git/git.git
6 years agopacked-ref cache: forbid dot-components in refnames
Jonathan Nieder [Fri, 26 Sep 2014 19:22:22 +0000 (12:22 -0700)]
packed-ref cache: forbid dot-components in refnames

Since v1.7.9-rc1~10^2 (write_head_info(): handle "extra refs" locally,
2012-01-06), this trick to keep track of ".have" refs that are only
valid on the wire and not on the filesystem is not needed any more.

Simplify by removing support for the REFNAME_DOT_COMPONENT flag.

This means we'll be slightly stricter with invalid refs found in a
packed-refs file or during clone.  read_loose_refs() already checks
for and skips refnames with .components so it is not affected.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agobranch -d: simplify by using RESOLVE_REF_READING
Ronnie Sahlberg [Thu, 11 Sep 2014 17:34:36 +0000 (10:34 -0700)]
branch -d: simplify by using RESOLVE_REF_READING

When "git branch -d" reads the branch it is about to delete, it used
to avoid passing the RESOLVE_REF_READING ('treat missing ref as
error') flag because a symref pointing to a nonexistent ref would show
up as missing instead of as something that could be deleted.  To check
if a ref is actually missing, we then check

 - is it a symref?
 - if not, did it resolve to null_sha1?

Now we pass RESOLVE_REF_NO_RECURSE and the correct information is
returned for a symref even when it points to a missing ref.  Simplify
by relying on RESOLVE_REF_READING.

No functional change intended.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agobranch -d: avoid repeated symref resolution
Jonathan Nieder [Thu, 11 Sep 2014 01:22:48 +0000 (18:22 -0700)]
branch -d: avoid repeated symref resolution

If a repository gets in a broken state with too much symref nesting,
it cannot be repaired with "git branch -d":

 $ git symbolic-ref refs/heads/nonsense refs/heads/nonsense
 $ git branch -d nonsense
 error: branch 'nonsense' not found.

Worse, "git update-ref --no-deref -d" doesn't work for such repairs
either:

 $ git update-ref -d refs/heads/nonsense
 error: unable to resolve reference refs/heads/nonsense: Too many levels of symbolic links

Fix both by teaching resolve_ref_unsafe a new RESOLVE_REF_NO_RECURSE
flag and passing it when appropriate.

Callers can still read the value of a symref (for example to print a
message about it) with that flag set --- resolve_ref_unsafe will
resolve one level of symrefs and stop there.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoreflog test: test interaction with detached HEAD
Junio C Hamano [Sat, 13 Sep 2014 17:52:25 +0000 (10:52 -0700)]
reflog test: test interaction with detached HEAD

A proposed patch produced broken HEAD reflog entries when checking out
anything other than a branch.  The testsuite still passed, so it took
a few days for the bug to be noticed.

Add tests checking the content of the reflog after detaching and
reattaching HEAD so we don't have to rely on manual testing to catch
such problems in the future.

[jn: using 'log -g --format=%H' instead of parsing --oneline output,
 resetting state in each test so they can be safely reordered or
 skipped]

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: change resolve_ref_unsafe reading argument to be a flags field
Ronnie Sahlberg [Tue, 15 Jul 2014 19:59:36 +0000 (12:59 -0700)]
refs.c: change resolve_ref_unsafe reading argument to be a flags field

resolve_ref_unsafe takes a boolean argument for reading (a nonexistent ref
resolves successfully for writing but not for reading).  Change this to be
a flags field instead, and pass the new constant RESOLVE_REF_READING when
we want this behaviour.

While at it, swap two of the arguments in the function to put output
arguments at the end.  As a nice side effect, this ensures that we can
catch callers that were unaware of the new API so they can be audited.

Give the wrapper functions resolve_refdup and read_ref_full the same
treatment for consistency.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: make write_ref_sha1 static
Ronnie Sahlberg [Mon, 28 Apr 2014 22:36:58 +0000 (15:36 -0700)]
refs.c: make write_ref_sha1 static

No external users call write_ref_sha1 any more so let's declare it static.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agofetch.c: change s_update_ref to use a ref transaction
Ronnie Sahlberg [Mon, 28 Apr 2014 20:49:07 +0000 (13:49 -0700)]
fetch.c: change s_update_ref to use a ref transaction

Change s_update_ref to use a ref transaction for the ref update.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: ref_transaction_commit: distinguish name conflicts from other errors
Ronnie Sahlberg [Fri, 16 May 2014 21:14:38 +0000 (14:14 -0700)]
refs.c: ref_transaction_commit: distinguish name conflicts from other errors

In _commit, ENOTDIR can happen in the call to lock_ref_sha1_basic, either
when we lstat the new refname or if the name checking function reports that
the same type of conflict happened.  In both cases, it means that we can not
create the new ref due to a name conflict.

Start defining specific return codes for _commit.  TRANSACTION_NAME_CONFLICT
refers to a failure to create a ref due to a name conflict with another ref.
TRANSACTION_GENERIC_ERROR is for all other errors.

When "git fetch" is creating refs, name conflicts differ from other errors in
that they are likely to be resolved by running "git remote prune <remote>".
"git fetch" currently inspects errno to decide whether to give that advice.
Once it switches to the transaction API, it can check for
TRANSACTION_NAME_CONFLICT instead.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: pass a list of names to skip to is_refname_available
Ronnie Sahlberg [Thu, 1 May 2014 18:16:07 +0000 (11:16 -0700)]
refs.c: pass a list of names to skip to is_refname_available

Change is_refname_available to take a list of strings to exclude when
checking for conflicts instead of just one single name. We can already
exclude a single name for the sake of renames. This generalizes that support.

ref_transaction_commit already tracks a set of refs that are being deleted
in an array.  This array is then used to exclude refs from being written to
the packed-refs file.  At some stage we will want to change this array to a
struct string_list and then we can pass it to is_refname_available via the
call to lock_ref_sha1_basic.  That will allow us to perform transactions
that perform multiple renames as long as there are no conflicts within the
starting or ending state.

