[CI]: Is t7527 known to be flakey?

The said test failed its linux-musl job in its first attempt, but re-running the failed job passed.

    https://github.com/git/git/actions/runs/3963948890/jobs/6792356234
    (seen@e096683 attempt #1 linux-musl)

    https://github.com/git/git/actions/runs/3963948890/jobs/6792850313
    (seen@e096683 attempt #2 linux-musl)

On 1/19/23 9:52 PM, Junio C Hamano wrote:

This is on Linux, so it would be using the linux inotify backend. Let me add Eric to the "To:" line for visibility. And see if he has experienced this during his development of it.

I've not looked at the inotify code so I can't say if there are races there or not. Tests that move directories feel like good candidates for race conditions -- since the daemon doesn't get a recursive view of the tree with inotify() and must simulate that and manage the individual directories, but again I don't want to assume that.

Jeff

Jeff Hostetler <git@jeffhostetler.com> writes:

Thanks for redirection.

On Thu, Jan 19, 2023 at 06:52:01PM -0800, Junio C Hamano wrote:

t7527 is quite slow, even with the right selection of test cases, but this little tweak makes it much faster:

  diff --git a/t/t7527-builtin-fsmonitor.sh b/t/t7527-builtin-fsmonitor.sh
  index 0e497ba98d..4210ef644c 100755
  --- a/t/t7527-builtin-fsmonitor.sh
  +++ b/t/t7527-builtin-fsmonitor.sh
  @@ -676,7 +676,10 @@ test_expect_success 'cleanup worktrees' '
   # cause incorrect results when the untracked-cache is enabled.
   
   test_lazy_prereq UNTRACKED_CACHE '
  -	git update-index --test-untracked-cache
  +	# This check takes a very long time, but I know it works on
  +	# my system, so let's fake it.
  +	#git update-index --test-untracked-cache
  +	true
   '
   
   test_expect_success 'Matrix: setup for untracked-cache,fsmonitor matrix' '

This with the right selection of test cases makes --stress practicable, and the test tends to fail after a handful of repetitions:

  ./t7527-builtin-fsmonitor.sh --stress -r 8,23-58

I saw different failures in multiple test cases, e.g.:

Unexpected output in case 55:

  expecting success of 7527.55 'Matrix[uc:true][fsm:true] move_directory_contents_deeper': 
  		matrix_clean_up_repo &&
  		$fn &&
  		if test $uc = false && test $fsm = false
  		then
  			git status --porcelain=v1 >.git/expect.$fn
  		else
  			git status --porcelain=v1 >.git/actual.$fn &&
  			test_cmp .git/expect.$fn .git/actual.$fn
  		fi
  	
  + matrix_clean_up_repo
  + git reset --hard HEAD
  HEAD is now at 1d1edcb initial
  + git clean -fd
  + move_directory_contents_deeper
  + mkdir T1/_new_
  + mv T1/F1 T1/F2 T1/T2 T1/_new_
  + test true = false
  + git status --porcelain=v1
  error: read error: Connection reset by peer
  error: could not read IPC response
  + test_cmp .git/expect.move_directory_contents_deeper .git/actual.move_directory_contents_deeper
  + test 2 -ne 2
  + eval diff -u "$@"
  + diff -u .git/expect.move_directory_contents_deeper .git/actual.move_directory_contents_deeper
  --- .git/expect.move_directory_contents_deeper	2023-01-21 09:47:12.677410349 +0000
  +++ .git/actual.move_directory_contents_deeper	2023-01-21 09:47:14.045448573 +0000
  @@ -7,3 +7,4 @@
    D T1/T2/T3/T4/F1
    D T1/T2/T3/T4/F2
   ?? T1/_new_/
  +?? dir1
  error: last command exited with $?=1
  not ok 55 - Matrix[uc:true][fsm:true] move_directory_contents_deeper

SIGPIPE in 'git status' cases 42, 43 and 55:

  expecting success of 7527.42 'Matrix[uc:false][fsm:true] move_directory_contents_deeper':
                  matrix_clean_up_repo &&
                  $fn &&
                  if test $uc = false && test $fsm = false
                  then
                          git status --porcelain=v1 >.git/expect.$fn
                  else
                          git status --porcelain=v1 >.git/actual.$fn &&
                          test_cmp .git/expect.$fn .git/actual.$fn
                  fi
  
