git-annex should use smudge/clean filters.

### implementation todo list

* Reconcile staged changes into the associated files database, whenever
  the database is queried. This is needed to handle eg:

  ```
  	git add largefile
	git mv largefile othername
	git annex move othername --to foo
	# fails to drop content from associated file othername,
	# because it doesn't know it has that name
	# git commit clears up this mess
  ```

* Dropping a smudged file causes git status (and git annex status)
  to show it as modified,  because the timestamp has changed. 
  Getting a smudged file can also cause this.
  Upgrading a direct mode repo also leaves files in this state.
  User can use `git add` to clear it up, but better to avoid this,
  by updating stat info in the index.
  (May need to use libgit2 to do this efficiently, cannot find
  any plumbing except git-update-index, which is very inneficient for
  smudged files.)

* Checking out a different branch causes git to smudge all changed files,
  and write their content. This does not honor annex.thin. A warning
  message is printed in this case.  

  This is particularly wasteful when checking out an adjusted unlocked
  branch, which causes 2x the space to be used.  
  
  "git annex proxy" could be used to handle this.
  Make it run the git command with smudge filters disabled and then
  scan through the changed files in the work tree, and update pointer files
  to be hard links to their content.
  
  git-annex adjust and git-annex sync could both use that internally
  when checking out the adjusted branch, and merging a branch into HEAD.
  
  Since this approach modifies work tree files, it again causes git status
  to think files are modified. So, the above todo item about that needs to
  be sorted out first; it would not do for git annex adjust to cause
  the whole work tree to be considered to be modified!

* Eventually (but not yet), make v6 the default for new repositories.
  Note that the assistant forces repos into direct mode; that will need to
  be changed then, and it should enable annex.thin instead.
* Later still, remove support for direct mode, and enable automatic
  v5 to v6 upgrades.

### historical notes

2013: Currently, this does not look likely to work. In particular,
the clean filter needs to consume all stdin from git, which consists of the
entire content of the file. It cannot optimise by directly accessing
the file in the repository, because git may be cleaning a different
version of the file during a merge. 

So every `git status` would need to read the entire content of all
available files, and checksum them, which is too expensive.

> Update from GitTogether: Peff thinks a new interface could be added to
> git to handle this sort of case in an efficient way.. just needs someone
> to do the work. --[[Joey]] 

>> Update 2015: git status only calls the clean filter for files
>> that the index says are modified, so this is no longer a problem.
>> --[[Joey]]

### background

The clean filter is run when files are staged for commit. So a user could copy
any file into the annex, git add it, and git-annex's clean filter causes
the file's key to be staged, while its value is added to the annex.

The smudge filter is run when files are checked out. Since git annex
repos have partial content, this would not git annex get the file content.
Instead, if the content is not currently available, it would need to do
something like return empty file content. (Sadly, it cannot create a
symlink, as git still wants to write the file afterwards.)

So the nice current behavior of unavailable files being clearly missing due
to dangling symlinks, would be lost when using smudge/clean filters.
(Contact git developers to get an interface to do this?)

Instead, we get the nice behavior of not having to remeber to `git annex
add` files, and just being able to use `git add` or `git commit -a`,
and have it use git-annex when .gitattributes says to. Also, annexed
files can be directly modified without having to `git annex unlock`.

### configuration

In .gitattributes, the user would put something like "* filter=git-annex".
This way they could control which files are annexed vs added normally.

It would also be good to allow using this without having to specify
the files in .gitattributes. Just use "* filter=git-annex" there, and then
let git-annex decide which files to annex and which to pass through the
smudge and clean filters as-is. The smudge filter can just read a little of
its input to see if it's a pointer to an annexed file. The clean filter
could apply annex.largefiles to decide whether to annex a file's content or
not.

For files not configured this way in .gitattributes, git-annex could
continue to use its symlink method -- this would preserve backwards
compatability, and even allow mixing the two methods in a repo as desired.
(But not switching an existing repo between indirect and direct modes;
the user decides which mode to use when adding files to the repo.)

### clean

The trick is doing it efficiently. Since git a2b665d, v1.7.4.1,
something like this works to provide a filename to the clean script:

	git config --global filter.huge.clean huge-clean %f

This could avoid it needing to read all the current file content from stdin
when doing eg, a git status or git commit. Instead it is passed the
filename that git is operating on, in the working directory.
(Update: No, doesn't work; git may be cleaning a different file content
than is currently on disk, and git requires all stdin be consumed too.)