For example, that would allow a single transaction that contains two
renames that are both individually conflicting:

   m -> n/n
   n -> m/m

No functional change intended yet.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: call lock_ref_sha1_basic directly from commit
Ronnie Sahlberg [Thu, 1 May 2014 17:43:39 +0000 (10:43 -0700)]
refs.c: call lock_ref_sha1_basic directly from commit

Skip using the lock_any_ref_for_update wrapper and call lock_ref_sha1_basic
directly from the commit function.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: refuse to lock badly named refs in lock_ref_sha1_basic
Ronnie Sahlberg [Thu, 1 May 2014 17:40:10 +0000 (10:40 -0700)]
refs.c: refuse to lock badly named refs in lock_ref_sha1_basic

Move the check for check_refname_format from lock_any_ref_for_update to
lock_ref_sha1_basic.  At some later stage we will get rid of
lock_any_ref_for_update completely.  This has no visible impact to callers
except for the inability to lock badly named refs, which is not possible
today already for other reasons.(*)

Keep lock_any_ref_for_update as a no-op wrapper.  It is the public facing
version of this interface and keeping it as a separate function will make
it easier to experiment with the internal lock_ref_sha1_basic signature.

(*) For example, if lock_ref_sha1_basic checks the refname format and
refuses to lock badly named refs, it will not be possible to delete
such refs because the first step of deletion is to lock the ref.  We
currently already fail in that case because these refs are not recognized
to exist:

 $ cp .git/refs/heads/master .git/refs/heads/echo...\*\*
 $ git branch -D .git/refs/heads/echo...\*\*
 error: branch '.git/refs/heads/echo...**' not found.

This has been broken for a while.  Later patches in the series will start
repairing the handling of badly named refs.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorename_ref: don't ask read_ref_full where the ref came from
Ronnie Sahlberg [Wed, 30 Apr 2014 19:41:04 +0000 (12:41 -0700)]
rename_ref: don't ask read_ref_full where the ref came from

We call read_ref_full with a pointer to flags from rename_ref but since
we never actually use the returned flags we can just pass NULL here instead.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: pass the ref log message to _create/delete/update instead of _commit
Ronnie Sahlberg [Wed, 30 Apr 2014 19:22:42 +0000 (12:22 -0700)]
refs.c: pass the ref log message to _create/delete/update instead of _commit

Change the ref transaction API so that we pass the reflog message to the
create/delete/update functions instead of to ref_transaction_commit.
This allows different reflog messages for each ref update in a multi-ref
transaction.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: add an err argument to delete_ref_loose
Ronnie Sahlberg [Thu, 15 May 2014 15:25:23 +0000 (08:25 -0700)]
refs.c: add an err argument to delete_ref_loose

Add an err argument to delete_ref_loose so that we can pass a descriptive
error string back to the caller. Pass the err argument from transaction
commit to this function so that transaction users will have a nice error
string if the transaction failed due to delete_ref_loose.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agowrapper.c: add a new function unlink_or_msg
Ronnie Sahlberg [Wed, 16 Jul 2014 18:20:36 +0000 (11:20 -0700)]
wrapper.c: add a new function unlink_or_msg

This behaves like unlink_or_warn except that on failure it writes the message
to its 'err' argument, which the caller can display in an appropriate way or
ignore.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorefs.c: lock_ref_sha1_basic is used for all refs
Ronnie Sahlberg [Thu, 2 Oct 2014 14:59:02 +0000 (07:59 -0700)]
refs.c: lock_ref_sha1_basic is used for all refs

lock_ref_sha1_basic is used to lock refs that sit directly in the .git
dir such as HEAD and MERGE_HEAD in addition to the more ordinary refs
under "refs/".  Remove the note claiming otherwise.

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agowrapper.c: remove/unlink_or_warn: simplify, treat ENOENT as success
Ronnie Sahlberg [Wed, 16 Jul 2014 18:01:18 +0000 (11:01 -0700)]
wrapper.c: remove/unlink_or_warn: simplify, treat ENOENT as success

Simplify the function warn_if_unremovable slightly. Additionally, change
behaviour slightly. If we failed to remove the object because the object
does not exist, we can still return success back to the caller since none of
the callers depend on "fail if the file did not exist".

Signed-off-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agomv test: recreate mod/ directory instead of relying on stale copy
Jonathan Nieder [Wed, 10 Sep 2014 21:01:46 +0000 (14:01 -0700)]
mv test: recreate mod/ directory instead of relying on stale copy

The tests for 'git mv moves a submodule' functionality often run
commands like

git mv sub mod/sub

to move a submodule into a subdirectory.  Just like plain /bin/mv,
this is supposed to succeed if the mod/ parent directory exists
and fail if it doesn't exist.

Usually these tests mkdir the parent directory beforehand, but some
instead rely on it being left behind by previous tests.

More precisely, when 'git reset --hard' tries to move to a state where
mod/sub is not present any more, it would perform the following
operations:

rmdir("mod/sub")
rmdir("mod")

The first fails with ENOENT because the test script removed mod/sub
with "rm -rf" already, so 'reset --hard' doesn't bother to move on to
the second, and the mod/ directory is kept around.

Better to explicitly remove and re-create the mod/ directory so later
tests don't have to depend on the directory left behind by the earlier
ones at all (making it easier to rearrange or skip some tests in the
file or to tweak 'reset --hard' behavior without breaking unrelated
tests).

Noticed while testing a patch that fixes the reset --hard behavior
described above.

Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoUpdate draft release notes to 2.2
Junio C Hamano [Tue, 14 Oct 2014 17:59:04 +0000 (10:59 -0700)]
Update draft release notes to 2.2

Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoMerge branch 'bc/asciidoc-pretty-formats-fix'
Junio C Hamano [Tue, 14 Oct 2014 17:50:14 +0000 (10:50 -0700)]
Merge branch 'bc/asciidoc-pretty-formats-fix'

* bc/asciidoc-pretty-formats-fix:
  Documentation: fix misrender of pretty-formats in Asciidoctor

6 years agoMerge branch 'rs/plug-leak-in-bundle'
Junio C Hamano [Tue, 14 Oct 2014 17:50:09 +0000 (10:50 -0700)]
Merge branch 'rs/plug-leak-in-bundle'

* rs/plug-leak-in-bundle:
  bundle: plug minor memory leak in is_tag_in_date_range()

6 years agoMerge branch 'rs/more-uses-of-skip-prefix'
Junio C Hamano [Tue, 14 Oct 2014 17:50:07 +0000 (10:50 -0700)]
Merge branch 'rs/more-uses-of-skip-prefix'

* rs/more-uses-of-skip-prefix:
  use skip_prefix() to avoid more magic numbers

6 years agoMerge branch 'rs/mailsplit'
Junio C Hamano [Tue, 14 Oct 2014 17:50:00 +0000 (10:50 -0700)]
Merge branch 'rs/mailsplit'

* rs/mailsplit:
  mailsplit: remove unnecessary unlink(2) call

6 years agoMerge branch 'rs/sha1-array-test'
Junio C Hamano [Tue, 14 Oct 2014 17:49:56 +0000 (10:49 -0700)]
Merge branch 'rs/sha1-array-test'

* rs/sha1-array-test:
  sha1-lookup: handle duplicates in sha1_pos()
  sha1-array: add test-sha1-array and basic tests

6 years agoMerge branch 'mh/lockfile-stdio'
Junio C Hamano [Tue, 14 Oct 2014 17:49:51 +0000 (10:49 -0700)]
Merge branch 'mh/lockfile-stdio'

* mh/lockfile-stdio:
  commit_packed_refs(): reimplement using fdopen_lock_file()
  dump_marks(): reimplement using fdopen_lock_file()
  fdopen_lock_file(): access a lockfile using stdio

6 years agoMerge branch 'mh/lockfile'
Junio C Hamano [Tue, 14 Oct 2014 17:49:45 +0000 (10:49 -0700)]
Merge branch 'mh/lockfile'

The lockfile API and its users have been cleaned up.

* mh/lockfile: (38 commits)
  lockfile.h: extract new header file for the functions in lockfile.c
  hold_locked_index(): move from lockfile.c to read-cache.c
  hold_lock_file_for_append(): restore errno before returning
  get_locked_file_path(): new function
  lockfile.c: rename static functions
  lockfile: rename LOCK_NODEREF to LOCK_NO_DEREF
  commit_lock_file_to(): refactor a helper out of commit_lock_file()
  trim_last_path_component(): replace last_path_elm()
  resolve_symlink(): take a strbuf parameter
  resolve_symlink(): use a strbuf for internal scratch space
  lockfile: change lock_file::filename into a strbuf
  commit_lock_file(): use a strbuf to manage temporary space
  try_merge_strategy(): use a statically-allocated lock_file object
  try_merge_strategy(): remove redundant lock_file allocation
  struct lock_file: declare some fields volatile
  lockfile: avoid transitory invalid states
  git_config_set_multivar_in_file(): avoid call to rollback_lock_file()
  dump_marks(): remove a redundant call to rollback_lock_file()
  api-lockfile: document edge cases
  commit_lock_file(): rollback lock file on failure to rename
  ...

6 years agoMerge branch 'sk/tag-contains-wo-recursion'
Junio C Hamano [Tue, 14 Oct 2014 17:49:41 +0000 (10:49 -0700)]
Merge branch 'sk/tag-contains-wo-recursion'

* sk/tag-contains-wo-recursion:
  t7004: give the test a bit more stack space

6 years agoMerge branch 'da/completion-show-signature'
Junio C Hamano [Tue, 14 Oct 2014 17:49:35 +0000 (10:49 -0700)]
Merge branch 'da/completion-show-signature'

* da/completion-show-signature:
  completion: add --show-signature for log and show

6 years agoMerge branch 'rs/daemon-fixes'
Junio C Hamano [Tue, 14 Oct 2014 17:49:22 +0000 (10:49 -0700)]
Merge branch 'rs/daemon-fixes'

"git daemon" (with NO_IPV6 build configuration) used to incorrectly
use the hostname even when gethostbyname() reported that the given
hostname is not found.

* rs/daemon-fixes:
  daemon: remove write-only variable maxfd
  daemon: fix error message after bind()
  daemon: handle gethostbyname() error

6 years agoMerge branch 'dt/cache-tree-repair'
Junio C Hamano [Tue, 14 Oct 2014 17:49:11 +0000 (10:49 -0700)]
Merge branch 'dt/cache-tree-repair'

This fixes a topic that has graduated to 'master'.

* dt/cache-tree-repair:
  t0090: avoid passing empty string to printf %d

6 years agoMerge branch 'so/rebase-doc-fork-point'
Junio C Hamano [Tue, 14 Oct 2014 17:49:06 +0000 (10:49 -0700)]
Merge branch 'so/rebase-doc-fork-point'

* so/rebase-doc-fork-point:
  Documentation/git-rebase.txt: document when --fork-point is auto-enabled

6 years agoMerge branch 'da/include-compat-util-first-in-c'
Junio C Hamano [Tue, 14 Oct 2014 17:49:00 +0000 (10:49 -0700)]
Merge branch 'da/include-compat-util-first-in-c'

Code clean-up.

* da/include-compat-util-first-in-c:
  cleanups: ensure that git-compat-util.h is included first

6 years agoDocumentation: fix misrender of pretty-formats in Asciidoctor
brian m. carlson [Wed, 8 Oct 2014 20:46:10 +0000 (20:46 +0000)]
Documentation: fix misrender of pretty-formats in Asciidoctor

Neither the AsciiDoc nor the Asciidoctor documentation specify whether
the same number of delimiter characters must be used to end a block as
to begin it, although both sets of documentation show exactly matching
pairs.  AsciiDoc allows mismatches, but AsciiDoctor apparently does not.
Adjust the pretty formats documentation to use matching pairs to prevent
a misrendering where the remainder of the document was rendered as a
listing block.

Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoUpdate draft release notes to 2.2
Junio C Hamano [Wed, 8 Oct 2014 20:08:55 +0000 (13:08 -0700)]
Update draft release notes to 2.2

Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoMerge branch 'sp/stream-clean-filter'
Junio C Hamano [Wed, 8 Oct 2014 20:05:32 +0000 (13:05 -0700)]
Merge branch 'sp/stream-clean-filter'

When running a required clean filter, we do not have to mmap the
original before feeding the filter.  Instead, stream the file
contents directly to the filter and process its output.

* sp/stream-clean-filter:
  sha1_file: don't convert off_t to size_t too early to avoid potential die()
  convert: stream from fd to required clean filter to reduce used address space
  copy_fd(): do not close the input file descriptor
  mmap_limit: introduce GIT_MMAP_LIMIT to allow testing expected mmap size
  memory_limit: use git_env_ulong() to parse GIT_ALLOC_LIMIT
  config.c: add git_env_ulong() to parse environment variable
  convert: drop arguments other than 'path' from would_convert_to_git()

6 years agoMerge branch 'bw/use-write-script-in-tests'
Junio C Hamano [Wed, 8 Oct 2014 20:05:29 +0000 (13:05 -0700)]
Merge branch 'bw/use-write-script-in-tests'

* bw/use-write-script-in-tests:
  t/lib-credential: use write_script

6 years agoMerge branch 'nd/archive-pathspec'
Junio C Hamano [Wed, 8 Oct 2014 20:05:25 +0000 (13:05 -0700)]
Merge branch 'nd/archive-pathspec'

"git archive" learned to filter what gets archived with pathspec.

* nd/archive-pathspec:
  archive: support filtering paths with glob

6 years agoMerge branch 'jc/push-cert'
Junio C Hamano [Wed, 8 Oct 2014 20:05:15 +0000 (13:05 -0700)]
Merge branch 'jc/push-cert'

Allow "git push" request to be signed, so that it can be verified and
audited, using the GPG signature of the person who pushed, that the
tips of branches at a public repository really point the commits
the pusher wanted to, without having to "trust" the server.

* jc/push-cert: (24 commits)
  receive-pack::hmac_sha1(): copy the entire SHA-1 hash out
  signed push: allow stale nonce in stateless mode
  signed push: teach smart-HTTP to pass "git push --signed" around
  signed push: fortify against replay attacks
  signed push: add "pushee" header to push certificate
  signed push: remove duplicated protocol info
  send-pack: send feature request on push-cert packet
  receive-pack: GPG-validate push certificates
  push: the beginning of "git push --signed"
  pack-protocol doc: typofix for PKT-LINE
  gpg-interface: move parse_signature() to where it should be
  gpg-interface: move parse_gpg_output() to where it should be
  send-pack: clarify that cmds_sent is a boolean
  send-pack: refactor inspecting and resetting status and sending commands
  send-pack: rename "new_refs" to "need_pack_data"
  receive-pack: factor out capability string generation
  send-pack: factor out capability string generation
  send-pack: always send capabilities
  send-pack: refactor decision to send update per ref
  send-pack: move REF_STATUS_REJECT_NODELETE logic a bit higher
  ...

6 years agoSync with maint
Junio C Hamano [Tue, 7 Oct 2014 20:41:03 +0000 (13:41 -0700)]
Sync with maint

* maint:
  git-tag.txt: Add a missing hyphen to `-s`

6 years agoMerge branch 'maint-2.0' into maint
Junio C Hamano [Tue, 7 Oct 2014 20:40:51 +0000 (13:40 -0700)]
Merge branch 'maint-2.0' into maint

* maint-2.0:
  git-tag.txt: Add a missing hyphen to `-s`

6 years agoMerge branch 'maint-1.9' into maint-2.0
Junio C Hamano [Tue, 7 Oct 2014 20:40:39 +0000 (13:40 -0700)]
Merge branch 'maint-1.9' into maint-2.0

* maint-1.9:
  git-tag.txt: Add a missing hyphen to `-s`

6 years agoMerge branch 'maint-1.8.5' into maint-1.9
Junio C Hamano [Tue, 7 Oct 2014 20:40:19 +0000 (13:40 -0700)]
Merge branch 'maint-1.8.5' into maint-1.9

* maint-1.8.5:
  git-tag.txt: Add a missing hyphen to `-s`

6 years agoMerge branch 'jk/mbox-from-line' into maint
Junio C Hamano [Tue, 7 Oct 2014 20:39:24 +0000 (13:39 -0700)]
Merge branch 'jk/mbox-from-line' into maint

Some MUAs mangled a line in a message that begins with "From " to
">From " when writing to a mailbox file and feeding such an input to
"git am" used to lose such a line.

* jk/mbox-from-line:
  mailinfo: work around -Wstring-plus-int warning
  mailinfo: make ">From" in-body header check more robust

6 years agocompletion: add --show-signature for log and show
David Aguilar [Sat, 4 Oct 2014 23:20:38 +0000 (16:20 -0700)]
completion: add --show-signature for log and show

Signed-off-by: David Aguilar <davvid@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agouse skip_prefix() to avoid more magic numbers
René Scharfe [Sat, 4 Oct 2014 18:54:50 +0000 (20:54 +0200)]
use skip_prefix() to avoid more magic numbers

Continue where ae021d87 (use skip_prefix to avoid magic numbers) left off
and use skip_prefix() in more places for determining the lengths of prefix
strings to avoid using dependent constants and other indirect methods.

Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agogit-tag.txt: Add a missing hyphen to `-s`
Wieland Hoffmann [Sat, 4 Oct 2014 16:27:16 +0000 (18:27 +0200)]
git-tag.txt: Add a missing hyphen to `-s`

Signed-off-by: Wieland Hoffmann <themineo@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agomailsplit: remove unnecessary unlink(2) call
René Scharfe [Sat, 4 Oct 2014 08:41:13 +0000 (10:41 +0200)]
mailsplit: remove unnecessary unlink(2) call

The output file hasn't been created at this point, yet, so there is no
need to delete it when exiting early.

