1.9.0.txt   [plain text]

Git v1.9.0 Release Notes

Backward compatibility notes

"git submodule foreach $cmd $args" used to treat "$cmd $args" the same
way "ssh" did, concatenating them into a single string and letting the
shell unquote. Careless users who forget to sufficiently quote $args
get their argument split at $IFS whitespaces by the shell, and got
unexpected results due to this. Starting from this release, the
command line is passed directly to the shell, if it has an argument.

Read-only support for experimental loose-object format, in which users
could optionally choose to write their loose objects for a short
while between v1.4.3 and v1.5.3 era, has been dropped.

The meanings of the "--tags" option to "git fetch" has changed; the
command fetches tags _in addition to_ what is fetched by the same
command line without the option.

The way "git push $there $what" interprets the $what part given on the
command line, when it does not have a colon that explicitly tells us
what ref at the $there repository is to be updated, has been enhanced.

A handful of ancient commands that have long been deprecated are
finally gone (repo-config, tar-tree, lost-found, and peek-remote).

Backward compatibility notes (for Git 2.0.0)

When "git push [$there]" does not say what to push, we have used the
traditional "matching" semantics so far (all your branches were sent
to the remote as long as there already are branches of the same name
over there).  In Git 2.0, the default will change to the "simple"
semantics, which pushes:

 - only the current branch to the branch with the same name, and only
   when the current branch is set to integrate with that remote
   branch, if you are pushing to the same remote as you fetch from; or

 - only the current branch to the branch with the same name, if you
   are pushing to a remote that is not where you usually fetch from.

Use the user preference configuration variable "push.default" to
change this.  If you are an old-timer who is used to the "matching"
semantics, you can set the variable to "matching" to keep the
traditional behaviour.  If you want to live in the future early, you
can set it to "simple" today without waiting for Git 2.0.

When "git add -u" (and "git add -A") is run inside a subdirectory and
does not specify which paths to add on the command line, it
will operate on the entire tree in Git 2.0 for consistency
with "git commit -a" and other commands.  There will be no
mechanism to make plain "git add -u" behave like "git add -u .".
Current users of "git add -u" (without a pathspec) should start
training their fingers to explicitly say "git add -u ."
before Git 2.0 comes.  A warning is issued when these commands are
run without a pathspec and when you have local changes outside the
current directory, because the behaviour in Git 2.0 will be different
from today's version in such a situation.

In Git 2.0, "git add <path>" will behave as "git add -A <path>", so
that "git add dir/" will notice paths you removed from the directory
and record the removal.  Versions before Git 2.0, including this
release, will keep ignoring removals, but the users who rely on this
behaviour are encouraged to start using "git add --ignore-removal <path>"
now before 2.0 is released.

The default prefix for "git svn" will change in Git 2.0.  For a long
time, "git svn" created its remote-tracking branches directly under
refs/remotes, but it will place them under refs/remotes/origin/ unless
it is told otherwise with its --prefix option.

Updates since v1.8.5

Foreign interfaces, subsystems and ports.

 * The HTTP transport, when talking GSS-Negotiate, uses "100
   Continue" response to avoid having to rewind and resend a large
   payload, which may not be always doable.

 * Various bugfixes to remote-bzr and remote-hg (in contrib/).

 * The build procedure is aware of MirBSD now.

 * Various "git p4", "git svn" and "gitk" updates.

UI, Workflows & Features

 * Fetching from a shallowly-cloned repository used to be forbidden,
   primarily because the codepaths involved were not carefully vetted
   and we did not bother supporting such usage. This release attempts
   to allow object transfer out of a shallowly-cloned repository in a
   more controlled way (i.e. the receiver becomes a shallow repository
   with a truncated history).

 * Just like we give a reasonable default for "less" via the LESS
   environment variable, we now specify a reasonable default for "lv"
   via the "LV" environment variable when spawning the pager.

 * Two-level configuration variable names in "branch.*" and "remote.*"
   hierarchies, whose variables are predominantly three-level, were
   not completed by hitting a <TAB> in bash and zsh completions.

 * Fetching a 'frotz' branch with "git fetch", while a 'frotz/nitfol'
   remote-tracking branch from an earlier fetch was still there, would
   error out, primarily because the command was not told that it is
   allowed to lose any information on our side.  "git fetch --prune"
   now can be used to remove 'frotz/nitfol' to make room for fetching and
   storing the 'frotz' remote-tracking branch.

