Now really only done in the startup scan.
It turns out to be quite hard for event handlers to know when the startup
scan is complete. I tried to make addWatch pass that info, but found
threading the state very difficult. For now, a quick hack, using the fast
flag.
Note that it's actually possible for inotify events to come in while the
startup scan is still ongoing. Due to my hack, the expensive check will
be done for files added in such inotify events.
This requires a relatively expensive test at file add time to see if it's
in git already. But it can be optimised to only happen during the startup
scan.
I thought this might be a lock conflict that explains the deadlock when
built with -threaded, but it seems not.. it still locks! It even locks
without the committer thread.
Indeed, it locks when running "git annex add"! -threaded is exposing some
other problem.
Still, this seems conceptually cleaner and did not add any inneficiencies.
Also added some high-level documentation about the threads used.
The commit thread now has access to a channel containing the times of
all uncommitted changes. This lets it be smart about detecting busy times
when a batch job is running (such as rm -rf, or untarring something, etc),
and avoid committing until it's done. While at the same time, instantly
committing one-off changes that the user is going to expect to see
immediately.
I had to use STM to implement the channel, because of
http://hackage.haskell.org/trac/ghc/ticket/4154
While this adds a dependency, I always wanted to use STM, so this actually
makes me happy. ;)
Also happy that shouldCommit is a pure function, so other commit smartness
strategies can easily be played with. Although the current one seems pretty
good.
There is one bug, for some reason it does double commits, every time.
While I was in there, I noticed and fixed a bug in the queue size
calculations. It was never encountered only because Queue.add was
only ever run with 1 file in the list.
This ensures that all special remotes show up in git annex status.
Before, a special remote that was not manually described, and was not
a current git remote, did not show up there, although initremote did list
it.
There's a race adding a new file to the annex: The file is moved to the
annex and replaced with a symlink, and then we git add the symlink. If
someone comes along in the meantime and replaces the symlink with
something else, such as a new large file, we add that instead. Which could
be bad..
This race is fixed by avoiding using git add, instead the symlink is
directly staged into the index.
It would be nice to make `git annex add` use this same technique.
I have not done so yet because it currently runs git update-index once per
file, which would slow does `git annex add`. A future enhancement would be
to extend the Git.Queue to include the ability to run update-index with
a list of Streamers.
Anything that tries to open the file for write, or delete the file,
or replace it with something else, will not affect the add.
Only if a process has the file open for write before add starts
can it still change it while (or after) it's added to the annex.
(fsck will catch this later of course)
Uses a MVar again, as there seems no other way to thread the state through
inotify events.
This is a rather unsatisfactory result. I had wanted to run them in
the same monad so that the git queue could be used to coleasce git commands
and speed things up. But, that led to fragility: If several files are
added, and one is removed before queue flush, git add will fail to add
any of them. So, the queue is still explicitly flushed after each add for
now.
TODO: Investigate using git add --ignore-errors. This would need to be done
in Command.Add. And, git add still exits nonzero with it, so would need
to avoid crashing on queue flush.
When a new file is annexed, a deletion event occurs when it's moved away
to be replaced by a symlink. Most of the time, there is no problimatic
race, because the same thread runs the add event as the deletion event.
So, once the symlink is in place, the deletion code won't run at all,
due to existing checks that a deleted file is really gone.
But there is a race at startup, as then the inotify thread is running
at the same time as the main thread, which does the initial tree walking
and annexing. It would be possible for the deletion inotify to run
in a perfect race with the addition, and remove the newly added symlink
from the git cache.
To solve this race, added event serialization via a MVar. We putMVar
before running each event, which blocks if an event is already running.
And when an event finishes (or crashes!), we takeMVar to free the lock.
Also, make rm -rf not spew warnings by passing --ignore-unmatch when
deleting directories.
And just like that, annexed files can be moved and copies around within
the tree, and are automatically fixed to point to the content, and staged
in git. Huzzah!
Delete still remains TODO, with its troublesome race during add..
Improved the inotify code, so it will also notice directory removal
and symlink creation.
In the watch code, optimised away a stat of a file that's being added,
that's done by Command.Add.start. This is the reason symlink creation is
handled separately from file creation, since during initial tree walk
at startup, a stat was already done, and can be reused.
Resetting an unlocked file to the branch head failed if it had just been
added, not committed, and unlocked, since the branch didbn't have it.
The code was concerned about dropping any changes that might be staged in the
index, but I cannot see why.
Baked into the code was an assumption that a repository's git directory
could be determined by adding ".git" to its work tree (or nothing for bare
repos). That fails when core.worktree, or GIT_DIR and GIT_WORK_TREE are
used to separate the two.
This was attacked at the type level, by storing the gitdir and worktree
separately, so Nothing for the worktree means a bare repo.