Suggested-by: Jeff King <peff@peff.net>
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agobundle: plug minor memory leak in is_tag_in_date_range()
René Scharfe [Fri, 3 Oct 2014 22:40:24 +0000 (00:40 +0200)]
bundle: plug minor memory leak in is_tag_in_date_range()

Free the buffer returned by read_sha1_file() even if no valid tagger
line is found.

Signed-off-by: Rene Scharfe <l.s.r@web.de>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agocommit_packed_refs(): reimplement using fdopen_lock_file()
Michael Haggerty [Wed, 1 Oct 2014 11:14:49 +0000 (13:14 +0200)]
commit_packed_refs(): reimplement using fdopen_lock_file()

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agodump_marks(): reimplement using fdopen_lock_file()
Michael Haggerty [Wed, 1 Oct 2014 11:14:48 +0000 (13:14 +0200)]
dump_marks(): reimplement using fdopen_lock_file()

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agofdopen_lock_file(): access a lockfile using stdio
Michael Haggerty [Wed, 1 Oct 2014 11:14:47 +0000 (13:14 +0200)]
fdopen_lock_file(): access a lockfile using stdio

Add a new function, fdopen_lock_file(), which returns a FILE pointer
open to the lockfile. If a stream is open on a lock_file object, it is
closed using fclose() on commit, rollback, or close_lock_file().

This change will allow callers to use stdio to write to a lockfile
without having to muck around in the internal representation of the
lock_file object (callers will be rewritten in upcoming commits).

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolockfile.h: extract new header file for the functions in lockfile.c
Michael Haggerty [Wed, 1 Oct 2014 10:28:42 +0000 (12:28 +0200)]
lockfile.h: extract new header file for the functions in lockfile.c

Move the interface declaration for the functions in lockfile.c from
cache.h to a new file, lockfile.h. Add #includes where necessary (and
remove some redundant includes of cache.h by files that already
include builtin.h).

Move the documentation of the lock_file state diagram from lockfile.c
to the new header file.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agohold_locked_index(): move from lockfile.c to read-cache.c
Michael Haggerty [Wed, 1 Oct 2014 10:28:41 +0000 (12:28 +0200)]
hold_locked_index(): move from lockfile.c to read-cache.c

lockfile.c contains the general API for locking any file. Code
specifically about the index file doesn't belong here.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agohold_lock_file_for_append(): restore errno before returning
Michael Haggerty [Wed, 1 Oct 2014 10:28:40 +0000 (12:28 +0200)]
hold_lock_file_for_append(): restore errno before returning

Callers who don't pass LOCK_DIE_ON_ERROR might want to examine errno
to see what went wrong, so restore errno before returning.

In fact this function only has one caller, add_to_alternates_file(),
and it *does* use LOCK_DIE_ON_ERROR, but, you know, think of future
generations.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoget_locked_file_path(): new function
Michael Haggerty [Wed, 1 Oct 2014 10:28:39 +0000 (12:28 +0200)]
get_locked_file_path(): new function

Add a function to return the path of the file that is locked by a
lock_file object. This reduces the knowledge that callers have to have
about the lock_file layout.

Suggested-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolockfile.c: rename static functions
Michael Haggerty [Wed, 1 Oct 2014 10:28:38 +0000 (12:28 +0200)]
lockfile.c: rename static functions

* remove_lock_file() -> remove_lock_files()
* remove_lock_file_on_signal() -> remove_lock_files_on_signal()

Suggested-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolockfile: rename LOCK_NODEREF to LOCK_NO_DEREF
Michael Haggerty [Wed, 1 Oct 2014 10:28:37 +0000 (12:28 +0200)]
lockfile: rename LOCK_NODEREF to LOCK_NO_DEREF

This makes it harder to misread the name as LOCK_NODE_REF.

Suggested-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agocommit_lock_file_to(): refactor a helper out of commit_lock_file()
Michael Haggerty [Wed, 1 Oct 2014 10:28:36 +0000 (12:28 +0200)]
commit_lock_file_to(): refactor a helper out of commit_lock_file()

commit_locked_index(), when writing to an alternate index file,
duplicates (poorly) the code in commit_lock_file(). And anyway, it
shouldn't have to know so much about the internal workings of lockfile
objects. So extract a new function commit_lock_file_to() that does the
work common to the two functions, and call it from both
commit_lock_file() and commit_locked_index().

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agotrim_last_path_component(): replace last_path_elm()
Michael Haggerty [Wed, 1 Oct 2014 10:28:35 +0000 (12:28 +0200)]
trim_last_path_component(): replace last_path_elm()

Rewrite last_path_elm() to take a strbuf parameter and to trim off the
last path name element in place rather than returning a pointer to the
beginning of the last path name element. This simplifies the function
a bit and makes it integrate better with its caller, which is now also
strbuf-based. Rename the function accordingly and a bit less tersely.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoresolve_symlink(): take a strbuf parameter
Michael Haggerty [Wed, 1 Oct 2014 10:28:34 +0000 (12:28 +0200)]
resolve_symlink(): take a strbuf parameter

Change resolve_symlink() to take a strbuf rather than a string as
parameter.  This simplifies the code and removes an arbitrary pathname
length restriction.  It also means that lock_file's filename field no
longer needs to be initialized to a large size.

Helped-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoresolve_symlink(): use a strbuf for internal scratch space
Michael Haggerty [Wed, 1 Oct 2014 10:28:33 +0000 (12:28 +0200)]
resolve_symlink(): use a strbuf for internal scratch space

Aside from shortening and simplifying the code, this removes another
place where the path name length is arbitrarily limited.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolockfile: change lock_file::filename into a strbuf
Michael Haggerty [Wed, 1 Oct 2014 10:28:32 +0000 (12:28 +0200)]
lockfile: change lock_file::filename into a strbuf

For now, we still make sure to allocate at least PATH_MAX characters
for the strbuf because resolve_symlink() doesn't know how to expand
the space for its return value.  (That will be fixed in a moment.)

