The latest feature release GIT 1.6.1 is available at the usual
places:
http://www.kernel.org/pub/software/scm/git/
git-1.6.1.tar.{gz,bz2} (source tarball)
git-htmldocs-1.6.1.tar.{gz,bz2} (preformatted docs)
git-manpages-1.6.1.tar.{gz,bz2} (preformatted docs)
The RPM binary packages for a few architectures are also found in the
vicinity.
RPMS/$arch/git-*-1.6.1-1.fc9.$arch.rpm (RPM)
----------------------------------------------------------------
GIT v1.6.1 Release Notes
========================
Updates since v1.6.0
--------------------
When some commands (e.g. "git log", "git diff") spawn pager internally, we
used to make the pager the parent process of the git command that produces
output. This meant that the exit status of the whole thing comes from the
pager, not the underlying git command. We swapped the order of the
processes around and you will see the exit code from the command from now
on.
(subsystems)
* gitk can call out to git-gui to view "git blame" output; git-gui in turn
can run gitk from its blame view.
* Various git-gui updates including updated translations.
* Various gitweb updates from repo.or.cz installation.
* Updates to emacs bindings.
(portability)
* A few test scripts used nonportable "grep" that did not work well on
some platforms, e.g. Solaris.
* Sample pre-auto-gc script has OS X support.
* Makefile has support for (ancient) FreeBSD 4.9.
(performance)
* Many operations that are lstat(3) heavy can be told to pre-execute
necessary lstat(3) in parallel before their main operations, which
potentially gives much improved performance for cold-cache cases or in
environments with weak metadata caching (e.g. NFS).
* The underlying diff machinery to produce textual output has been
optimized, which would result in faster "git blame" processing.
* Most of the test scripts (but not the ones that try to run servers)
can be run in parallel.
* Bash completion of refnames in a repository with massive number of
refs has been optimized.
* Cygwin port uses native stat/lstat implementations when applicable,
which leads to improved performance.
* "git push" pays attention to alternate repositories to avoid sending
unnecessary objects.
* "git svn" can rebuild an out-of-date rev_map file.
(usability, bells and whistles)
* When you mistype a command name, git helpfully suggests what it guesses
you might have meant to say. help.autocorrect configuration can be set
to a non-zero value to accept the suggestion when git can uniquely
guess.
* The packfile machinery hopefully is more robust when dealing with
corrupt packs if redundant objects involved in the corruption are
available elsewhere.
* "git add -N path..." adds the named paths as an empty blob, so that
subsequent "git diff" will show a diff as if they are creation events.
* "git add" gained a built-in synonym for people who want to say "stage
changes" instead of "add contents to the staging area" which amounts
to the same thing.
* "git apply" learned --include=paths option, similar to the existing
--exclude=paths option.
* "git bisect" is careful about a user mistake and suggests testing of
merge base first when good is not a strict ancestor of bad.
* "git bisect skip" can take a range of commits.
* "git blame" re-encodes the commit metainfo to UTF-8 from i18n.commitEncoding
by default.
* "git check-attr --stdin" can check attributes for multiple paths.
* "git checkout --track origin/hack" used to be a syntax error. It now
DWIMs to create a corresponding local branch "hack", i.e. acts as if you
said "git checkout --track -b hack origin/hack".
* "git checkout --ours/--theirs" can be used to check out one side of a
conflicting merge during conflict resolution.
* "git checkout -m" can be used to recreate the initial conflicted state
during conflict resolution.
* "git cherry-pick" can also utilize rerere for conflict resolution.
* "git clone" learned to be verbose with -v
* "git commit --author=$name" can look up author name from existing
commits.
* output from "git commit" has been reworded in a more concise and yet
more informative way.
* "git count-objects" reports the on-disk footprint for packfiles and
their corresponding idx files.
* "git daemon" learned --max-connections=<count> option.
* "git daemon" exports REMOTE_ADDR to record client address, so that
spawned programs can act differently on it.
* "git describe --tags" favours closer lightweight tags than farther
annotated tags now.