A complication arose because we don't learn where a repository is bare
until its configuration is read. So another Location type handles
repositories that have not had their config read yet. I am not entirely
happy with this being a Location type, rather than representing them
entirely separate from the Git type. The new code is not worse than the
old, but better types could enforce more safety.
Added support for core.worktree. Overriding it with -c isn't supported
because it's not really clear what to do if a git repo's config is read, is
not bare, and is then overridden to bare. What is the right git directory
in this case? I will worry about this if/when someone has a use case for
overriding core.worktree with -c. (See Git.Config.updateLocation)
Also removed and renamed some functions like gitDir and workTree that
misused git's terminology.
One minor regression is known: git annex add in a bare repository does not
print a nice error message, but runs git ls-files in a way that fails
earlier with a less nice error message. This is because before --work-tree
was always passed to git commands, even in a bare repo, while now it's not.
annex.ssh-options, annex.rsync-options, annex.bup-split-options.
And adjust types to avoid the bugs that broke several config settings
recently. Now "annex." prefixing is enforced at the type level.
This is incomplete, it does not honor it yet for hash directories
and other annex bookkeeping files. Some of that is not needed for a bare
repo; some of it may be.
So far this only handles auto-annexing new files that are created inside
the repository while it's running. To make this really useful,
it needs to at least:
- notice deleted files and stage the deletion
(tricky; there's a race with add..)
- notice renamed files, auto-fix the symlink, and stage the new file location
- periodically auto-commit staged changes
- honor .gitignore, not adding files it excludes
Also nice to have would be:
- Somehow sync remotes, possibly using a push sync like dvcs-autosync
does, so they are immediately updated.
- Somehow get content that is unavilable. This is problimatic with inotify,
since we only get an event once the user has tried (and failed) to read
from the file. Perhaps instead, automatically copy content that is added
out to remotes, with the goal of all repos eventually getting a copy,
if df allows.
- Drop files that have not been used lately, or meet some other criteria
(as long as there's a copy elsewhere).
- Perhaps automatically dropunused files that have been deleted,
although I cannot see a way to do that, since by the time the inotify
deletion event arrives, the file is deleted, and we cannot see what
its symlink pointed to! Alternatievely, perhaps automatically
do an expensive unused/dropunused cleanup process.
Some of this probably needs the currently stateless threads to maintain
a common state.
getConfig got a remote-specific config, and this confusing name caused it
to be used a couple of places that only were interested in global configs.
Rename to getRemoteConfig and make getConfig only get global configs.
There are no behavior changes here, but remote.<name>.annex-web-options
never actually worked (and per-remote web options is a very unlikely to be
useful case so I didn't make it work), so fix the documentation for it.
Add tuning, docs, etc.
Not sure if status is the right place to remote size.. perhaps unused
should report the size and also warn if it sees more keys than the bloom
filter allows?
Can't trust the key size to be accurate for tmp and bad keys, so check
actual file size. In the wild I saw the old code be wrong by a factor
of about 100!
If all tmp/bad keys are empty, they're not shown in status at all.
Showing 0 bytes and suggesting to clean it up seemed weird..
Stale and bad files are rare, so it's more efficient to use inAnnex to see
if they can be deleted, rather than keeping the list of all present keys
around for them.
Before, it leaked space due to caching lists of keys. Now all necessary
data about keys is calculated as they stream in.
The "nearly constant" is due to getKeysPresent, which builds up a lot
of [] thunks as it traverses .git/annex/objects/. Will deal with it later.
Much of the memory bloat turned out to be due to getKeysReferenced
containing a mapM, which is strict and buffered the whole list
rather than streaming it.
The other half of the bloat was due to building a temporary Set
in order to call S.difference. While that is more cpu efficient,
I switched to successive S.delete, since with it, I can run a whole
git annex unused in less than 8 mb of memory.
The whole Set of keys with content available is still stored in memory,
so running unused in a repo with a whole lot of file content will still
use more memory. In a repo containing 6000 files, it needed 40 mb.
Note that the status command still uses the bloatful getKeysReferenced.
This has two benefits.
1. When a lot of refs are going to be received, get them via lower cost
connection when possible.
2. Allows ctrl-c of sync after the cheaper remotes have been pulled from
(or pushed to).
Rather than go through the location log to see which files are present on
the remote, it simply looks at the disk contents directly.
I benchmarked this speeding up scanning 834 files, from an annex on my
phone's SSD, from 11.39 seconds to 1.31 seconds. (No files actually moved.)
Also benchmarked 8139 files, from an annex on spinning storage,
speeding up from 103.17 to 13.39 seconds.