Another alternative would be to just use a strbuf as scratch space in
lock_file() but then store a pointer to the naked string in struct
lock_file.  But lock_file objects are often reused.  By reusing the
same strbuf, we can avoid having to reallocate the string most times
when a lock_file object is reused.

Helped-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agocommit_lock_file(): use a strbuf to manage temporary space
Michael Haggerty [Wed, 1 Oct 2014 10:28:31 +0000 (12:28 +0200)]
commit_lock_file(): use a strbuf to manage temporary space

Avoid relying on the filename length restrictions that are currently
checked by lock_file().

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agotry_merge_strategy(): use a statically-allocated lock_file object
Michael Haggerty [Wed, 1 Oct 2014 10:28:30 +0000 (12:28 +0200)]
try_merge_strategy(): use a statically-allocated lock_file object

Even the one lockfile object needn't be allocated each time the
function is called.  Instead, define one statically-allocated
lock_file object and reuse it for every call.

Suggested-by: Jeff King <peff@peff.net>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agotry_merge_strategy(): remove redundant lock_file allocation
Michael Haggerty [Wed, 1 Oct 2014 10:28:29 +0000 (12:28 +0200)]
try_merge_strategy(): remove redundant lock_file allocation

By the time the "if" block is entered, the lock_file instance from the
main function block is no longer in use, so re-use that one instead of
allocating a second one.

Note that the "lock" variable in the "if" block shadowed the "lock"
variable at function scope, so the only change needed is to remove the
inner definition.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agostruct lock_file: declare some fields volatile
Michael Haggerty [Wed, 1 Oct 2014 10:28:28 +0000 (12:28 +0200)]
struct lock_file: declare some fields volatile

The function remove_lock_file_on_signal() is used as a signal handler.
It is not realistic to make the signal handler conform strictly to the
C standard, which is very restrictive about what a signal handler is
allowed to do.  But let's increase the likelihood that it will work:

The lock_file_list global variable and several fields from struct
lock_file are used by the signal handler.  Declare those values
"volatile" to (1) force the main process to write the values to RAM
promptly, and (2) prevent updates to these fields from being reordered
in a way that leaves an opportunity for a jump to the signal handler
while the object is in an inconsistent state.

We don't mark the filename field volatile because that would prevent
the use of strcpy(), and it is anyway unlikely that a compiler
re-orders a strcpy() call across other expressions.  So in practice it
should be possible to get away without "volatile" in the "filename"
case.

Suggested-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolockfile: avoid transitory invalid states
Michael Haggerty [Wed, 1 Oct 2014 10:28:27 +0000 (12:28 +0200)]
lockfile: avoid transitory invalid states

Because remove_lock_file() can be called any time by the signal
handler, it is important that any lock_file objects that are in the
lock_file_list are always in a valid state.  And since lock_file
objects are often reused (but are never removed from lock_file_list),
that means we have to be careful whenever mutating a lock_file object
to always keep it in a well-defined state.

This was formerly not the case, because part of the state was encoded
by setting lk->filename to the empty string vs. a valid filename.  It
is wrong to assume that this string can be updated atomically; for
example, even

    strcpy(lk->filename, value)

is unsafe.  But the old code was even more reckless; for example,

    strcpy(lk->filename, path);
    if (!(flags & LOCK_NODEREF))
            resolve_symlink(lk->filename, max_path_len);
    strcat(lk->filename, ".lock");

During the call to resolve_symlink(), lk->filename contained the name
of the file that was being locked, not the name of the lockfile.  If a
signal were raised during that interval, then the signal handler would
have deleted the valuable file!

We could probably continue to use the filename field to encode the
state by being careful to write characters 1..N-1 of the filename
first, and then overwrite the NUL at filename[0] with the first
character of the filename, but that would be awkward and error-prone.

So, instead of using the filename field to determine whether the
lock_file object is active, add a new field "lock_file::active" for
this purpose.  Be careful to set this field only when filename really
contains the name of a file that should be deleted on cleanup.

Helped-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agogit_config_set_multivar_in_file(): avoid call to rollback_lock_file()
Michael Haggerty [Wed, 1 Oct 2014 10:28:26 +0000 (12:28 +0200)]
git_config_set_multivar_in_file(): avoid call to rollback_lock_file()

After commit_lock_file() is called, then the lock_file object is
necessarily either committed or rolled back.  So there is no need to
call rollback_lock_file() again in either of these cases.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agodump_marks(): remove a redundant call to rollback_lock_file()
Michael Haggerty [Wed, 1 Oct 2014 10:28:25 +0000 (12:28 +0200)]
dump_marks(): remove a redundant call to rollback_lock_file()

When commit_lock_file() fails, it now always calls
rollback_lock_file() internally, so there is no need to call that
function here.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoapi-lockfile: document edge cases
Michael Haggerty [Wed, 1 Oct 2014 10:28:24 +0000 (12:28 +0200)]
api-lockfile: document edge cases

* Document the behavior of commit_lock_file() when it fails, namely
  that it rolls back the lock_file object and sets errno
  appropriately.

* Document the behavior of rollback_lock_file() when called for a
  lock_file object that has already been committed or rolled back,
  namely that it is a NOOP.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agocommit_lock_file(): rollback lock file on failure to rename
Michael Haggerty [Wed, 1 Oct 2014 10:28:23 +0000 (12:28 +0200)]
commit_lock_file(): rollback lock file on failure to rename

If rename() fails, call rollback_lock_file() to delete the lock file
(in case it is still present) and reset the filename field to the
empty string so that the lockfile object is left in a valid state.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoclose_lock_file(): if close fails, roll back
Michael Haggerty [Wed, 1 Oct 2014 10:28:22 +0000 (12:28 +0200)]
close_lock_file(): if close fails, roll back

If closing an open lockfile fails, then we cannot be sure of the
contents of the lockfile, so there is nothing sensible to do but
delete it. This change also insures that the lock_file object is left
in a defined state in this error path (namely, unlocked).