* "git diff" learned to mimic --suppress-blank-empty from GNU diff via a
configuration option.
* "git diff" learned to put more sensible hunk headers for Python,
HTML and ObjC contents.
* "git diff" learned to vary the a/ vs b/ prefix depending on what are
being compared, controlled by diff.mnemonicprefix configuration.
* "git diff" learned --dirstat-by-file to count changed files, not number
of lines, when summarizing the global picture.
* "git diff" learned "textconv" filters --- a binary or hard-to-read
contents can be munged into human readable form and the difference
between the results of the conversion can be viewed (obviously this
cannot produce a patch that can be applied, so this is disabled in
format-patch among other things).
* "--cached" option to "git diff has an easier to remember synonym "--staged",
to ask "what is the difference between the given commit and the
contents staged in the index?"
* "git for-each-ref" learned "refname:short" token that gives an
unambiguously abbreviated refname.
* Auto-numbering of the subject lines is the default for "git
format-patch" now.
* "git grep" learned to accept -z similar to GNU grep.
* "git help" learned to use GIT_MAN_VIEWER environment variable before
using "man" program.
* "git imap-send" can optionally talk SSL.
* "git index-pack" is more careful against disk corruption while
completing a thin pack.
* "git log --check" and "git log --exit-code" passes their underlying diff
status with their exit status code.
* "git log" learned --simplify-merges, a milder variant of --full-history;
"gitk --simplify-merges" is easier to view than with --full-history.
* "git log" learned "--source" to show what ref each commit was reached
from.
* "git log" also learned "--simplify-by-decoration" to show the
birds-eye-view of the topology of the history.
* "git log --pretty=format:" learned "%d" format element that inserts
names of tags that point at the commit.
* "git merge --squash" and "git merge --no-ff" into an unborn branch are
noticed as user errors.
* "git merge -s $strategy" can use a custom built strategy if you have a
command "git-merge-$strategy" on your $PATH.
* "git pull" (and "git fetch") can be told to operate "-v"erbosely or
"-q"uietly.
* "git push" can be told to reject deletion of refs with receive.denyDeletes
configuration.
* "git rebase" honours pre-rebase hook; use --no-verify to bypass it.
* "git rebase -p" uses interactive rebase machinery now to preserve the merges.
* "git reflog expire branch" can be used in place of "git reflog expire
refs/heads/branch".
* "git remote show $remote" lists remote branches one-per-line now.
* "git send-email" can be given revision range instead of files and
maildirs on the command line, and automatically runs format-patch to
generate patches for the given revision range.
* "git submodule foreach" subcommand allows you to iterate over checked
out submodules.
* "git submodule sync" subcommands allows you to update the origin URL
recorded in submodule directories from the toplevel .gitmodules file.
* "git svn branch" can create new branches on the other end.
* "gitweb" can use more saner PATH_INFO based URL.
(internal)
* "git hash-object" learned to lie about the path being hashed, so that
correct gitattributes processing can be done while hashing contents
stored in a temporary file.
* various callers of git-merge-recursive avoid forking it as an external
process.
* Git class defined in "Git.pm" can be subclasses a bit more easily.
* We used to link GNU regex library as a compatibility layer for some
platforms, but it turns out it is not necessary on most of them.
* Some path handling routines used fixed number of buffers used alternately
but depending on the call depth, this arrangement led to hard to track
bugs. This issue is being addressed.
Fixes since v1.6.0
------------------
All of the fixes in v1.6.0.X maintenance series are included in this
release, unless otherwise noted.
* Porcelains implemented as shell scripts were utterly confused when you
entered to a subdirectory of a work tree from sideways, following a
symbolic link (this may need to be backported to older releases later).
* Tracking symbolic links would work better on filesystems whose lstat()
returns incorrect st_size value for them.
* "git add" and "git update-index" incorrectly allowed adding S/F when S
is a tracked symlink that points at a directory D that has a path F in
it (we still need to fix a similar nonsense when S is a submodule and F
is a path in it).
* "git am" after stopping at a broken patch lost --whitespace, -C, -p and
--3way options given from the command line initially.