So, WORM could just look at that file and easily tell if it is one
it already knows (same mtime and size). If so, it can short-circuit and
do nothing, file content is already cached.

SHA1 has a harder job. Would not want to re-sha1 the file every time,
probably. So it'd need a local cache of file stat info, mapped to known
objects.

But: Even with %f, git actually passes the full file content to the clean
filter, and if it fails to consume it all, it will crash (may only happen
if the file is larger than some chunk size; tried with 500 mb file and 
saw a SIGPIPE.) This means unnecessary works needs to be done, 
and it slows down *everything*, from `git status` to `git commit`.
**showstopper** I have sent a patch to the git mailing list to address
this. <http://marc.info/?l=git&m=131465033512157&w=2> (Update: apparently
can't be fixed.)

> Update: I tried this again (2015) and it seems that git status and git
> add avoid re-sending the file content to the clean filter, as long as the
> file stat has not changed. I'm not sure when git started doing that,
> but it seems to avoid this problem.
> --[[Joey]]

### smudge

The smudge script can also be provided a filename with %f, but it
cannot directly write to the file or git gets unhappy.

> Still the case in 2015. Means an unnecesary read and pipe of the file
> even if the content is already locally available on disk. --[[Joey]]

### partial checkouts

.. Are very important, otherwise a repo can't scale past the size of the
smallest client's disk!

It would be nice if the smudge filter could hard link a work
tree file to the annex object.

But currently, the smudge filter can't modify the work tree file on its own
-- git always modifies the file after getting the output of the smudge
filter, and will stumble over any modifications that the smudge filter
makes. And, it's important that the smudge filter never fail as that will
leave the repo in a bad state.

Seems the best that can be done is for the smudge filter to copy from the
annex object when the object is present. When it's not present, the smudge
filter should provide a pointer to its content.

The clean filter should detect when it's operating on that pointer file.

I've a demo implementation of this technique in the scripts below.

### deduplication

.. Is nice; needing 2 copies of every annexed file is annoying.

Unfortunately, when using smudge/clean, `git merge` does not preserve a
smudged file in the work tree when renaming it. It instead deletes the old
file and asks the smudge filter to smudge the new filename.

So, copies need to be maintained in .git/annex/objects, though it's ok
to use hard links to the work tree files. (Although somewhat unsafe
since modification of the file will lose the old version. annex.thin
setting can enable this.)

Even if hard links are used, smudge needs to output the content of an
annexed file, which will result in duplication when merging in renames of
files.

### design

Goal: Get rid of current direct mode, using smudge/clean filters instead to
cover the same use cases, more flexibly and robustly.

Use case 1:

A user wants to be able to edit files, and git-add, git commit,
without needing to worry about using git-annex to unlock files, add files,
etc.

Use case 2:

Using git-annex on a crippled filesystem that does not support symlinks.

Data:

* An annex pointer file has as its first line the git-annex key
  that it's standing in for (prefixed with "annex/objects/", similar to
  an annex symlink target). Subsequent lines of the file might
  be a message saying that the file's content is not currently available.
  An annex pointer file is checked into the git repository the same way
  that an annex symlink is checked in.
* A file map is  maintained by git-annex, to keep track of the keys
  that are used by files in the working tree.

Configuration: 

* .gitattributes tells git which files to use git-annex's smudge/clean
  filters with. Typically, all files except for dotfiles:

	* filter=annex
	.* !filter

* annex.largefiles tells git-annex which files should in fact be put in 
  the annex. Other files are passed through the smudge/clean as-is and
  have their contents stored in git.

* annex.direct is repurposed to configure how git-annex adds files.
  When set to false, it adds symlinks and when true it adds pointer files.

git-annex clean:

* Run by `git add` (and diff and status, etc), and passed the
  filename, as well as fed the file content on stdin.

  Look at configuration to decide if this file's content belongs in the
  annex. If not, output the file content to stdout.

  Generate annex key from filename and content from stdin.

  Hard link (annex.thin) or copy .git/annex/objects to the file,
  if it doesn't already exist.

  This is done to prevent losing the only copy of a file when eg
  doing a git checkout of a different branch, or merging a commit that
  renames or deletes a file. But, with annex.thin no attempt is made to 
  protect the object from being modified. If a user wants to
  protect object contents from modification, they should use
  `git annex add`, not `git add`, or they can `git annex lock` after adding,
  or not enable annex.thin.

  Update file map.

  Output the pointer file content to stdout.

git-annex smudge:

* Run by eg `git checkout`
  and passed the filename, as well as fed the pointer file content on stdin.

