The variable was documented in git-mailinfo.txt, but not in config.txt.
The detailed documentation is still the one of --scissors in
git-mailinfo.txt, but we give enough information here to let the user
understand what it is about, and to make it easy to find it (e.g.
searching ">8" and "8<" finds it).
Signed-off-by: Matthieu Moy <Matthieu.Moy@imag.fr>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Provide a callback function for strbuf_expand() instead of using the
helper strbuf_expand_dict_cb(). While the resulting code is longer, it
only looks up the canonical hostname and IP address if at least one of
the placeholders %CH and %IP are used with --interpolated-path.
Use a struct for passing the directory to the callback function instead
of passing it directly to avoid having to cast away its const qualifier.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Look up canonical hostname and IP address using getaddrinfo(3) or
gethostbyname(3) only if --interpolated-path or --access-hook were
specified.
Do that by introducing getter functions for canon_hostname and
ip_address and using them for all read accesses. These wrappers call
the new helper lookup_hostname(), which sets the variables only at its
first call.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We use the daemon_avoid_alias function to make sure that the
pathname the user gives us is sane. However, after applying
that check, we might then interpolate the path using a
string given by the server admin, but which may contain more
untrusted data from the client. We should be sure to
sanitize this data, as well.
We cannot use daemon_avoid_alias here, as it is more strict
than we need in requiring a leading '/'. At the same time,
we can be much more strict here. We are interpreting a
hostname, which should not contain slashes or excessive runs
of dots, as those things are not allowed in DNS names.
Note that in addition to cleansing the hostname field, we
must check the "canonical hostname" (%CH) as well as the
port (%P), which we take as a raw string. For the canonical
hostname, this comes from an actual DNS lookup on the
accessed IP, which makes it a much less likely vector for
problems. But it does not hurt to sanitize it in the same
way. Unfortunately we cannot test this case easily, as it
would involve a custom hostname lookup.
We do not need to check %IP, as it comes straight from
inet_ntop, so must have a sane form.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We did not test this at all; let's just give a basic sanity
check that we can find a path based on virtual hosting, and
that the downcase canonicalization works.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we connect to a git-daemon at a given host and port, we
actually send the string "localhost:9418" to the other side,
which allows it to do virtual-hosting lookups. For testing
and debugging, we'd like to be able to send arbitrary
strings, rather than the hostname we actually connected to.
Using "insteadOf" config does not work for this purpose, as
the hostname determination happens at a very low level,
right before we feed the hostname to our lookup routines.
You could use /etc/hosts or similar to get around this, but
we cannot do that portably from our test suite.
Instead, this patch provides an environment variable that
can be used to send an arbitrary string.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This fixes a memory leak when building the cache entries as
refresh_cache_entry may decide to return NULL, but it does not
free the cache entry structure which was passed in as an argument.
Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Only Perl version 5.8.0 or later is required, but that comes with
an older Getopt::Long (2.32) that does not support the 'no-'
prefix. Support for that was added in Getopt::Long version 2.33.
Since the help only mentions the 'no-' prefix and not the 'no'
prefix, add explicit support for the 'no-' prefix to support
older GetOptions versions.
Reported-by: Tom G. Christensen <tgc@statsbiblioteket.dk>
Signed-off-by: Kyle J. McKay <mackyle@gmail.com>
Tested-by: Tom G. Christensen <tgc@statsbiblioteket.dk>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
What we wanted out of the SANITY precondition is that the filesystem
behaves sensibly with permission bits settings.
- You should not be able to remove a file in a read-only directory,
- You should not be able to tell if a file in a directory exists if
the directory lacks read or execute permission bits.
We used to cheat by approximating that condition with "is the /
writable?" test and/or "are we running as root?" test. Neither test
is sufficient or appropriate in environments like Cygwin.
Signed-off-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Update this ancient test script to a more modern style in which the
expected result is prepared inside the body of the test that uses
it. Also, instead of using $tree, a shell variable, throughout the
test script, create a tag that points at it, to make it easier to
manually debug the test script in its trash directory.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The "sanitize" helper wanted to strip the similarity and
dissimilarity scores when making comparison, but it was
stripping away the object names as well.
While we do not want to require the exact object names the tests
expect to be maintained, as it would be seen as an extra burden,
this would have prevented us catching a silly bug such as showing
non 0{40} object name on the preimage side of an addition or on the
postimage side of a deletion, because all [0-9a-f]{40} strings were
considered equally OK.