The only caller that is ultimately affected by this change is
try_merge_strategy() -> write_locked_index(), which can call
close_lock_file() via various execution paths. This caller uses a
static lock_file object which previously could have been reused after
a failed close_lock_file() even though it was still in locked state.
This change causes the lock_file object to be unlocked on failure,
thus fixing this error-handling path.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agocommit_lock_file(): die() if called for unlocked lockfile object
Michael Haggerty [Wed, 1 Oct 2014 10:28:21 +0000 (12:28 +0200)]
commit_lock_file(): die() if called for unlocked lockfile object

It was previously a bug to call commit_lock_file() with a lock_file
object that was not active (an illegal access would happen within the
function).  It was presumably never done, but this would be an easy
programming error to overlook.  So before continuing, do a consistency
check that the lock_file object really is locked.

Helped-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agocommit_lock_file(): inline temporary variable
Michael Haggerty [Wed, 1 Oct 2014 10:28:20 +0000 (12:28 +0200)]
commit_lock_file(): inline temporary variable

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoremove_lock_file(): call rollback_lock_file()
Michael Haggerty [Wed, 1 Oct 2014 10:28:19 +0000 (12:28 +0200)]
remove_lock_file(): call rollback_lock_file()

It does just what we need.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolock_file(): exit early if lockfile cannot be opened
Michael Haggerty [Wed, 1 Oct 2014 10:28:18 +0000 (12:28 +0200)]
lock_file(): exit early if lockfile cannot be opened

This is a bit easier to read than the old version, which nested part
of the non-error code in an "if" block.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoprepare_index(): declare return value to be (const char *)
Michael Haggerty [Wed, 1 Oct 2014 10:28:17 +0000 (12:28 +0200)]
prepare_index(): declare return value to be (const char *)

Declare the return value to be const to make it clear that we aren't
giving callers permission to write over the string that it points at.
(The return value is the filename field of a struct lock_file, which
can be used by a signal handler at any time and therefore shouldn't be
tampered with.)

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agodelete_ref_loose(): don't muck around in the lock_file's filename
Michael Haggerty [Wed, 1 Oct 2014 10:28:16 +0000 (12:28 +0200)]
delete_ref_loose(): don't muck around in the lock_file's filename

It's bad manners. Especially since there could be a signal during the
call to unlink_or_warn(), in which case the signal handler will see
the wrong filename and delete the reference file, leaving the lockfile
behind.

So make our own copy to work with.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agocache.h: define constants LOCK_SUFFIX and LOCK_SUFFIX_LEN
Michael Haggerty [Wed, 1 Oct 2014 10:28:15 +0000 (12:28 +0200)]
cache.h: define constants LOCK_SUFFIX and LOCK_SUFFIX_LEN

There are a few places that use these values, so define constants for
them.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolockfile.c: document the various states of lock_file objects
Michael Haggerty [Wed, 1 Oct 2014 10:28:14 +0000 (12:28 +0200)]
lockfile.c: document the various states of lock_file objects

Document the valid states of lock_file objects, how they get into each
state, and how the state is encoded in the object's fields.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolock_file(): always initialize and register lock_file object
Michael Haggerty [Wed, 1 Oct 2014 10:28:13 +0000 (12:28 +0200)]
lock_file(): always initialize and register lock_file object

The purpose of this change is to make the state diagram for
lock_file objects simpler and deterministic.

If locking fails, lock_file() sometimes leaves the lock_file object
partly initialized, but sometimes not. It sometimes registers the
object in lock_file_list, but sometimes not. This makes the state
diagram for lock_file objects effectively indeterministic and hard
to reason about. A future patch will also change the filename field
into a strbuf, which needs more involved initialization, so it will
become even more important that the state of a lock_file object is
well-defined after a failed attempt to lock.

The ambiguity doesn't currently have any ill effects, because
lock_file objects cannot be removed from the lock_file_list anyway.
But to make it easier to document and reason about the code, make
this behavior consistent: *always* initialize the lock_file object
and *always* register it in lock_file_list the first time it is
used, regardless of whether an error occurs.

While we're at it, make sure that all of the lock_file fields are
initialized to values appropriate for an unlocked object; the caller
is only responsible for making sure that on_list is set to zero before
the first time it is used.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agohold_lock_file_for_append(): release lock on errors
Michael Haggerty [Wed, 1 Oct 2014 10:28:12 +0000 (12:28 +0200)]
hold_lock_file_for_append(): release lock on errors

If there is an error copying the old contents to the lockfile, roll
back the lockfile before exiting so that the lockfile is not held
until process cleanup.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agolockfile: unlock file if lockfile permissions cannot be adjusted
Michael Haggerty [Wed, 1 Oct 2014 10:28:11 +0000 (12:28 +0200)]
lockfile: unlock file if lockfile permissions cannot be adjusted

If the call to adjust_shared_perm() fails, lock_file returns -1, which
to the caller looks like any other failure to lock the file.  So in
this case, roll back the lockfile before returning so that the lock
file is deleted immediately and the lockfile object is left in a
predictable state (namely, unlocked).  Previously, the lockfile was
retained until process cleanup in this situation.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorollback_lock_file(): set fd to -1
Michael Haggerty [Wed, 1 Oct 2014 10:28:10 +0000 (12:28 +0200)]
rollback_lock_file(): set fd to -1

When rolling back the lockfile, call close_lock_file() so that the
lock_file's fd field gets set back to -1. This keeps the lock_file
object in a valid state, which is important because these objects are
allowed to be reused. It also makes it unnecessary to check whether
the file has already been closed, because close_lock_file() takes care
of that.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorollback_lock_file(): exit early if lock is not active
Michael Haggerty [Wed, 1 Oct 2014 10:28:09 +0000 (12:28 +0200)]
rollback_lock_file(): exit early if lock is not active

Eliminate a layer of nesting.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agorollback_lock_file(): do not clear filename redundantly
Michael Haggerty [Wed, 1 Oct 2014 10:28:08 +0000 (12:28 +0200)]
rollback_lock_file(): do not clear filename redundantly

It is only necessary to clear the lock_file's filename field if it was
not already clear.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoclose_lock_file(): exit (successfully) if file is already closed
Michael Haggerty [Wed, 1 Oct 2014 10:28:07 +0000 (12:28 +0200)]
close_lock_file(): exit (successfully) if file is already closed

Suggested-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoapi-lockfile: revise and expand the documentation
Michael Haggerty [Wed, 1 Oct 2014 10:28:06 +0000 (12:28 +0200)]
api-lockfile: revise and expand the documentation

Document a couple more functions and the flags argument as used by
hold_lock_file_for_update() and hold_lock_file_for_append().
Reorganize the document to make it more accessible.

