior was right?



We've not had much push back as of yet,
and my current thinking is it's not that disruptive a change.


I suspect that's because it has not yet been widely deployed, which 
makes now the time to fix it.


Michael Stone

Bug#1058752: bug#62572: cp --no-clobber behavior has changed

2023-12-15 Thread Pádraig Brady 

On 15/12/2023 15:56, Michael Stone wrote:

I tend to think this was a serious mistake: it breaks the behavior of
existing scripts with no deprecation period. A stated advantage is
better compatibility with freebsd, but I don't understand why that is
more desirable than compatibility with all deployed gnu/linux systems? I
also don't think it's sufficient to try to lawyer out by saying that the
current behavior was undocumented: the previous documentation said that
-n would "silently do nothing" and that the return code would be zero on
success. Logically, unless cp fails to "do nothing", it should exit with
a zero code.

Such a drastic change in behavior demands a new flag, not a radical
repurposing of a widely used existing flag.

I was hoping to see more action on this bug, but that hasn't happened.
I'm not sure I see a way forward for debian other than reverting to the
old behavior. I am reluctant to do so as that will likely lead to
divergent behavior between distributions, but breaking scripts without a
compelling reason is also not good. I would encourage coreutils to
reconsider the change and finding a non-breaking way forward.


Yes it's a fair point.
It's an awkward case, and worth discussing.

To summarise:

  coreutils >= 7.1 had -n skip existing in dest (2009)
  coreutils >= 9.2 has -n immediately fail if existing in dest
  coreutils >= 9.3 has --update=none to skip existing in dest

  FreeBSD >= 4.7/macos has -n immediately fail if existing in dest

  bash has noclobber as a file protection mechanism,
  and fails immediately upon trying to overwrite a file.
  This is more consistent with the new coreutils behavior.

I see a reasonable amount of cp -n usage across github:
https://github.com/search?q=/cp+.*+-n+.*/+path:*.sh=code

Now it's not clear which behavior these github usages expect,
and the original docs didn't make it clear which behavior to expect.
A quick scan of the github usages also seem mainly to expect
a protection rather than an update use case, so failing
immediately would be the most appropriate action there too.
Also the original coreutils bug report here expected the new behaviour.

So we probably all agree that failing immediately is the
most appropriate / consistent -n behavior,
but GNU had diverged from that so there are about 10 years
of scripts that may expect the silent skip behavior.

Two options I see are:

- Leave as is and fix -n usages that expected the skip behavior
- Deprecate -n entirely and prompt to use --update={fail,none}

Advantages of leaving as is:
We get consistency of "noclobber" behavior across systems / shells.
We fix cases where previously scripts could have proceeded with
stale old files in place.

Disadvantages of leaving as is:
Users expecting the skip behavior, have to change to --update=none.

There is no potential for data loss etc. so it just comes
down to how disruptive it is, or how often -n was used
with the "skip behavior" assumption.

We've not had much push back as of yet,
and my current thinking is it's not that disruptive a change.
So I'd be 55:45 if favor of keeping things as is.

thanks,
Pádraig.