Re: core.autocrlf=true causes `git apply` to fail on patch generated with `git diff-index HEAD --patch`

From: Anthony Sottile <asottile@umich.edu>
Date: Aug 1, 2017, 20:58 UTC
Message-ID: <CA+dzEB=3OMw_YM4K_a8dyDG_FwGavU382stXrEOkbYoyM4DSZQ@mail.gmail.com>
In-Reply-To: <287407ac-b0d0-ef24-4950-0982a2db9bed@web.de>

Here's where I'm hitting the problem described: https://github.com/pre-commit/pre-commit/issues/570

Note that `git -c core.autocrlf=false` apply patch fixes this situation, but breaks others.

Here's a testcase where `git -c core.autocrlf=false apply patch` causes a *different* patch failure:

``` #!/bin/bash set -ex

rm -rf foo git init foo cd foo

git config --local core.autocrlf true

# Commit lf into repository python3 -c 'open("foo", "wb").write(b"1\r\n2\r\n")' git add foo python3 -c 'open("foo", "wb").write(b"3\n4\n")'

# Generate a patch, check it out, restore it git diff --ignore-submodules --binary --no-color --no-ext-diff > patch python3 -c 'print(open("patch", "rb").read())' git checkout -- . git -c core.autocrlf=false apply patch ```

output:

```
+ rm -rf foo
+ git init foo
Initialized empty Git repository in /tmp/foo/.git/
+ cd foo
+ git config --local core.autocrlf true
+ python3 -c 'open("foo", "wb").write(b"1\r\n2\r\n")'
+ git add foo
+ python3 -c 'open("foo", "wb").write(b"3\n4\n")'
+ git diff --ignore-submodules --binary --no-color --no-ext-diff
warning: LF will be replaced by CRLF in foo.
The file will have its original line endings in your working directory.
+ python3 -c 'print(open("patch", "rb").read())'
b'diff --git a/foo b/foo\nindex 1191247..b944734 100644\n---
a/foo\n+++ b/foo\n@@ -1,2 +1,2 @@\n-1\n-2\n+3\n+4\n'
+ git checkout -- .
+ git -c core.autocrlf=false apply patch
error: patch failed: foo:1
```

My current workaround is:
- try `git apply patch`
- try `git -c core.autocrlf=false apply patch`

which seems to work pretty well.

Anthony

On Tue, Aug 1, 2017 at 1:47 PM, Torsten Bögershausen <tboegi@web.de> wrote:

Show 108 quoted lines

>
>
> On 08/01/2017 08:24 PM, Anthony Sottile wrote:
>>
>> Here's my minimal reproduction -- it's slightly far-fetched in that it
>> involves*committing crlf*  and
>>
>> then using `autocrlf=true` (commit lf, check out crlf).
>>
>> ```
>> #!/bin/bash
>> set -ex
>>
>> rm -rf foo
>> git init foo
>> cd foo
>>
>> # Commit crlf into repository
>> git config --local core.autocrlf false
>> python3 -c 'open("foo", "wb").write(b"1\r\n2\r\n")'
>> git add foo
>> git commit -m "Initial commit with crlf"
>>
>> # Change whitespace mode to autocrlf, "commit lf, checkout crlf"
>> git config --local core.autocrlf true
>> python3 -c 'open("foo", "wb").write(b"1\r\n2\r\n\r\n\r\n\r\n")'
>>
>> # Generate a patch, check it out, restore it
>> git diff --ignore-submodules --binary --no-color --no-ext-diff > patch
>> python3 -c 'print(open("patch", "rb").read())'
>> git checkout -- .
>> # I expect this to succeed, it fails
>> git apply patch
>> ```
>>
>> And here's the output:
>>
>> ```
>> + rm -rf foo
>> + git init foo
>> Initialized empty Git repository in/tmp/foo/.git/
>> + cd foo
>> + git config --local core.autocrlf false
>> + python3 -c 'open("foo", "wb").write(b"1\r\n2\r\n")'
>> + git add foo
>> + git commit -m 'Initial commit with crlf'
>> [master (root-commit) 02d3246] Initial commit with crlf
>>   1 file changed, 2 insertions(+)
>>   create mode 100644 foo
>> + git config --local core.autocrlf true
>> + python3 -c 'open("foo", "wb").write(b"1\r\n2\r\n\r\n\r\n\r\n")'
>> + git diff --ignore-submodules --binary --no-color --no-ext-diff
>> + python3 -c 'print(open("patch", "rb").read())'
>> b'diff --git a/foo b/foo\nindex bd956ea..fbf7936 100644\n---
>> a/foo\n+++ b/foo\n@@ -1,2 +1,5 @@\n 1\r\n 2\r\n+\r\n+\r\n+\r\n'
>> + git checkout -- .
>> + git apply patch
>> patch:8: trailing whitespace.
>>
>> patch:9: trailing whitespace.
>>
>> patch:10: trailing whitespace.
>>
>> error: patch failed: foo:1
>> error: foo: patch does not apply
>> ```
>>
>> I also tried with `git apply --ignore-whitespace`, but this causes the
>> line endings of the existing contents to be changed to*lf*  (there may
>> be two bugs here?)
>>
>> Thanks,
>>
>> Anthony
>
>
>
> I can reproduce you test case here.
>
> The line
>
> git apply patch
>
> would succeed, if you temporally (for the runtime of the apply command) set
> core.autocrlf to false:
>
> git -c core.autocrlf=false apply patch
>
> So this seems to be a bug (in a corner case ?):
>
> Typically repos which had been commited with CRLF should be normalized,
> which means that the CRLF in the repo are replaced by LF.
> So you test script is a corner case, for which Git has not been designed,
> It seems as if "git apply" gets things wrong here.
> Especially, as the '\r' is not a whitespace as a white space. but part
> of the line ending.
> So in my understanding the "--ignore-whitespace" option shouldn't affect
> the line endings at all.
>
> Fixes are possible, does anyone have a clue, why the '\r' is handled
> like this in apply ?
>
> And out of interest: is this a real life problem ?
>
>
>
>
>