 * "diff.orderfile=<file>" configuration variable can be used to
   pretend as if the "-O<file>" option were given from the command
   line of "git diff", etc.

 * The negative pathspec syntax allows "git log -- . ':!dir'" to tell
   us "I am interested in everything but 'dir' directory".

 * "git difftool" shows how many different paths there are in total,
   and how many of them have been shown so far, to indicate progress.

 * "git push origin master" used to push our 'master' branch to update
   the 'master' branch at the 'origin' repository.  This has been
   enhanced to use the same ref mapping "git push origin" would use to
   determine what ref at the 'origin' to be updated with our 'master'.
   For example, with this configuration

   [remote "origin"]
      push = refs/heads/*:refs/review/*

   that would cause "git push origin" to push out our local branches
   to corresponding refs under refs/review/ hierarchy at 'origin',
   "git push origin master" would update 'refs/review/master' over
   there.  Alternatively, if push.default is set to 'upstream' and our
   'master' is set to integrate with 'topic' from the 'origin' branch,
   running "git push origin" while on our 'master' would update their
   'topic' branch, and running "git push origin master" while on any
   of our branches does the same.

 * "gitweb" learned to treat ref hierarchies other than refs/heads as
   if they are additional branch namespaces (e.g. refs/changes/ in

 * "git for-each-ref --format=..." learned a few formatting directives;
   e.g. "%(color:red)%(HEAD)%(color:reset) %(refname:short) %(subject)".

 * The command string given to "git submodule foreach" is passed
   directly to the shell, without being eval'ed.  This is a backward
   incompatible change that may break existing users.

 * "git log" and friends learned the "--exclude=<glob>" option, to
   allow people to say "list history of all branches except those that
   match this pattern" with "git log --exclude='*/*' --branches".

 * "git rev-parse --parseopt" learned a new "--stuck-long" option to
   help scripts parse options with an optional parameter.

 * The "--tags" option to "git fetch" no longer tells the command to
   fetch _only_ the tags. It instead fetches tags _in addition to_
   what are fetched by the same command line without the option.

Performance, Internal Implementation, etc.

 * When parsing a 40-hex string into the object name, the string is
   checked to see if it can be interpreted as a ref so that a warning
   can be given for ambiguity. The code kicked in even when the
   core.warnambiguousrefs is set to false to squelch this warning, in
   which case the cycles spent to look at the ref namespace were an
   expensive no-op, as the result was discarded without being used.

 * The naming convention of the packfiles has been updated; it used to
   be based on the enumeration of names of the objects that are
   contained in the pack, but now it also depends on how the packed
   result is represented---packing the same set of objects using
   different settings (or delta order) would produce a pack with
   different name.

 * "git diff --no-index" mode used to unnecessarily attempt to read
   the index when there is one.

 * The deprecated parse-options macro OPT_BOOLEAN has been removed;
   use OPT_BOOL or OPT_COUNTUP in new code.

 * A few duplicate implementations of prefix/suffix string comparison
   functions have been unified to starts_with() and ends_with().

 * The new PERLLIB_EXTRA makefile variable can be used to specify
   additional directories Perl modules (e.g. the ones necessary to run
   git-svn) are installed on the platform when building.

 * "git merge-base" learned the "--fork-point" mode, that implements
   the same logic used in "git pull --rebase" to find a suitable fork
   point out of the reflog entries for the remote-tracking branch the
   work has been based on.  "git rebase" has the same logic that can be
   triggered with the "--fork-point" option.

 * A third-party "receive-pack" (the responder to "git push") can
   advertise the "no-thin" capability to tell "git push" not to use
   the thin-pack optimization. Our receive-pack has always been
   capable of accepting and fattening a thin-pack, and will continue
   not to ask "git push" to use a non-thin pack.

Also contains various documentation updates and code clean-ups.

Fixes since v1.8.5

Unless otherwise noted, all the fixes since v1.8.5 in the maintenance
track are contained in this release (see the maintenance releases' notes
for details).

 * The pathspec matching code, while comparing two trees (e.g. "git
   diff A B -- path1 path2") was too aggressive and failed to match
   some paths when multiple pathspecs were involved.

 * "git repack --max-pack-size=8g" stopped being parsed correctly when
   the command was reimplemented in C.

 * An earlier update in v1.8.4.x to "git rev-list --objects" with
   negative ref had a performance regression.
   (merge 200abe7 jk/mark-edges-uninteresting later to maint).