  Update file map.

  When an object is present in the annex, outputs its content to stdout.
  Otherwise, outputs the file pointer content.

git annex direct/indirect:

  Previously these commands switched in and out of direct mode.
  Now they become no-ops.

git annex lock/unlock:

  Makes sense for these to change to switch files between using
  git-annex symlinks and pointers. So, this provides both a way to
  transition repositories to using pointers, and a cleaner unlock/lock
  for repos using symlinks.

  unlock will stage a pointer file, and will link the content of the object
  from .git/annex/objects to the work tree file.
  
  lock will replace the current work tree file with the symlink, and stage it,
  and lock down the permissions of the annex object.

#### file map

The file map needs to map from `Key -> [File]`. `File -> Key`
seems useful to have, but in practice is not worthwhile.

Drop and get operations need to know what files in the work tree use a
given key in order to update the work tree. And, we don't want to
overwrite a work tree file if it's been modified when dropping or getting.

git-annex commands that look at annex symlinks to get keys to act on will
need fall back to either consulting the file map, or looking at the staged
file to see if it's a pointer to a key. So a `File -> Key` map is a possible
optimisation.

Question: If the smudge/clean filters update the file map incrementally
based on the pointer files they generate/see, will the result
always be consistent with the content of the working tree?

This depends on when git calls the smudge/clean filters and on what.
In particular:

* Does the clean filter always get called when adding a relevant 
  file to git? Yes.
* Is the clean filter called at any other time? Yes, for example
  git diff will clean relevant modified files to generate the diff.
  So, the clean filter may see file versions that have not yet been staged
  in git.
* Is the clean filter ever passed content not in the work tree?
  I don't think so, but not 100% sure.
* Is the smudge filter always called when git updates a relevant file
  in the work tree? Yes.
* Is the smudge filter called at any other time? Seems unlikely but then
  there could be situations with a detached work tree or such.
* Does git call any useful hooks when removing a file from the work tree,
  or converting it to not be annexed, or for `git mv` of an annexed file?
  No!

From this analysis, any file map generated by the smudge/clean filters
is necessary potentially innaccurate. It may list deleted files.
It may or may not reflect current unstaged changes from the work tree.


Follows that any use of the file map needs to verify the info from it,
and throw out bad cached info (updating the map to match reality). 

When downloading a key, check if the files listed in the file map are
still pointer files in the work tree, and only replace them with the
content if so. 

When dropping a key, check if the files listed for it in the file map are
unmodified in the work tree, and are staged as pointers to the key,
and only reset them to the pointers if so. Note that this means that
a modified work tree file that has not yet been staged, but that
corresponds to a key, won't be reset when the key is dropped.
This is probably not a big deal; the user will either add the
file, which will add the key back, or reset the file.

Does the `File -> Key` map have any benefits given this innaccuracy?
Answer seems to be no; any answer that map gives may be innaccurate and
needs to be verified by looking at actual repo content, so might as well
just look at the repo content in the first place..

#### Upgrading

annex.version changes to 6

git config for filter.annex.smudge and filter.annex.clean is set up.

.gitattributes is updated with a stock configuration,
unless it already mentions "filter=annex".

Upgrading a direct mode repo needs to switch it out of bare mode, and
needs to run `git annex unlock` on all files (or reach the same result).
So will need to stage changes to all annexed files.

When a repo has some clones indirect and some direct, the upgraded repo
will have all files unlocked, necessarily in all clones. This happens
automatically, because when the direct repos are upgraded that causes the
files to be unlocked, while the indirect upgrades don't touch the files.

----

### test files

huge-smudge:

<pre>
#!/bin/sh
read f
file="$1"
echo "smudging $f" >&2
if [ -e ~/$f ]; then
	cat ~/$f # possibly expensive copy here
else
	echo "$f not available"
fi
</pre>

huge-clean:

<pre>
#!/bin/sh
file="$1"
cat >/tmp/file
# in real life, this should be done more efficiently, not trying to read
# the whole file content!
if grep -q 'not available' /tmp/file; then
	awk '{print $1}' /tmp/file # provide what we would if the content were avail!
	exit 0
fi
echo "cleaning $file" >&2
# XXX store file content here
echo $file
</pre>

.gitattributes:

<pre>
*.huge filter=huge
</pre>

in .git/config:

<pre>
[filter "huge"]
        clean = huge-clean %f
        smudge = huge-smudge %f
<pre>