Re: [PATCH v2 3/5] scalar: enable built-in FSMonitor on `register`

[Victoria Dye <vdye@github.com>]

Derrick Stolee wrote:
> On 8/16/2022 7:58 PM, Matthew John Cheetham via GitGitGadget wrote:
> 
>> +#ifdef HAVE_FSMONITOR_DAEMON_BACKEND
>> +		/*
>> +		 * Enable the built-in FSMonitor on supported platforms.
>> +		 */
>> +		{ "core.fsmonitor", "true" },
>> +#endif
>> +	if (fsmonitor_ipc__is_supported() && start_fsmonitor_daemon())
>> +		return error(_("could not start the FSMonitor daemon"));
>> +
> 
> I initially worried if fsmonitor_ipc__is_supported() could use some
> run-time information to detect if FS Monitor is supported (say, existence
> of a network share or something). However, that implementation is
> currently defined as a constant depending on
> HAVE_FSMONITOR_DAEMON_BACKEND.
> 
> The reason I was worried is that we could enable core.fsmonitor=true based
> on the compile-time macro, but then avoid starting the daemon based on the
> run-time results. If we get into this state, would the user's 'git status'
> calls start complaining about the core.fsmonitor=true config because it is
> not supported?
> 
> The most future-proof thing to do might be to move the config write out of
> the set_recommended_config() and into start_fsmonitor_daemon(). Perhaps
> rename it to enable_fsmonitor() so it can fail due to writing the config
> _or_ for starting the daemon. The error message would change, then, too.

I spent some time digging into this, and I think gating both the config and
subsequent 'git fsmonitor--daemon start' on having platform *and* repository
support is a good idea. I'll update the next version to both set the
'core.fsmonitor' config and start the daemon only if the built-in FSMonitor
is fully supported.

(warning: long-winded tangent mostly unrelated to FSMonitor)

In the process of testing FSMonitor behavior, I think found other issues
with Scalar registration. Specifically, the test I wrote attempted to
'scalar register' a bare repo, since bare directories are incompatible with
FSMonitor. After seeing that FSMonitor was *not* incompatible with the
repository, I found that Scalar was 1) ignoring the bare repository and, as
a result, 2) identifying my Git clone (way above GIT_CEILING_DIRECTORIES) as
the "enlistment root". I think 1) might be fine as-is - uniformly ignoring
bare repos seems like a reasonable choice - but 2) seems like more of a
problem. 

Right now, 'setup_enlistment_directory()' searches for the repo root
beginning at directory '<dir>', which is either a user-provided path or
current working directory. It checks whether '<dir>' or '<dir>/src' is a
repo root: if so, it sets the enlistment info; otherwise, it repeats the
process with the parent of '<dir>' until the repo root is found. For
example, given the following directory structure:

somedir
└── enlistment
    ├── src
    │   └── .git
    └── test
        └── data

'scalar register somedir/enlistment/test/data' will search:

  * somedir/enlistment/test/data/src
  * somedir/enlistment/test/data
  * somedir/enlistment/test/src
  * somedir/enlistment/test
  * somedir/enlistment/src

The current usage of GIT_CEILING_DIRECTORIES relies on the fact that, when
invoking a normal 'git' command, 'setup_git_directory()' only searches
upwards from the current working directory to find the repo root; it's a
clear "yes" or "no" as to whether that search passes a ceiling directory.
Scalar isn't as clear, since it searches for the repo root both "downwards"
into '<dir>/src' *and* upwards through the parents of '<dir>'. It's not
totally clear to me what the "right" behavior for Scalar is, but my current
thought is to follow the same rules as 'setup_git_directory()', but for the
*enlistment* root rather than the repository root. It's more restrictive
than GIT_CEILING_DIRECTORIES on a normal git repo, e.g.:

1. 'GIT_CEILING_DIRECTORIES=somedir/enlistment git -C somedir/enlistment/src status' 
   is valid.
2. 'GIT_CEILING_DIRECTORIES=somedir/enlistment scalar register somedir/enlistment/src'
   is not valid.

but since Scalar works on the entire enlistment (not just the repo inside of
it), I think it makes sense to prevent it from crossing a ceiling directory
boundary.

What do you think? Hopefully my rambling wasn't too confusing (if it is,
please let me know what I can clarify). 

> 
> Or maybe I'm making a mountain out of a mole hill and what exists here is
> perfectly fine.
> 
>> +test_lazy_prereq BUILTIN_FSMONITOR '
>> +	git version --build-options | grep -q "feature:.*fsmonitor--daemon"
>> +'
> 
> It looks like we already have a FSMONITOR_DAEMON prereq in test-lib.sh.
> Should we use that instead?

Works for me, happy to reuse code wherever possible. :)

> 
> Thanks,
> -Stolee