In the longer term, when a test only wants to see the status of the
change without having to worry about object names, it should be
rewritten not to inspect the raw format.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These two files have been modified since the tests started using
as test input, making the exact object names they expect to be
different from what actually happens in the trash repository they
use to run tests.
Instead, take a snapshot of these two files and keep them in
t/diff-lib/ so that we can update the real ones without having to
worry about breaking tests.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The output the test expects is bogus.
It was left unnoticed only because compare_diff_raw, which only
cares about the add/delete/rename/copy was used to check the result.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The output the test #36 expects is bogus. There are no blob objects
whose names are 36a590... or 046d037... when this test was run.
It was left unnoticed only because compare_diff_raw, which only
cares about the add/delete/rename/copy was used to check the result.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A complete rewrite of a single file was originally designed to be
expressed as a deletion immediately followed by a creation of the
same file, and the comments in the test updated here were written to
reflect that design decision made in f345b0a0 (Add -B flag to diff-*
brothers., 2005-05-30). However, we later realized that a complete
rewrite is merely how a textual diff should be represented at
366175ef (Rework -B output., 2005-06-19), and updated the actual
tests. But we forgot to update the introductory text while doing
so.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
41 bytes is the exact number of bytes needed for having the returned
hex string represented. 50 seems to be an arbitrary number, such
that there are no benefits from alignment to certain address boundaries.
Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A few files include the same header file directly more than once.
As all these headers protect themselves against repeated inclusion
by the "#ifndef FOO_H / #define FOO_H / ... / #endif" idiom, leave
only the first inclusion and remove the later inclusion as a no-op
clean-up.
Signed-off-by: Дилян Палаузов <git-dpa@aegee.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
run_setup_gently() is called before merge-file. This may result in changing
current working directory, which wasn't taken into account when opening a file
for writing.
Fix by prepending the passed prefix. Previous var is left so that error
messages keep referring to the file from the user's working directory
perspective.
Signed-off-by: Aleksander Boruch-Gruszecki <aleksander.boruchgruszecki@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Because Git tracks symbolic links as symbolic links, a path that
has a symbolic link in its leading part (e.g. path/to/dir/file,
where path/to/dir is a symbolic link to somewhere else, be it
inside or outside the working tree) can never appear in a patch
that validly applies, unless the same patch first removes the
symbolic link to allow a directory to be created there.
Detect and reject such a patch.
Things to note:
- Unfortunately, we cannot reuse the has_symlink_leading_path()
from dir.c, as that is only about the working tree, but "git
apply" can be told to apply the patch only to the index or to
both the index and to the working tree.
- We cannot directly use has_symlink_leading_path() even when we
are applying only to the working tree, as an early patch of a
valid input may remove a symbolic link path/to/dir and then a
later patch of the input may create a path path/to/dir/file, but
"git apply" first checks the input without touching either the
index or the working tree. The leading symbolic link check must
be done on the interim result we compute in-core (i.e. after the
first patch, there is no path/to/dir symbolic link and it is
perfectly valid to create path/to/dir/file).
Similarly, when an input creates a symbolic link path/to/dir and
then creates a file path/to/dir/file, we need to flag it as an
error without actually creating path/to/dir symbolic link in the
filesystem.
Instead, for any patch in the input that leaves a path (i.e. a non
deletion) in the result, we check all leading paths against the
resulting tree that the patch would create by inspecting all the
patches in the input and then the target of patch application
(either the index or the working tree).
This way, we catch a mischief or a mistake to add a symbolic link
path/to/dir and a file path/to/dir/file at the same time, while
allowing a valid patch that removes a symbolic link path/to/dir and
then adds a file path/to/dir/file.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We should reject a patch, whether it renames/copies dir/file to
elsewhere with or without modificiation, or updates dir/file in
place, if "dir/" part is actually a symbolic link to elsewhere,
by making sure that the code to read the preimage does not read
from a path that is beyond a symbolic link.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We currently read the preimage to apply a patch from the index only
when the --cached option is given. Do so also when the command is
running under the --index option. With --index, the index entry and
the working tree file for a path that is involved in a patch must be
identical, so this should not affect the result, but by reading from
the index, we will get the protection to avoid reading an unintended
path beyond a symbolic link automatically.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
By default, a patch that affects outside the working area (either a
Git controlled working tree, or the current working directory when
"git apply" is used as a replacement of GNU patch) is rejected as a
mistake (or a mischief). Git itself does not create such a patch,
unless the user bends over backwards and specifies a non-standard
prefix to "git diff" and friends.
