Interactive rebase uses 'git cherry-pick' and 'git merge' to replay
commits. Both invoke the 'rerere' machinery when they fail due to merge
conflicts. Note that all code paths with these two commands also invoke
the shell function die_with_patch when the commands fail.

Since commit 629716d2 ("rerere: do use multiple variants") the second
operation of the rerere machinery can be observed by a duplicated
message "Recorded preimage for 'file'". This second operation records
the same preimage as the first one and, hence, only wastes cycles.
Remove the 'git rerere' invocation from die_with_patch.

Shell function die_with_patch can be called after the failure of
"git commit", too, which also calls into the rerere machinery, but it
does so only after a successful commit to record the resolution.
Therefore, it is wrong to call 'git rerere' from die_with_patch after
"git commit" fails.
Signed-off-by: NJohannes Sixt <j6t@kdbg.org>
Acked-by: NJohannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Bugfix patches were backported from the 'master' front to plug heap
corruption holes, to catch integer overflow in the computation of
pathname lengths, and to get rid of the name_path API.  Both of
these would have resulted in writing over an under-allocated buffer
when formulating pathnames while tree traversal.

* jk/path-name-safety-2.4:
  list-objects: pass full pathname to callbacks
  list-objects: drop name_path entirely
  list-objects: convert name_path to a strbuf
  show_object_with_name: simplify by using path_name()
  http-push: stop using name_path
  tree-diff: catch integer overflow in combine_diff_path allocation
  add helpers for detecting size_t overflow

When we find a blob at "a/b/c", we currently pass this to
our show_object_fn callbacks as two components: "a/b/" and
"c". Callbacks which want the full value then call
path_name(), which concatenates the two. But this is an
inefficient interface; the path is a strbuf, and we could
simply append "c" to it temporarily, then roll back the
length, without creating a new copy.

So we could improve this by teaching the callsites of
path_name() this trick (and there are only 3). But we can
also notice that no callback actually cares about the
broken-down representation, and simply pass each callback
the full path "a/b/c" as a string. The callback code becomes
even simpler, then, as we do not have to worry about freeing
an allocated buffer, nor rolling back our modification to
the strbuf.

This is theoretically less efficient, as some callbacks
would not bother to format the final path component. But in
practice this is not measurable. Since we use the same
strbuf over and over, our work to grow it is amortized, and
we really only pay to memcpy a few bytes.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

In the previous commit, we left name_path as a thin wrapper
around a strbuf. This patch drops it entirely. As a result,
every show_object_fn callback needs to be adjusted. However,
none of their code needs to be changed at all, because the
only use was to pass it to path_name(), which now handles
the bare strbuf.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

The "struct name_path" data is examined in only two places:
we generate it in process_tree(), and we convert it to a
single string in path_name(). Everyone else just passes it
through to those functions.

We can further note that process_tree() already keeps a
single strbuf with the leading tree path, for use with
tree_entry_interesting().

Instead of building a separate name_path linked list, let's
just use the one we already build in "base". This reduces
the amount of code (especially tricky code in path_name()
which did not check for integer overflows caused by deep
or large pathnames).

It is also more efficient in some instances.  Any time we
were using tree_entry_interesting, we were building up the
strbuf anyway, so this is an immediate and obvious win
there. In cases where we were not, we trade off storing
"pathname/" in a strbuf on the heap for each level of the
path, instead of two pointers and an int on the stack (with
one pointer into the tree object). On a 64-bit system, the
latter is 20 bytes; so if path components are less than that
on average, this has lower peak memory usage.  In practice
it probably doesn't matter either way; we are already
holding in memory all of the tree objects leading up to each
pathname, and for normal-depth pathnames, we are only
talking about hundreds of bytes.

This patch leaves "struct name_path" as a thin wrapper
around the strbuf, to avoid disrupting callbacks. We should
fix them, but leaving it out makes this diff easier to view.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

When "git rev-list" shows an object with its associated path
name, it does so by walking the name_path linked list and
printing each component (stopping at any embedded NULs or
newlines).

We'd like to eventually get rid of name_path entirely in
favor of a single buffer, and dropping this custom printing
code is part of that. As a first step, let's use path_name()
to format the list into a single buffer, and print that.
This is strictly less efficient than the original, but it's
a temporary step in the refactoring; our end game will be to
get the fully formatted name in the first place.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

The graph traversal code here passes along a name_path to
build up the pathname at which we find each blob. But we
never actually do anything with the resulting names, making
it a waste of code and memory.