  + matrix_clean_up_repo
  + git reset --hard HEAD
  HEAD is now at 1d1edcb initial
  + git clean -fd
  + move_directory_contents_deeper
  + mkdir T1/_new_
  + mv T1/F1 T1/F2 T1/T2 T1/_new_
  + test false = false
  + test true = false
  + git status --porcelain=v1
  error: last command exited with $?=141
  not ok 42 - Matrix[uc:false][fsm:true] move_directory_contents_deeper

  expecting success of 7527.43 'Matrix[uc:false][fsm:true] move_directory_up': 
  		matrix_clean_up_repo &&
  		$fn &&
  		if test $uc = false && test $fsm = false
  		then
  			git status --porcelain=v1 >.git/expect.$fn
  		else
  			git status --porcelain=v1 >.git/actual.$fn &&
  			test_cmp .git/expect.$fn .git/actual.$fn
  		fi
  	
  + matrix_clean_up_repo
  + git reset --hard HEAD
  HEAD is now at 1d1edcb initial
  + git clean -fd
  Removing T1/_new_/
  + move_directory_up
  + mv T1/T2/T3 T1
  + test false = false
  + test true = false
  + git status --porcelain=v1
  error: last command exited with $?=141
  not ok 43 - Matrix[uc:false][fsm:true] move_directory_up

  expecting success of 7527.55 'Matrix[uc:true][fsm:true] move_directory_contents_deeper':
                  matrix_clean_up_repo &&
                  $fn &&
                  if test $uc = false && test $fsm = false
                  then
                          git status --porcelain=v1 >.git/expect.$fn
                  else
                          git status --porcelain=v1 >.git/actual.$fn &&
                          test_cmp .git/expect.$fn .git/actual.$fn
                  fi
  
  + matrix_clean_up_repo
  + git reset --hard HEAD
  HEAD is now at 1d1edcb initial
  + git clean -fd
  + move_directory_contents_deeper
  + mkdir T1/_new_
  + mv T1/F1 T1/F2 T1/T2 T1/_new_
  + test true = false
  + git status --porcelain=v1
  error: last command exited with $?=141

I find it interesting that the output of 'git status' is redirected to a file in all these cases, and yet it gets a SIGPIPE.

I also saw the test hang a couple of times, e.g.:

  $ ./t7527-builtin-fsmonitor.sh --stress -r 8,23-58
  OK    6.1
  OK    7.1
  OK    1.1
  OK    2.1
  OK    3.1
  OK    5.1
  OK    4.1
  OK    0.1
  OK    6.2
  OK    1.2
  OK    2.2
  OK    7.2
  OK    5.2
  OK    0.2
  OK    4.2
  OK    6.3
  OK    7.3
  OK    2.3
  OK    0.3
  OK    4.3
  OK    6.4
  OK    7.4
  OK    2.4
  OK    0.4
  OK    4.4
  OK    6.5
  OK    7.5
  OK    2.5
  OK    0.5
  OK    4.5
  OK    6.6
  OK    7.6
  OK    2.6
  OK    0.6
  OK    4.6
  OK    6.7
  OK    7.7
  OK    2.7
  OK    0.7
  OK    4.7
  OK    6.8
  OK    7.8
  OK    2.8
  OK    0.8
  OK    4.8
  OK    6.9
  OK    7.9
  OK    2.9
  OK    0.9
  OK    4.9
  OK    6.10
  OK    7.10
  OK    2.10
  OK    0.10
  OK    4.10
  OK    6.11
  OK    7.11
  OK    2.11
  OK    0.11
  OK    4.11
  OK    6.12
  OK    7.12
  OK    2.12
  OK    0.12
  OK    4.12
  OK    6.13
  OK    7.13
  OK    2.13
  OK    0.13
  OK    6.14
  OK    4.13
  OK    7.14
  OK    2.14
  OK    0.14
  OK    6.15
  OK    4.14
  OK    2.15
  OK    7.15
  OK    0.15
  FAIL  7.16
  OK    6.16
  OK    2.16
  OK    4.15
  OK    0.16

At this point the test script should print the log of the failed job, but it hangs instead, as there are a number of stuck fsmonitor--daemon and status processes (notice how the stress test starts with 8 jobs, but the last repetition only has 4):

  $ ps aux |grep git
  szeder   1857100  0.0  0.1  72272  4452 pts/2    Sl+  21:40   0:00 /home/szeder/src/git/git fsmonitor--daemon run --detach --ipc-threads=8
  szeder   1857779  0.0  0.1   6560  4152 pts/2    S+   21:40   0:00 /home/szeder/src/git/git status --porcelain=v1
  szeder   1860020  0.0  0.1  88664  4312 pts/2    Sl+  21:40   0:00 /home/szeder/src/git/git fsmonitor--daemon run --detach --ipc-threads=8
  szeder   1860668  0.0  0.1   6560  4040 pts/2    S+   21:40   0:00 /home/szeder/src/git/git status --porcelain=v1
  szeder   1860749  0.0  0.1  96860  4528 pts/2    Sl+  21:40   0:00 /home/szeder/src/git/git fsmonitor--daemon run --detach --ipc-threads=8
  szeder   1861281  0.0  0.1   6560  4272 pts/2    S+   21:40   0:00 /home/szeder/src/git/git status --porcelain=v1

On 1/21/23 5:23 AM, SZEDER Gábor wrote:

The `git status` command talks to the running `git fsmonitor--daemon` on a Unix domain socket (or a Named Pipe on Windows), so a SIGPIPE is possible.