 * A recent update to "git send-email" broke platforms where
   /etc/ssl/certs/ directory exists but cannot be used as SSL_ca_path
   (e.g. Fedora rawhide).

 * A handful of bugs around interpreting $branch@{upstream} notation
   and its lookalike, when $branch part has interesting characters,
   e.g. "@", and ":", have been fixed.

 * "git clone" would fail to clone from a repository that has a ref
   directly under "refs/", e.g. "refs/stash", because different
   validation paths do different things on such a refname.  Loosen the
   client side's validation to allow such a ref.

 * "git log --left-right A...B" lost the "leftness" of commits
   reachable from A when A is a tag as a side effect of a recent
   bugfix.  This is a regression in 1.8.4.x series.

 * documentations to "git pull" hinted there is an "-m" option because
   it incorrectly shared the documentation with "git merge".

 * "git diff A B submod" and "git diff A B submod/" ought to have done
   the same for a submodule "submod", but didn't.

 * "git clone $origin foo\bar\baz" on Windows failed to create the
   leading directories (i.e. a moral-equivalent of "mkdir -p").

 * "submodule.*.update=checkout", when propagated from .gitmodules to
   .git/config, turned into a "submodule.*.update=none", which did not
   make much sense.
   (merge efa8fd7 fp/submodule-checkout-mode later to maint).

 * The implementation of 'git stash $cmd "stash@{...}"' did not quote
   the stash argument properly and left it split at IFS whitespace.

 * The "--[no-]informative-errors" options to "git daemon" were parsed
   a bit too loosely, allowing any other string after these option

 * There is no reason to have a hardcoded upper limit for the number of
   parents of an octopus merge, created via the graft mechanism, but
   there was.

 * The basic test used to leave unnecessary trash directories in the
   t/ directory.
   (merge 738a8be jk/test-framework-updates later to maint).

 * "git merge-base --octopus" used to leave cleaning up suboptimal
   result to the caller, but now it does the clean-up itself.

 * A "gc" process running as a different user should be able to stop a
   new "gc" process from starting, but it didn't.

 * An earlier "clean-up" introduced an unnecessary memory leak.

 * "git add -A" (no other arguments) in a totally empty working tree
   used to emit an error.

 * "git log --decorate" did not handle a tag pointed by another tag

 * When we figure out how many file descriptors to allocate for
   keeping packfiles open, a system with non-working getrlimit() could
   cause us to die(), but because we make this call only to get a
   rough estimate of how many are available and we do not even attempt
   to use up all available file descriptors ourselves, it is nicer to
   fall back to a reasonable low value rather than dying.

 * read_sha1_file(), that is the workhorse to read the contents given
   an object name, honoured object replacements, but there was no
   corresponding mechanism to sha1_object_info() that was used to
   obtain the metainfo (e.g. type & size) about the object.  This led
   callers to weird inconsistencies.
   (merge 663a856 cc/replace-object-info later to maint).

 * "git cat-file --batch=", an admittedly useless command, did not
   behave very well.

 * "git rev-parse <revs> -- <paths>" did not implement the usual
   disambiguation rules the commands in the "git log" family used in
   the same way.

 * "git mv A B/", when B does not exist as a directory, should error
   out, but it didn't.

 * A workaround to an old bug in glibc prior to glibc 2.17 has been
   retired; this would remove a side effect of the workaround that
   corrupts system error messages in non-C locales.

 * SSL-related options were not passed correctly to underlying socket
   layer in "git send-email".

 * "git commit -v" appends the patch to the log message before
   editing, and then removes the patch when the editor returned
   control. However, the patch was not stripped correctly when the
   first modified path was a submodule.

 * "git fetch --depth=0" was a no-op, and was silently ignored.
   Diagnose it as an error.

 * Remote repository URLs expressed in scp-style host:path notation are
   parsed more carefully (e.g. "foo/bar:baz" is local, "[::1]:/~user" asks
   to connect to user's home directory on host at address ::1.

 * "git diff -- ':(icase)makefile'" was unnecessarily rejected at the
   command line parser.

 * "git cat-file --batch-check=ok" did not check the existence of
   the named object.

 * "git am --abort" sometimes complained about not being able to write
   a tree with an 0{40} object in it.

 * Two processes creating loose objects at the same time could have
   failed unnecessarily when the name of their new objects started
   with the same byte value, due to a race condition.