Dealing with a race without using locking is exceedingly difficult and tricky.
Fully tested, I hope.
There are three places left where the branch can be updated, that are not
covered by the race recovery code. Let's prove they're all immune to the
race:
1. tryFastForwardTo checks to see if a fast-forward can be done,
and then does git-update-ref on the branch to fast-forward it.
If a push comes in before the check, then either no fast-forward
will be done (ok), or the push set the branch to a ref that can
still be fast-forwarded (also ok)
If a push comes in after the check, the git-update-ref will
undo the ref change made by the push. It's as if the push did not come
in, and the next git-push will see this, and try to re-do it.
(acceptable)
2. When creating the branch for the very first time, an empty index
is created, and a commit of it made to the branch. The commit's ref
is recorded as the current state of the index. If a push came in
during that, it will be noticed the next time a commit is made to the
branch, since the branch will have changed. (ok)
3. Creating the branch from an existing remote branch involves making
the branch, and then getting its ref, and recording that the index
reflects that ref.
If a push creates the branch first, git-branch will fail (ok).
If the branch is created and a racing push is then able to change it
(highly unlikely!) we're still ok, because it first records the ref into
the index.lck, and then updating the index. The race can cause the
index.lck to have the old branch ref, while the index has the newly pushed
branch merged into it, but that only results in an unnecessary update of
the index file later on.
The last branch ref that the index was updated to is stored in
.git/annex/index.lck, and the index only updated when the current
branch ref differs.
(The .lck file should later be used for locking too.)
Some more optimization is still needed, since there is some redundancy in
calls to git show-ref.
Always merge the git-annex branch into .git/annex/index before making a
commit from the index.
This ensures that, when the branch has been changed in any way
(by a push being received, or changes pulled directly into it, or
even by the user checking it out, and committing a change), the index
reflects those changes.
This is much too slow; it needs to be optimised to only update the
index when the branch has really changed, not every time.
Also, there is an unhandled race, when a change is made to the branch
right after the index gets updated. I left it in for now because it's
unlikely and I didn't want to complicate things with additional locking
yet.
The only fully supported thing is to have the main repository on one disk,
and .git/annex on another. Only commands that move data in/out of the annex
will need to copy it across devices.
There is only partial support for putting arbitrary subdirectories of
.git/annex on different devices. For one thing, but this can require more
copies to be done. For example, when .git/annex/tmp is on one device, and
.git/annex/journal on another, every journal write involves a call to
mv(1). Also, there are a few places that make hard links between various
subdirectories of .git/annex with createLink, that are not handled.
In the common case without cross-device, the new moveFile is actually
faster than renameFile, avoiding an unncessary stat to check that a file
(not a directory) is being moved. Of course if a cross-device move is
needed, it is as slow as mv(1) of the data.
The branch may not exist, if .git/annex has been copied over from another
repo (or a corrupted repo). I suppose it could also have gotten deleted
somehow. Without this, there is a confusing failure.
In git, a Ref can be a Sha, or a Branch, or a Tag. I added type aliases for
those. Note that this does not prevent mixing up of eg, refs and branches
at the type level. Since git really doesn't care, except rare cases like
git update-ref, or git tag -d, that seems ok for now.
There's also a tree-ish, but let's just use Ref for it. A given Sha or Ref
may or may not be a tree-ish, depending on the object type, so there seems
no point in trying to represent it at the type level.
Before, a merge was first calculated, by running various actions that
called git and built up a list of lines, which were at the end sent
to git update-index. This necessarily used space proportional to the size
of the diff between the trees being merged.
Now, lines are streamed into git update-index from each of the actions in
turn.
Runtime size of git-annex merge when merging 50000 location log files
drops from around 100 mb to a constant 4 mb.
Presumably it runs quite a lot faster, too.
Avoids doing auto-merging in commands that don't need fully current
information from the git-annex branch. In particular, git annex add no
longer needs to auto-merge. Affected commands: Anything that doesn't
look up data from the branch, but does write a change to it.
It might seem counterintuitive that we can change a value without first
making sure we have the current value. This optimisation works because
these two sequences are equivilant:
1. pull from remote
2. union merge
3. read file from branch
4. modify file and write to branch
vs.
1. read file from branch
2. modify file and write to branch
3. pull from remote
4. union merge
After either sequence, the git-annex branch contains the same logical content
for the modified file. (Possibly with lines in a different order or
additional old lines of course).
Many functions took the repo as their first parameter. Changing it
consistently to be the last parameter allows doing some useful things with
currying, that reduce boilerplate.
In particular, g <- gitRepo is almost never needed now, instead
use inRepo to run an IO action in the repo, and fromRepo to get
a value from the repo.
This also provides more opportunities to use monadic and applicative
combinators.
Thanks Valentin Haenel for a test case showing how non-fast-forward merges
could result in an ongoing pull/merge/push cycle.
While the git-annex branch is fast-forwarded, git-annex's index file is still
updated using the union merge strategy as before. There's no other way to
update the index that would be any faster.
It is possible that a union merge and a fast-forward result in different file
contents: Files should have the same lines, but a union merge may change
their order. If this happens, the next commit made to the git-annex branch
will have some unnecessary changes to line orders, but the consistency
of data should be preserved.
Note that when the journal contains changes, a fast-forward is never attempted,
which is fine, because committing those changes would be vanishingly unlikely
to leave the git-annex branch at a commit that already exists in one of
the remotes.
The real difficulty is handling the case where multiple remotes have all
changed. git-annex does find the best (ie, newest) one and fast forwards
to it. If the remotes are diverged, no fast-forward is done at all. It would
be possible to pick one, fast forward to it, and make a merge commit to
the rest, I see no benefit to adding that complexity.
Determining the best of N changed remotes requires N*2+1 calls to git-log, but
these are fast git-log calls, and N is typically small. Also, typically
some or all of the remote refs will be the same, and git-log is not called to
compare those. In the real world I expect this will almost always add only
1 git-log call to the merge process. (Which already makes N anyway.)
Specifically, disabled trying to update the git-annex branch on the remote,
since that data is never used by operations that act on such remotes.
Also, when copying content to such a remote, skip committing the presence
information changes to its git-annex branch. Leaving it in the journal there
is ok: Any command run on the remote that needs the info will flush the
journal.
This may partially solve this bug:
http://git-annex.branchable.com/bugs/fails_to_handle_lot_of_files/
Although I still see unreaped git processes piling up when doing a copy --to.
Another process may stage journalled files before the lock is
taken, so need to get the list of journalled files afterwards.
It's unfortunate this means getting the directory contents twice,
but it seems better to do that than sometimes take the lock
unnecessarily.