Was this on Linux or MacOS ?

Jeff

On Mon, Jan 23, 2023 at 11:56:53AM -0500, Jeff Hostetler wrote:

On an average-ish Linux (an Ubuntu LTS variant). So the issue is not specific to musl.

On 1/23/23 1:12 PM, SZEDER Gábor wrote:

OK, thanks. I wasn't worried about "musl", but rather whether you were running the stress test on a Linux or Mac.

Since they have different backends (inotify vs FSEvent) and all the code that touches the filesystem is different, it would best to start on the correct OS when trying to repro it.

Can you tell from your stess test whether the fsmonitor-daemon is crashing? (It might be subtle since the daemon is auto-started if necessary, so it might be crashing and silently getting restarted by the next command.)

I ask because a SIGPIPE in the client would make me think that the server suddenly closed the connection unexpectedly, like if it had SIGSEGV'd or something.

I won't have time to spin up a Linux VM until next week, so I won't be able to investigate this for a bit.

Jeff

On Wed, Jan 25, 2023 at 02:02:40PM -0500, Jeff Hostetler wrote:

Last time around I only looked at the failing test case, and didn't notice anything that might have indicated the cause of the SIGPIPE. This time I chanced to look a bit further up in the test log, and:

  expecting success of 7527.55 'Matrix[uc:true][fsm:true] move_directory_contents_deeper':
                  matrix_clean_up_repo &&
                  $fn &&
                  if test $uc = false && test $fsm = false
                  then
                          git status --porcelain=v1 >.git/expect.$fn
                  else
                          git status --porcelain=v1 >.git/actual.$fn &&
                          test_cmp .git/expect.$fn .git/actual.$fn
                  fi
  
  + matrix_clean_up_repo
  + git reset --hard HEAD
  HEAD is now at 1d1edcb initial
  + git clean -fd
  + move_directory_contents_deeper
  + mkdir T1/_new_
  + mv T1/F1 T1/F2 T1/T2 T1/_new_
  + test true = false
  + git status --porcelain=v1
  error: read error: Connection reset by peer
  error: could not read IPC response
  + test_cmp .git/expect.move_directory_contents_deeper .git/actual.move_directory_contents_deeper
  + test 2 -ne 2
  + eval diff -u "$@"
  + diff -u .git/expect.move_directory_contents_deeper .git/actual.move_directory_contents_deeper
  ok 55 - Matrix[uc:true][fsm:true] move_directory_contents_deeper
  
  expecting success of 7527.56 'Matrix[uc:true][fsm:true] move_directory_up':
                  matrix_clean_up_repo &&
                  $fn &&
                  if test $uc = false && test $fsm = false
                  then
                          git status --porcelain=v1 >.git/expect.$fn
                  else
                          git status --porcelain=v1 >.git/actual.$fn &&
                          test_cmp .git/expect.$fn .git/actual.$fn
                  fi
  
  + matrix_clean_up_repo
  + git reset --hard HEAD
  HEAD is now at 1d1edcb initial
  + git clean -fd
  Removing T1/_new_/
  + move_directory_up
  + mv T1/T2/T3 T1
  + test true = false
  + git status --porcelain=v1
  error: last command exited with $?=141
  not ok 56 - Matrix[uc:true][fsm:true] move_directory_up

Notice that "error: read error: Connection reset by peer" in the previous, still successful test case! I ran it a couple of times, and saw the same error message in the still successful '42 - Matrix[uc:false][fsm:true] move_directory_contents_deeper' followed by a SIGPIPE caused failure in the next test case. And now that I knew what to look for, I noticed this error message in the very first test failure I reported the other day, which didn't fail because of SIGPIPE, and in that case the error message was printed in the failed test case.

And there were a few cases that failed because of SIGPIPE but there were no error messages at all.

I can't say what caused these errors, but I doubt that anything segfaulted, because segfaults are logged in syslog, and I haven't found any such syslog entries coinciding with stress testing.