The stuff_change() function makes diff_filespecs out of
blobs. The term we generally use for filespecs is "path",
not "name", so let's be consistent here. That will make
things less confusing when the next patch starts caring
about the path/name distinction inside the pending object
array.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When diffing blobs directly, git-diff picks the blobs out of
the rev_info's pending array and copies the relevant bits to
a custom "struct blobinfo". But the pending array entry
already has all of this information (and more, which we'll
use in future patches). Let's just pass the original entry
instead.
In practice, these two blobs are probably adjacent in the
revs->pending array, and we could just pass the whole array.
But the current code is careful to pick each blob out
separately and put it into another array, so we'll continue
to do so and make our own array-of-pointers.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If the revision parser sees an argument like tree:path, we
parse it down to the correct blob (or tree), but throw away
the "path" portion. Let's ask get_sha1_with_context() to
record it, and pass it along in the pending array.
This will let programs like git-diff which rely on the
revision-parser show more accurate paths.
Note that the implementation is a little tricky; we have to
make sure we free oc.path in all code paths. For handle_dotdot(),
we can piggy-back on the existing cleanup-wrapper pattern.
The real work happens in handle_dotdot_1(), but the
handle_dotdot() wrapper makes sure that the path is freed no
matter how we exit the function (and for that reason we make
sure that the object_context struct is zero'd, so if we fail
to even get to the get_sha1_with_context() call, we just end
up calling free(NULL)).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The "a..b" revision syntax was designed to handle commits,
so it doesn't bother to record any mode we find while
traversing a "tree:path" endpoint. These days "git diff" can
diff blobs using either "a:path..b:path" (with dots) or
"a:path b:path" (without), but the two behave
inconsistently, as the with-dots version fails to notice the
mode.
Let's teach the dot-dot range parser to record modes; it
doesn't cost us anything, and it makes this case work.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The git-diff command can directly compare two blobs (or a
blob and a file), but we don't test this at all. Let's add
some basic tests that reveal a few problems.
There are basically four interesting inputs:
1. sha1 against sha1 (where diff has no information beyond
the contents)
2. tree:path against tree:path (where it can get
information via get_sha1_with_context)
3. Same as (2), but using the ".." range syntax
4. tree:path against a filename
And beyond generating a sane diff, we care about a few
little bits: which paths they show in the diff header, and
whether they correctly pick up a mode change.
They should all be able to show a mode except for (1),
though note that case (3) is currently broken.
For the headers, we would ideally show the path within the
tree if we have it, making:
git diff a:path b:path
look the same as:
git diff a b -- path
We can't always do that (e.g., in the direct sha1/sha1 diff,
we have no path to show), in which case we should fall back
to the name that resolved to the blob (which is nonsense
from the repository's perspective, but is the best we can
do).
Aside from the fallback case in (1), none of the cases get
this right. Cases (2) and (3) always show the full
tree:path, even though we should be able to know just the
"path" portion.
Case (4) picks up the filename path, but assigns it to
_both_ sides of the diff. So this works for:
git diff tree:path path
but not for:
git diff tree:other_path path
The appropriate tests are marked to expect failure.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a sha1 lookup returns the tree path via "struct
object_context", it just copies it into a fixed-size buffer.
This means the result can be truncated, and it means our
"struct object_context" consumes a lot of stack space.
Instead, let's allocate a string on the heap. Because most
callers don't care about this information, we'll avoid doing
it by default (so they don't all have to start calling
free() on the result). There are basically two options for
the caller to signal to us that it's interested:
1. By setting a pointer to storage in the object_context.
2. By passing a flag in another parameter.
Doing (1) would match the way that sha1_object_info_extended()
works. But it would mean that every caller would have to
initialize the object_context, which they don't currently
have to do.
This patch does (2), and adds a new bit to the function's
flags field. All of the callers that look at the "path"
field are updated to pass the new flag.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The get_sha1_with_context() function zeroes out the
oc->symlink_path strbuf, but doesn't use strbuf_init() to
set up the usual invariants (like pointing to the slopbuf).
We don't actually write to the oc->symlink_path strbuf
unless we call get_tree_entry_follow_symlinks(), and that
function does initialize it. However, readers may still look
at the zero'd strbuf.
In practice this isn't a triggerable bug. The only caller
that looks at it only does so when the mode we found is 0.
This doesn't happen for non-tree-entries (where we return
S_IFINVALID). A broken tree entry could have a mode of 0,
but canon_mode() quietly rewrites that into S_IFGITLINK.
So the "0" mode should only come up when we did indeed find
a symlink.