* "git diff --stdin" used to take two trees on a line and compared them,
but we dropped support for such a use case long time ago. This has
been resurrected.
* "git filter-branch" failed to rewrite a tag name with slashes in it.
* "git http-push" did not understand URI scheme other than opaquelocktoken
when acquiring a lock from the server (this may need to be backported to
older releases later).
* After "git rebase -p" stopped with conflicts while replaying a merge,
"git rebase --continue" did not work (may need to be backported to older
releases).
* "git revert" records relative to which parent a revert was made when
reverting a merge. Together with new documentation that explains issues
around reverting a merge and merging from the updated branch later, this
hopefully will reduce user confusion (this may need to be backported to
older releases later).
* "git rm --cached" used to allow an empty blob that was added earlier to
be removed without --force, even when the file in the work tree has
since been modified.
* "git push --tags --all $there" failed with generic usage message without
telling saying these two options are incompatible.
* "git log --author/--committer" match used to potentially match the
timestamp part, exposing internal implementation detail. Also these did
not work with --fixed-strings match at all.
* "gitweb" did not mark non-ASCII characters imported from external HTML fragments
correctly.
On Wed, Dec 24, 2008 at 10:36:27PM -0800, Junio C Hamano wrote:
> * The packfile machinery hopefully is more robust when dealing with
> corrupt packs if redundant objects involved in the corruption are
> available elsewhere.
Has anyone written a summary of how Git's redundancy operates?
* What would be the probability for a single bit flip to corrupt the
repository?
* And what is the situation where a single bit flip can not corrupt the
database?
* When (which commands/functions) is error detection done?
Heikki
["Followup-To:" header set to gmane.comp.version-control.git.]
On 2008-12-25, Heikki Orsila <[email protected]> wrote:
> * What would be the probability for a single bit flip to corrupt the
> repository?
> * And what is the situation where a single bit flip can not corrupt the
> database?
These two questions are not specific to git; whatever
answers apply to sha1 also apply here. Sha1 will always be
strong enough to detect any combination of random errors.
As for deliberate attacks, it is considered strong enough to
resist all but the most computationally intensive attacks
(the kind that require a worldwide effort or maybe NSA type
facilities).
> * When (which commands/functions) is error detection done?
"git fsck --full" will do it of course, but I'm not sure
what other operations also start off or end up doing a
check.
On Thu, 25 Dec 2008, Heikki Orsila wrote:
> On Wed, Dec 24, 2008 at 10:36:27PM -0800, Junio C Hamano wrote:
> > * The packfile machinery hopefully is more robust when dealing with
> > corrupt packs if redundant objects involved in the corruption are
> > available elsewhere.
>
> Has anyone written a summary of how Git's redundancy operates?
>
> * What would be the probability for a single bit flip to corrupt the
> repository?
This is not something that git itself could answer. The probability
depends on the quality of your hardware. Once that probability has
occurred though, it is clear that your repository is then corrupted as
there is hardly any redundant bits in a git repository.
> * And what is the situation where a single bit flip can not corrupt the
> database?
The database can be resilient against most kind of corruptions if you
have a redundant copy of the affected object. It will still be
corrupted, but git is now able to detect corruptions gracefully and
function correctly with some fallback objects. Those objects must exist
in some other related repository though, and copied over to the affected
repository manually. It's then possible and recommended to "fix" the
corruption simply by repacking the repository at that point.
So there is no magic involved: you need to have some kind of backups,
either using traditional backup solutions, or by simply having your
repository cloned somewhere else. The idea of having a repository fixed
with redundant objects is for those cases where you need to salvage new
work that has no corresponding backup, but although corresponding
objects are not corrupted, they could be delta objects which base is
against old objects which happen to be corrupted.
> * When (which commands/functions) is error detection done?
Error detection is performed all the time. When it's not the more
expensive SHA1 checksum, at least the zlib CRC32 is verified. What the
latest git version does amongst other corruption related things is to
close some small holes where some specific kind of corruptions could
have been undetected and propagated from one pack to another when
repacking.
Nicolas