When `git apply` is used as a "better GNU patch", the user can pass
the `--unsafe-paths` option to override this safety check. This
option has no effect when `--index` or `--cached` is in use.
The new test was stolen from Jeff King with slight enhancements.
Note that a few new tests for touching outside the working area by
following a symbolic link are still expected to fail at this step,
but will be fixed in later steps.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When an import has finished, we run end_packfile() to
finalize the data and move the packfile into place. If this
process fails, we call die() and end up in our die_nicely()
handler. Which unfortunately includes running end_packfile
to save any progress we made. We enter the function again,
and start operating on the pack_data struct while it is in
an inconsistent state, leading to a segfault.
One way to trigger this is to simply start two identical
fast-imports at the same time. They will both create the
same packfiles, which will then try to create identically
named ".keep" files. One will win the race, and the other
will die(), and end up with the segfault.
Since 3c078b9, we already reset the pack_data pointer to
NULL at the end of end_packfile. That covers the case of us
calling die() right after end_packfile, before we have
reinitialized the pack_data pointer. This new problem is
quite similar, except that we are worried about calling
die() _during_ end_packfile, not right after. Ideally we
would simply set pack_data to NULL as soon as we enter the
function, and operate on a copy of the pointer.
Unfortunately, it is not so easy. pack_data is a global, and
end_packfile calls into other functions which operate on the
global directly. We would have to teach each of these to
take an argument, and there is no guarantee that we would
catch all of the spots.
Instead, we can simply use a static flag to avoid
recursively entering the function. This is a little less
elegant, but it's short and fool-proof.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since ea02ffa3 (mailmap: simplify map_user() interface, 2013-01-05),
find_alignment() has been invoking commit_info_destroy() on an
uninitialized auto 'struct commit_info' (when METAINFO_SHOWN is not
set). commit_info_destroy() calls strbuf_release() for each
'commit_info' strbuf member, which randomly invokes free() on
whatever random stack value happens to reside in strbuf.buf, thus
leading to periodic crashes.
Reported-by: Dilyan Palauzov <dilyan.palauzov@aegee.org>
Signed-off-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The string in 'base' contains a path suffix to a specific object;
when its value is used, the suffix must either be filled (as in
stat_sha1_file, open_sha1_file, check_and_freshen_nonlocal) or
cleared (as in prepare_packed_git) to avoid junk at the end.
660c889e (sha1_file: add for_each iterators for loose and packed
objects, 2014-10-15) introduced loose_from_alt_odb(), but this did
neither and treated 'base' as a complete path to the "base" object
directory, instead of a pointer to the "base" of the full path
string.
The trailing path after 'base' is still initialized to NUL, hiding
the bug in some common cases. Additionally the descendent
for_each_file_in_obj_subdir() function swallows ENOENT, so an error
only shows if the alternate's path was last filled with a valid
object (where statting /path/to/existing/00/0bjectfile/00 fails).
Signed-off-by: Jonathon Mah <me@JonathonMah.com>
Helped-by: Kyle J. McKay <mackyle@gmail.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We feed a root "objdir" path to this iterator function,
which then copies the result into a strbuf, so that it can
repeatedly append the object sub-directories to it. Let's
make it easy for callers to just pass us a strbuf in the
first place.
We leave the original interface as a convenience for callers
who want to just pass a const string like the result of
get_object_directory().
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
MAC_OS_X_VERSION_MIN_REQUIRED may be defined by the builder to a
specific version in order to produce compatible binaries for a
particular system. Blindly defining it to MAC_OS_X_VERSION_10_6
is bad.
Additionally MAC_OS_X_VERSION_10_6 will not be defined on older
systems and should AvailabilityMacros.h be included on such as
system an error will result. However, using the explicit value
of 1060 (which is what MAC_OS_X_VERSION_10_6 is defined to) does
not solve the problem.
The changes that introduced stepping on MAC_OS_X_VERSION_MIN were
made in b195aa00 (git-compat-util: suppress unavoidable
Apple-specific deprecation warnings) to avoid deprecation
warnings.
Instead of blindly setting MAC_OS_X_VERSION_MIN to 1060 change
the definition of DEPRECATED_ATTRIBUTE to empty to avoid the
warnings. This preserves any MAC_OS_X_VERSION_MIN_REQUIRED
setting while avoiding the warnings as intended by b195aa00.