This is mostly just an accident of how the code happens to
work, though. Let's future-proof ourselves to make sure the
strbuf is properly initialized for all calls (it's only a
few struct member assignments, not a heap allocation).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
An early version of the patch to add object_context used the
name object_resolve_context. This was later shortened to
just object_context, but the "orc" variable name stuck in a
few places. Let's use "oc", which is used elsewhere in the
code.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The handle_revision_arg function is rather long, and a big
chunk of it is handling the range operators. Let's pull that
out to a separate helper. While we're doing so, we can clean
up a few of the rough edges that made the flow hard to
follow:
- instead of manually restoring *dotdot (that we overwrote
with a NUL), do the real work in a sub-helper, which
makes it clear that the munge/restore lines are a
matched pair
- eliminate a goto which wasn't actually used for control
flow, but only to avoid duplicating a few lines
(instead, those lines are pushed into another helper
function)
- use early returns instead of deep nesting
- consistently name all variables for the left-hand side
of the range as "a" (rather than "this" or "from") and
the right-hand side as "b" (rather than "next", or using
the unadorned "sha1" or "flags" from the main function).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since 003c84f6d (specifying ranges: we did not mean to make
".." an empty set, 2011-05-02), we treat the argument ".."
specially. We detect it by noticing that both sides of the
range are empty, and that this is a non-symmetric two-dot
range.
While correct, this makes the code overly complicated. We
can just detect ".." up front before we try to do further
parsing. This avoids having to de-munge the NUL from dotdot,
and lets us eliminate an extra const array (which we needed
only to do direct pointer comparisons).
It also removes the one code path from the range-parsing
conditional that requires us to return -1. That will make it
simpler to pull the dotdot parsing out into its own
function.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The handle_revision_arg() function has a "dotdot" variable
that it uses to find a ".." or "..." in the argument. If we
don't find one, we look for other marks, like "^!". But we
just keep re-using the "dotdot" variable, which is
confusing.
Let's introduce a separate "mark" variable that can be used
for these other marks. They still reuse the same variable,
but at least the name is no longer actively misleading.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The "dotdot" range parser avoids calling
lookup_commit_reference() if we are directly fed two
commits. But its casts are unnecessarily complex; that
function will just return a commit we pass into it.
Just calling the function all the time is much simpler, and
doesn't do any significant extra work (the object is already
parsed, and deref_tag() on a non-tag is a noop; we do incur
one extra lookup_object() call, but that's fairly trivial).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we are parsing a range like "a..b", we write a
temporary NUL over the first ".", so that we can access the
names "a" and "b" as C strings. But our restoration of the
original "." is done at inconsistent times, which can lead
to confusing results.
For most calls, we restore the "." after we resolve the
names, but before we call verify_non_filename(). This means
that when we later call add_pending_object(), the name for
the left-hand "a" has been re-expanded to "a..b". You can
see this with:
git log --source a...b
where "b" will be correctly marked with "b", but "a" will be
marked with "a...b". Likewise with "a..b" (though you need
to use --boundary to even see "a" at all in that case).
To top off the confusion, when the REVARG_CANNOT_BE_FILENAME
flag is set, we skip the non-filename check, and leave the
NUL in place.
That means we do report the correct name for "a" in the
pending array. But some code paths try to show the whole
"a...b" name in error messages, and these erroneously show
only "a" instead of "a...b". E.g.:
$ git cherry-pick HEAD:foo...HEAD:foo
error: object d95f3ad14dee633a758d2e331151e950dd13e4ed is a blob, not a commit
error: object d95f3ad14dee633a758d2e331151e950dd13e4ed is a blob, not a commit
fatal: Invalid symmetric difference expression HEAD:foo
(That last message should be "HEAD:foo...HEAD:foo"; I used
cherry-pick because it passes the CANNOT_BE_FILENAME flag).
As an interesting side note, cherry-pick actually looks at
and re-resolves the arguments from the pending->name fields.
So it would have been visibly broken by the first bug, but
the effect was canceled out by the second one.
This patch makes the whole function consistent by re-writing
the NUL immediately after calling verify_non_filename(), and
then restoring the "." as appropriate in some error-printing
and early-return code paths.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a remote server uses git-shell, the client side will
connect to it like:
ssh server "git-upload-pack 'foo.git'"
and we literally exec ("git-upload-pack", "foo.git"). In
early versions of upload-pack and receive-pack, we took a
repository argument and nothing else. But over time they
learned to accept dashed options. If the user passes a
repository name that starts with a dash, the results are
confusing at best (we complain of a bogus option instead of
a non-existent repository) and malicious at worst (the user
can start an interactive pager via "--help").
We could pass "--" to the sub-process to make sure the
user's argument is interpreted as a branch name. I.e.:
git-upload-pack -- -foo.git
But adding "--" automatically would make us inconsistent
with a normal shell (i.e., when git-shell is not in use),
where "-foo.git" would still be an error. For that case, the
client would have to specify the "--", but they can't do so
reliably, as existing versions of git-shell do not allow
more than a single argument.
The simplest thing is to simply disallow "-" at the start of
the repo name argument. This hasn't worked either with or
without git-shell since version 1.0.0, and nobody has
complained.
Note that this patch just applies to do_generic_cmd(), which
runs upload-pack, receive-pack, and upload-archive. There
are two other types of commands that git-shell runs:
- do_cvs_cmd(), but this already restricts the argument to
be the literal string "server"
- admin-provided commands in the git-shell-commands
directory. We'll pass along arbitrary arguments there,
so these commands could have similar problems. But these
commands might actually understand dashed arguments, so
we cannot just block them here. It's up to the writer of
the commands to make sure they are safe. With great
power comes great responsibility.
Reported-by: Timo Schmid <tschmid@ernw.de>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