Helped-by: Jonathan Nieder <jrnieder@gmail.com>
Helped-by: Junio Hamano <gitster@pobox.com>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agounable_to_lock_die(): rename function from unable_to_lock_index_die()
Michael Haggerty [Wed, 1 Oct 2014 10:28:05 +0000 (12:28 +0200)]
unable_to_lock_die(): rename function from unable_to_lock_index_die()

This function is used for other things besides the index, so rename it
accordingly.

Suggested-by: Jeff King <peff@peff.net>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Ronnie Sahlberg <sahlberg@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agodaemon: remove write-only variable maxfd
René Scharfe [Wed, 1 Oct 2014 10:21:57 +0000 (12:21 +0200)]
daemon: remove write-only variable maxfd

It became unused when 6573faff (NO_IPV6 support for git daemon) replaced
select() with poll().

Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agodaemon: fix error message after bind()
René Scharfe [Wed, 1 Oct 2014 10:18:15 +0000 (12:18 +0200)]
daemon: fix error message after bind()

Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agodaemon: handle gethostbyname() error
René Scharfe [Wed, 1 Oct 2014 10:16:17 +0000 (12:16 +0200)]
daemon: handle gethostbyname() error

If the user-supplied hostname can't be found then we should not use it.
We already avoid doing that in the non-NO_IPV6 case by checking if the
return value of getaddrinfo() is zero (success).  Do the same in the
NO_IPV6 case and make sure the return value of gethostbyname() isn't
NULL before dereferencing this pointer.

Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agosha1-lookup: handle duplicates in sha1_pos()
René Scharfe [Wed, 1 Oct 2014 15:02:37 +0000 (17:02 +0200)]
sha1-lookup: handle duplicates in sha1_pos()

If the first 18 bytes of the SHA1's of all entries are the same then
sha1_pos() dies and reports that the lower and upper limits of the
binary search were the same that this wasn't supposed to happen.  This
is wrong because the remaining two bytes could still differ.

Furthermore: It wouldn't be a problem if they actually were the same,
i.e. if all entries have the same SHA1.  The code already handles
duplicates just fine.  Simply remove the erroneous check.

Signed-off-by: Rene Scharfe <l.s.r@web.de>
Acked-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agosha1-array: add test-sha1-array and basic tests
René Scharfe [Wed, 1 Oct 2014 15:00:33 +0000 (17:00 +0200)]
sha1-array: add test-sha1-array and basic tests

Helped-by: Jeff King <peff@peff.net>
Helped-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Acked-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agot0090: avoid passing empty string to printf %d
René Scharfe [Tue, 30 Sep 2014 17:42:03 +0000 (19:42 +0200)]
t0090: avoid passing empty string to printf %d

FreeBSD's printf(1) doesn't accept empty strings for numerical format
specifiers:

$ printf "%d\n" "" >/dev/null; echo $?
printf: : expected numeric value
1

Initialize the AWK variable c to make sure the shell variable
subtree_count always contains a numerical value, in order to keep the
subsequently called printf happy.

Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
6 years agoSync with 2.1.2
Junio C Hamano [Tue, 30 Sep 2014 05:17:57 +0000 (22:17 -0700)]
Sync with 2.1.2

* maint:
  Git 2.1.2

6 years agoMerge branch 'jt/itimer-autoconf'
Junio C Hamano [Tue, 30 Sep 2014 05:17:23 +0000 (22:17 -0700)]
Merge branch 'jt/itimer-autoconf'

setitmer(2) and related API elements can be configured from
Makefile but autoconf did not know about it.

* jt/itimer-autoconf:
  autoconf: check for setitimer()
  autoconf: check for struct itimerval
  git-compat-util.h: add missing semicolon after struct itimerval

6 years agoMerge branch 'jc/test-lazy-prereq'
Junio C Hamano [Tue, 30 Sep 2014 05:17:22 +0000 (22:17 -0700)]
Merge branch 'jc/test-lazy-prereq'

Test-script clean-up.

* jc/test-lazy-prereq:
  tests: drop GIT_*_TIMING_TESTS environment variable support

6 years agoMerge branch 'sb/merge-recursive-copy-paste-fix'
Junio C Hamano [Tue, 30 Sep 2014 05:17:21 +0000 (22:17 -0700)]
Merge branch 'sb/merge-recursive-copy-paste-fix'

"git merge-recursive" had a small bug that could have made it
mishandle "one side deleted, the other side did not touch it" in a
rare corner case, where the other side actually did touch to cause
the blob object names to be different but both blobs before and
after the change normalize to the same (e.g. correcting mistake to
check in a blob with CRLF line endings by replacing it with another
blob that records the same contents with LF line endings).

* sb/merge-recursive-copy-paste-fix:
  merge-recursive: remove stale commented debugging code
  merge-recursive: fix copy-paste mistake

6 years agoMerge branch 'pr/use-default-sigpipe-setting'
Junio C Hamano [Tue, 30 Sep 2014 05:17:20 +0000 (22:17 -0700)]
Merge branch 'pr/use-default-sigpipe-setting'

We used to get confused when a process called us with SIGPIPE
ignored; we do want to die with SIGPIPE when the output is not
read by default, and do ignore the signal when appropriate.

* pr/use-default-sigpipe-setting:
  mingw.h: add dummy functions for sigset_t operations
  unblock and unignore SIGPIPE