Signed-off-by: Kyle J. McKay <mackyle@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Our config code simulates a stdio stream around a buffer,
but our fake ungetc() does not behave quite like the real
one. In particular, we only rewind the position by one
character, but do _not_ actually put the character from the
caller into position.
It turns out that this does not matter, because we only ever
push back the character we just read. In other words, such
an assignment would be a noop. But because the function is
called ungetc, and because it takes a character parameter,
it is a mistake waiting to happen.
Actually assigning the character into the buffer would be
ideal, but our pointer is actually a "const" copy of the
buffer. We do not know who the real owner of the buffer is
in this code, and would not want to munge their contents.
Instead, we can simply add an assertion that matches what
the current caller does, and will let us know if new callers
are added that violate the contract.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The decimal_width function originally appeared in blame.c as
"lineno_width", and was designed for calculating the
print-width of small-ish integer values (line numbers in
text files). In ec7ff5b, it was made into a reusable
function, and in dc801e7, we started using it to align
diffstats.
Binary files in a diffstat show byte counts rather than line
numbers, meaning they can be quite large (e.g., consider
adding or removing a 2GB file). decimal_width is not up to
the challenge for two reasons:
1. It takes the value as an "int", whereas large files may
easily surpass this. The value may be truncated, in
which case we will produce an incorrect value.
2. It counts "up" by repeatedly multiplying another
integer by 10 until it surpasses the value. This can
cause an infinite loop when the value is close to the
largest representable integer.
For example, consider using a 32-bit signed integer,
and a value of 2,140,000,000 (just shy of 2^31-1).
We will count up and eventually see that 1,000,000,000
is smaller than our value. The next step would be to
multiply by 10 and see that 10,000,000,000 is too
large, ending the loop. But we can't represent that
value, and we have signed overflow.
This is technically undefined behavior, but a common
behavior is to lose the high bits, in which case our
iterator will certainly be less than the number. So
we'll keep multiplying, overflow again, and so on.
This patch changes the argument to a uintmax_t (the same
type we use to store the diffstat information for binary
filese), and counts "down" by repeatedly dividing our value
by 10.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we are parsing a config value, if we see a carriage
return, we fgetc the next character to see if it is a
line feed (in which case we silently drop the CR). If it
isn't, we then ungetc the character, and take the literal
CR.
But we never check whether we in fact got a character at
all. If the config file ends in CR, we will get EOF here,
and try to ungetc EOF. This works OK for a real stdio
stream. The ungetc returns an error, and the next fgetc will
then return EOF again.
However, our custom buffer-based stream is not so fortunate.
It happily rewinds the position of the stream by one
character, ignoring the fact that we fed it EOF. The next
fgetc call returns the final CR again, over and over, and we
end up in an infinite loop.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The __builtin_ctzll function was added in gcc 3.4.0.
This extends the check for gcc so that use of __builtin_ctzll is only
enabled if gcc >= 3.4.0.
Signed-off-by: Tom G. Christensen <tgc@statsbiblioteket.dk>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
curl 7.11.0 through 7.12.2 when built from their official release
archives will present a 5 digit version number instead of the documented
6 digits which breaks the version check in the Makefile.
Correct these broken version numbers on the fly when extracting them to
ensure the comparison works correctly.
[jc: shortened the new sed scripts a bit]
Signed-off-by: Tom G. Christensen <tgc@statsbiblioteket.dk>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we add a new submodule the path of the submodule is being
normalized. We fail to normalize multiple adjacent '/./', though.
Thus 'path/to/././submodule' will become 'path/to/./submodule' where
it should be 'path/to/submodule' instead.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Acked-by: Jens Lehmann <Jens.Lehmann@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We may want to say something about command line option names in the
new section as well, but for now, let's make sure everybody is clear
on how to structure and name their configuration variables.
The text for the rules are partly taken from the log message of
Jonathan's 6b3020a2 (add: introduce add.ignoreerrors synonym for
add.ignore-errors, 2010-12-01).
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It seems to be a common mistake to try using a single remote
(e.g. 'origin') to fetch from one place (i.e. upstream) while
pushing to another (i.e. your publishing point).
That will never work satisfactorily, and it is easy to understand
why if you think about what refs/remotes/origin/* would mean in such
a world. It fundamentally cannot reflect the reality. If it
follows the state of your upstream, it cannot match what you have
published, and vice versa.
It may be that misinformation is spread by some people. Let's
counter them by adding a few words to our documentation.