Note that benchmarking with an encrypted annex on flash actually showed a
minor slowdown with this optimisation -- from 13.93 to 14.50 seconds. Seems
the overhead of doing the crypto needed to get the filenames to directly
check can be higher than the overhead of looking up data in the location
log. (Which says good things about how well the location log and git have
been optimised!) It *may* make sense to make encrypted local remotes not
have hasKeyCheap set; further benchmarking is called for.
Eventually, git-annex might try running this after making changes to
a remote. I have not yet thought of a good way for it to tell which
remotes it needs to run it on though. It can't just do it when
shutting down a cached ssh connection, because ssh connection caching
is optional, and that would not handle local remotes not accessed over ssh
either.
Now changes are staged into the branch's index, but not committed,
which avoids growing a large journal. And sync and merge always
explicitly commit, ensuring that even when they do nothing else,
they commit the staged changes.
Added a flag file to indicate that the branch's journal contains
uncommitted changes. (Could use git ls-files, but don't want to run
that every time.)
In the future, this ability to have uncommitted changes staged in the
journal might be used on remotes after a series of oneshot commands.
Now gitattributes are looked up, efficiently, in only the places that
really need them, using the same approach used for cat-file.
The old CheckAttr code seemed very fragile, in the way it streamed files
through git check-attr.
I actually found that cad8824852
was still deadlocking with ghc 7.4, at the end of adding a lot of files.
This should fix that problem, and avoid future ones.
The best part is that this removes withAttrFilesInGit and withNumCopies,
which were complicated Seek methods, as well as simplfying the types
for several other Seek methods that had a Backend tupled in.
Can be used to specify what file the url is added to. This can be used to
override the default filename that is used when adding an url, which is
based on the url. Or, when the file already exists, the url is recorded as
another location of the file.
Under ghc 7.4, this seems to be able to handle all filename encodings
again. Including filename encodings that do not match the LANG setting.
I think this will not work with earlier versions of ghc, it uses some ghc
internals.
Turns out that ghc 7.4 has a special filesystem encoding that it uses when
reading/writing filenames (as FilePaths). This encoding is documented
to allow "arbitrary undecodable bytes to be round-tripped through it".
So, to get FilePaths from eg, git ls-files, set the Handle that is reading
from git to use this encoding. Then things basically just work.
However, I have not found a way to make Text read using this encoding.
Text really does assume unicode. So I had to switch back to using String
when reading/writing data to git. Which is a pity, because it's some
percent slower, but at least it works.
Note that stdout and stderr also have to be set to this encoding, or
printing out filenames that contain undecodable bytes causes a crash.
IMHO this is a misfeature in ghc, that the user can pass you a filename,
which you can readFile, etc, but that default, putStr of filename may
cause a crash!
Git.CheckAttr gave me special trouble, because the filenames I got back
from git, after feeding them in, had further encoding breakage.
Rather than try to deal with that, I just zip up the input filenames
with the attributes. Which must be returned in the same order queried
for this to work.
Also of note is an apparent GHC bug I worked around in Git.CheckAttr. It
used to forkProcess and feed git from the child process. Unfortunatly,
after this forkProcess, accessing the `files` variable from the parent
returns []. Not the value that was passed into the function. This screams
of a bad bug, that's clobbering a variable, but for now I just avoid
forkProcess there to work around it. That forkProcess was itself only added
because of a ghc bug, #624389. I've confirmed that the test case for that
bug doesn't reproduce it with ghc 7.4. So that's ok, except for the new ghc
bug I have not isolated and reported. Why does this simple bit of code
magnet the ghc bugs? :)
Also, the symlink touching code is currently broken, when used on utf-8
filenames in a non-utf-8 locale, or probably on any filename containing
undecodable bytes, and I temporarily commented it out.
I had not realized what a memory leak the lazy state monad could be,
although I have not seen much evidence of actual leaking in git-annex.
However, if running git-annex on a great many files, this could matter.
The additional Utility.State.changeState adds even more strictness,
avoiding a problem I saw in github-backup where repeatedly modifying
state built up a huge pile of thunks.
Done by adding a oneshot mode, in which location log changes are written to
the journal, but not committed. Taking advantage of git-annex's existing
ability to recover in this situation.
This is used by git-annex-shell and other places where changes are made to
a remote's location log.
Fscking a remote is now supported. It's done by retrieving
the contents of the specified files from the remote, and checking them,
so can be an expensive operation.
(Several optimisations are possible, to speed it up, of course.. This is
the slow and stupid remote fsck to start with.)
Still, if the remote is a special remote, or a git repository that you
cannot run fsck in locally, it's nice to have the ability to fsck it.
If you have any directory special remotes, now would be a good time to
fsck them, in case you were hit by the data loss bug fixed in the
previous release!
