Bug report about symlinks

Greetings from Russia, comrads!

I've noticed something strange with git status when replacing a folder with symlink to another folder. There is a git repo with script with demo in the attachment.

Yours sincerely, NickKolok aka Nikolay Avdeev.

Доброго времени суток, товарищи!

При замене папки на симлинк на другую папку с похожими, но не одинаковыми файлами git status ведёт себя странно. Прилагаю архив с репозиторием. В скрипте - пример и пояснения.

С уважением, Николай Авдеев aka NickKolok.

Am 31.07.2014 um 21:50 schrieb Nikolay Avdeev:

Let's try and make this a bit easier for folks to follow along.

# Create test repo with two directories with two files each.
$ git init
Initialized empty Git repository in /tmp/.git/
$ mkdir a b
$ echo x >a/equal
$ echo x >b/equal
$ echo y >a/different
$ echo z >b/different
$ git add a b
$ git commit -minitial
[master (root-commit) 6d36895] initial
  4 files changed, 4 insertions(+)
  create mode 100644 a/different
  create mode 100644 a/equal
  create mode 100644 b/different
  create mode 100644 b/equal

# Replace one directory with a symlink to the other. $ rm -rf b $ ln -s a b

# First git status call.
$ git status
On branch master
Changes not staged for commit:
   (use "git add/rm <file>..." to update what will be committed)
   (use "git checkout -- <file>..." to discard changes in working directory)

         deleted:    b/different

Untracked files:
   (use "git add <file>..." to include in what will be committed)

         b

no changes added to commit (use "git add" and/or "git commit -a")

# Stage the changes. $ git add b

# Second git status call.
$ git status
On branch master
Changes to be committed:
   (use "git reset HEAD <file>..." to unstage)

         new file:   b
         deleted:    b/different
         deleted:    b/equal

# Commit the staged changes.
$ git commit -msymlinked
[master 4498f28] symlinked
  3 files changed, 1 insertion(+), 2 deletions(-)
  create mode 120000 b
  delete mode 100644 b/different
  delete mode 100644 b/equal

That the first and second status call report different results is a feature; staging changes using git add alters the status. A commit after the first status call would be empty.

It could be debated if the first git status call should also report b/equal as deleted.

René

René Scharfe <l.s.r@web.de> writes:

Hmph.

I wonder if "could be" is an understatement. The difference of reporting between b/equal and b/different may indicate that somebody is mistakenly comparing the index with these paths, without first checking each path with has_symlink_leading_path(), or something, no?

Or am I misreading the report and codepath?

Puzzled...

Am 01.08.2014 um 18:23 schrieb Junio C Hamano:

How about the patch below? Before it checks if an index entry exists in the work tree, it checks if its path includes a symlink. After the patch, git status reports b/equal as deleted, too. The test suite still passes.

And do we need to use the threaded_ variant of the function here?

René Scharfe <l.s.r@web.de> writes:

Honestly, I didn't expect the fix to be in the refresh-index code path, but doing this there sort of makes sense.

I think you, who dug to find out where to add the check, already know this, and I am writing this mainly for myself and for the list archive, but when the knee-jerk "has-syjmlink-leading-path missing?" reaction came to me, two obvious optimizations also came to my mind:

 - When checking a cache entry "a/b/c/d/e" and we find out "a/b/c"
   is a symbolic link, we mark it as ~CE_VALID, but at the same
   time, we learn "a/b/c/any/thing" are also ~CE_VALID with that
   check, so we _could_, after the has_symlink_leading_path once
   returns true, so there may be an opportunity to fast-forward the
   scan, marking all paths under "a/b/c" as ~CE_VALID.

 - After finding out "a/b/c" is a symbolic link to cause anything
   underneath marked as ~CE_VALID, we also know "a/" and "a/b/"
   exists as real directories, so a later check for "a/b/some/thing"
   can start checking at "a/b/some/" without checking "a/" and "a/b/".

The latter is more important as it is a much more common case that the shape of the working tree not to change.

But I think the implementation of has_symlink_leading_path() already has these optimizations built inside around the (unfortunately named) "struct cache_def", so it probably would not give us much boost to implement such a fast-forwarding of the scan.

Hmmm, this is a tangent, but you comment made me wonder if we also need to adjust preload_thread() in preload-index.c somehow, but we do not touch CE_UPTODATE there, so it probably is not necessary.

The caller of refresh_cache_ent() is walking an array of sorted pathnames aka istate->cache[] in a single-threaded fashion, possibly with a pathspec to limit the scan. Do you mean that we may want to make istate->cache[] scanned by multiple threads? I am not sure.

Am 03.08.2014 um 19:19 schrieb Junio C Hamano:

I found it through observation, not understanding. Just looked for stat/lstat calls executed by git status for b/different and b/equal using strace; debugging printfs told me where they came from.

The function calls ce_mark_uptodate(), which does set CE_UPTODATE. It calls threaded_has_symlink_leading_path() before lstat() already, however. (Since f62ce3de: Make index preloading check the whole path to the file.)

There are two direct callers (refresh_index(), refresh_cache_entry()) and several indirect ones. Do we have a way to detect unsynchronized parallel access to the has_symlink_leading_path()-cache? Checking the full callers-of-callers tree manually looks a bit scary to me.

No, I didn't want to suggest any performance improvements. I'm only interested in correctness for now.

René

René Scharfe <l.s.r@web.de> writes:

Yeah, by "we do not touch", I meant "for paths that is beyond a symlink, we do not touch" (i.e. we have that "continue" before lstat-match-then-mark sequence).

The threaded variant is not used anybody outside preload-index, so currently we should be OK, I would think.

Don't add paths with leading symlinks to the index while refreshing; we only track those symlinks themselves. We already ignore them while preloading (see read_index_preload.c).

Reported-by: Nikolay Avdeev <avdeev@math.vsu.ru>
Signed-off-by: Rene Scharfe <l.s.r@web.de>
---
 read-cache.c               |  8 ++++++++
 t/t7515-status-symlinks.sh | 43 +++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 51 insertions(+)
 create mode 100755 t/t7515-status-symlinks.sh

On Mon, Aug 4, 2014 at 12:19 AM, Junio C Hamano <gitster@pobox.com> wrote:

Just checking, you meant CE_UPTODATE, not CE_VALID, right? CE_VALID is only used with --assume-unchanged

Yeah my first thought is another flood of lstat(). But it looks like has_symlink_leading_path() tries hard to reduce lstat() already. We could disable this call in filesytems that do not support symlinks (e.g. vfat) but I guess that's just a minority.

Duy Nguyen <pclouds@gmail.com> writes:

Yup.  Thanks.