- The description was referring to <oldurl> and <newurl>, but never
mentioned <name> argument you give from the command line. By
mentioning "remote <name>", stress the fact that it is configuring
a single remote.
- Add a reminder that explicitly states that this is about a single
remote, which the triangular workflow is not about.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some older versions of gpg (reportedly v1.2.6 from RHEL4) cannot
import the keyrings found in our test suite, and thus cannot even
make a signature. The previous change works it around, but we
cannot anticipate breakages update to GPG would cause in the future.
Do a test-sign before declaring the GPG prerequisite fulfilled
to future-proof our tests.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since 1e3eefb (tests: replace binary GPG keyrings with
ASCII-armored keys, 2014-12-12), we import our test GPG keys
from a single file. Each keypair in the import stream
contains both the secret and public keys. However, older
versions of gpg reportedly fail to import the public half of
the key. We can solve this by including duplicates of the
public keys separately. The duplicates are ignored by modern
gpg, and this makes older versions work.
Reported by Tom G. Christensen <tgc@statsbiblioteket.dk> on
gpg 1.2.6 (from RHEL4).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
b6d8f309 (diff-raw format update take #2., 2005-05-23) started
documenting the diff format, and it said
...
(8) sha1 for "dst"; 0{40} if creation, unmerged or "look at work tree".
(9) status, followed by similarlity index number only for C and R.
(10) a tab or a NUL when '-z' option is used.
...
because C and R _were_ the only ones that came with a number back
then. This was corrected by ddafa7e9 (diff-helper: Fix R/C score
parsing under -z flag., 2005-05-29) and we started saying "score"
instead of "similarlity index" (because we can have other kind of
score there), and stopped saying "only for C and R" (because Git is
an ever evolving system). Later f345b0a0 (Add -B flag to diff-*
brothers., 2005-05-30) introduced a new concept, "dissimilarity"
score; it did not have to fix any documentation.
The current text that says only C and R can have scores came
independently from a5a323f3 (Add reference for status letters in
documentation., 2008-11-02) and it was wrong from the day one.
Noticed-by: Mike Hommey
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As per the code, the --repo <repo> option is equivalent to the
<repo> argument to 'git push', but somehow it was documented as
something that is more than that. [It exists for historical
reasons, back from the time when options had to come before
arguments.]
Say so. [But not that.]
Signed-off-by: Michael J Gruber <git@drmicha.warpmail.net>
Helped-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There is no point in checking "!ref->name" when ref is a
"struct ref". The name field is a flex-array, and there
always has a non-zero address. This is almost certainly not
hurting anything, but it does cause clang-3.6 to complain.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we are chomping newlines from the end of a strbuf, we
must check "sb.len != 0" before accessing "sb.buf[sb.len - 1]".
However, this code mistakenly checks "&sb.len", which is
always true (it is a part of an auto struct, so the address
is always non-zero). This could lead to us accessing memory
outside the strbuf when we read an empty file.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The old text gave an impression that even in a new repository using
old form might be safer. Only Git from pre 1.7.0 days choke on the
correctly named variable, which is ancient by today's standard.
We have no intention to remove the support for deprecated ones, but
let's make sure that we do not give room for confused questions such
as "why does core.sparse-checkout not work, when add.ignore-errors
does?"
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Once upon a time, dumb http always fetched .idx files
directly into their final location, and then checked their
validity with parse_pack_index. This was refactored in
commit 750ef42 (http-fetch: Use temporary files for
pack-*.idx until verified, 2010-04-19), which uses the
following logic:
1. If we have the idx already in place, see if it's
valid (using parse_pack_index). If so, use it.
2. Otherwise, fetch the .idx to a tempfile, check
that, and if so move it into place.
3. Either way, fetch the pack itself if necessary.
However, it got step 1 wrong. We pass a NULL path parameter
to parse_pack_index, so an existing .idx file always looks
broken. Worse, we do not treat this broken .idx as an
opportunity to re-fetch, but instead return an error,
ignoring the pack entirely. This can lead to a dumb-http
fetch failing to retrieve the necessary objects.
This doesn't come up much in practice, because it must be a
packfile that we found out about (and whose .idx we stored)
during an earlier dumb-http fetch, but whose packfile we
_didn't_ fetch. I.e., we did a partial clone of a
repository, didn't need some packfiles, and now a followup
fetch needs them.
Discovery and tests by Charles Bailey <charles@hashpling.org>.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
