On 2017-04-24 19:00, Torsten Bögershausen wrote:

On 2017-04-24 18:45, Lars Schneider wrote:

quoted

Hi,

"t0025.3 - crlf=true causes a CRLF file to be normalized" failed 
sporadically on next and master recently: 
https://travis-ci.org/git/git/jobs/225084459#L2382
https://travis-ci.org/git/git/jobs/223830505#L2342

Are you aware of a race condition in the code
or in the test?

Not yet - I'll have a look

So,
The test failed under Linux & pu of the day, using Peff's
stress script.

not ok 3 - crlf=true causes a CRLF file to be normalized

The good case (simplified):
$ git status
   modified:   CRLFonly

Untracked files:
        .gitattributes


$ git diff | tr "\015" Q
warning: CRLF will be replaced by LF in CRLFonly.
The file will have its original line endings in your working directory.

diff --git a/CRLFonly b/CRLFonly
index 44fc21c..666dbf4 100644
--- a/CRLFonly
+++ b/CRLFonly

@@ -1,7 +1,7 @@
-IQ
-amQ
-veryQ
-veryQ
-fineQ
-thankQ
-youQ
+I
+am
+very
+very
+fine
+thank
+you

--------------------

The failed case:
$ git status
Untracked files:
        .gitattributes

-----------------------
$ ls -al -i
total 28
3430195 drwxr-xr-x 3 tb tb 4096 Apr 24 21:19 .
3427617 drwxr-xr-x 3 tb tb 4096 Apr 24 21:19 ..
3429958 -rw-r--r-- 1 tb tb   37 Apr 24 21:19 CRLFonly
3432574 drwxr-xr-x 8 tb tb 4096 Apr 24 21:27 .git
3425599 -rw-r--r-- 1 tb tb   14 Apr 24 21:19 .gitattributes
3430089 -rw-r--r-- 1 tb tb   24 Apr 24 21:19 LFonly
3430174 -rw-r--r-- 1 tb tb   36 Apr 24 21:19 LFwithNUL

-----------------
#After
$ mv CRLFonly tmp
$ cp tmp CRLFonly
$ ls -al -i
3430195 drwxr-xr-x 3 tb tb 4096 Apr 24 21:36 .
3427617 drwxr-xr-x 3 tb tb 4096 Apr 24 21:19 ..
3401599 -rw-r--r-- 1 tb tb   37 Apr 24 21:36 CRLFonly
3432574 drwxr-xr-x 8 tb tb 4096 Apr 24 21:36 .git
3425599 -rw-r--r-- 1 tb tb   14 Apr 24 21:19 .gitattributes
3430089 -rw-r--r-- 1 tb tb   24 Apr 24 21:19 LFonly
3430174 -rw-r--r-- 1 tb tb   36 Apr 24 21:19 LFwithNUL
3429958 -rw-r--r-- 1 tb tb   37 Apr 24 21:19 tmp

$ git status
	modified:   CRLFonly
Untracked files:
        .gitattributes
        tmp


So all in all it seams as if there is a very old race condition here,
which we "never" have seen yet.
Moving the file to a different inode number fixes the test case,
Git doesn't treat it as unchanged any more.