"error": "sync: creat(xxx) failed: Not a directory: not a directory"

[benpeart]

When testing watchman with git submodules, I execute the following query:

["query", "C:/git-sdk-64/usr/src/git/t/trash directory.t7506-status-submodule/sub", {"since": 1499713020, "fields": ["name"], "expression": ["not", ["allof", ["since", 1499713020, "cclock"], ["not", "exists"]]]}]

Which results in the following error:

{
"version": "4.9.0",
"error": "sync: creat(C:/git-sdk-64/usr/src/git/t/trash directory.t7506-status-submodule/sub/.git/.watchman-cookie-BenPeart-HP-58768-13) failed: Not a directory: not a directory"
}

$ watchman version
{
"version": "4.9.0",
"buildinfo": "981452aff98ee5064c06a3a8cd39ff2394846114 2017-07-03T21:52:36.0000000Z"
}

It appears that there is some git specific code in Watchman (note the path contains "/sub/.git/.watchman" which in the case of sub modules, is invalid. Sub-modules take advantage of the git feature that allows the .git folder to be replaced with a .git file that contains the path to the actual folder. Here is the contents of the sub/.git file:

$ cat .git
gitdir: C:/git-sdk-64/usr/src/git/t/trash directory.t7506-status-submodule/sub/../.real

I'm surprised/worried that Watchman has git specific logic but if it needs to continue to exist, Watchman will need to handle the case where .git is a file instead of a folder as this can happen in any .git repo in addition to being used for sub-modules.

I also wonder if this (creating a cookie) is contributing to the random errors where a file is in use within a folder which prevents the folder from being deleted. I'll report that separately.

[wez]

Hmm, sounds fishy. We don't have git specific logic, but we do have this bit of logic:

watchman/root/vcs.cpp

Lines 52 to 55 in a1bd090

	auto info = getFileInformation(fullname.c_str(), case_sensitive);
	if (info.isDir()) {
	// root/{.hg,.git,.svn}
	cookies.setCookieDir(fullname);

The idea is that we want to place our sync cookies in a directory that doesn't participate in the SCM status operation in case we get killed and leave behind a junk file in the repo. We do this based on the vcs_ignore_dirs setting; this is just a list of dirs that we partially ignore (we watch only its direct contents, but not anything lower).

We should only pick one of those dirs if it shows up as a dir, and that should make that "not a directory" error impossible.

However, we do have this long standing issue that surfaced from an earlier attempt at integrating watchman with git: #25

Is it possible that that path in the error message was originally a nested git repo and that it was later swapped out for a subrepo?

[benpeart]

I looked into the test that was failing and yes, it does start out with a .git folder and then replaces it in a later test with a .git file. I read your proposed fix and it seems like it would solve the problem.