This usage came in aa1dbc98 (Update http-push functionality,
2006-03-07), and originally the result was passed to
"add_object" (which stored it, but didn't really use it,
either). But we stopped using that function in 1f1e895f (Add
"named object array" concept, 2006-06-19) in favor of
storing just the objects themselves.

Moreover, the generation of the name in process_tree() is
buggy. It sticks "name" onto the end of the name_path linked
list, and then passes it down again as it recurses (instead
of "entry.path"). So it's a good thing this was unused, as
the resulting path for "a/b/c/d" would end up as "a/a/a/a".
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

A combine_diff_path struct has two "flex" members allocated
alongside the struct: a string to hold the pathname, and an
array of parent pointers. We use an "int" to compute this,
meaning we may easily overflow it if the pathname is
extremely long.

We can fix this by using size_t, and checking for overflow
with the st_add helper.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Performing computations on size_t variables that we feed to
xmalloc and friends can be dangerous, as an integer overflow
can cause us to allocate a much smaller chunk than we
realized.

We already have unsigned_add_overflows(), but let's add
unsigned_mult_overflows() to that. Furthermore, rather than
have each site manually check and die on overflow, we can
provide some helpers that will:

  - promote the arguments to size_t, so that we know we are
    doing our computation in the same size of integer that
    will ultimately be fed to xmalloc

  - check and die on overflow

  - return the result so that computations can be done in
    the parameter list of xmalloc.

These functions are a lot uglier to use than normal
arithmetic operators (you have to do "st_add(foo, bar)"
instead of "foo + bar"). To at least limit the damage, we
also provide multi-valued versions. So rather than:

  st_add(st_add(a, b), st_add(c, d));

you can write:

  st_add4(a, b, c, d);

This isn't nearly as elegant as a varargs function, but it's
a lot harder to get it wrong. You don't have to remember to
add a sentinel value at the end, and the compiler will
complain if you get the number of arguments wrong. This
patch adds only the numbered variants required to convert
the current code base; we can easily add more later if
needed.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Signed-off-by: NJunio C Hamano <gitster@pobox.com>

The previous commit enforces MAX_XDIFF_SIZE at the
interfaces to xdiff: xdi_diff (which calls xdl_diff) and
ll_xdl_merge (which calls xdl_merge).

But we have another direct call to xdl_merge in
merge-file.c. If it were written today, this probably would
just use the ll_merge machinery. But it predates that code,
and uses slightly different options to xdl_merge (e.g.,
ZEALOUS_ALNUM).

We could try to abstract out an xdi_merge to match the
existing xdi_diff, but even that is difficult. Rather than
simply report error, we try to treat large files as binary,
and that distinction would happen outside of xdi_merge.

The simplest fix is to just replicate the MAX_XDIFF_SIZE
check in merge-file.c.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

The xdiff code is not prepared to handle extremely large
files. It uses "int" in many places, which can overflow if
we have a very large number of lines or even bytes in our
input files. This can cause us to produce incorrect diffs,
with no indication that the output is wrong. Or worse, we
may even underallocate a buffer whose size is the result of
an overflowing addition.

We're much better off to tell the user that we cannot diff
or merge such a large file. This patch covers both cases,
but in slightly different ways:

  1. For merging, we notice the large file and cleanly fall
     back to a binary merge (which is effectively "we cannot
     merge this").

  2. For diffing, we make the binary/text distinction much
     earlier, and in many different places. For this case,
     we'll use the xdi_diff as our choke point, and reject
     any diff there before it hits the xdiff code.

     This means in most cases we'll die() immediately after.
     That's not ideal, but in practice we shouldn't
     generally hit this code path unless the user is trying
     to do something tricky. We already consider files
     larger than core.bigfilethreshold to be binary, so this
     code would only kick in when that is circumvented
     (either by bumping that value, or by using a
     .gitattribute to mark a file as diffable).

     In other words, we can avoid being "nice" here, because
     there is already nice code that tries to do the right
     thing. We are adding the suspenders to the nice code's
     belt, so notice when it has been worked around (both to
     protect the user from malicious inputs, and because it
     is better to die() than generate bogus output).

The maximum size was chosen after experimenting with feeding
large files to the xdiff code. It's just under a gigabyte,
which leaves room for two obvious cases:

  - a diff3 merge conflict result on files of maximum size X
    could be 3*X plus the size of the markers, which would
    still be only about 3G, which fits in a 32-bit int.

  - some of the diff code allocates arrays of one int per
    record. Even if each file consists only of blank lines,
    then a file smaller than 1G will have fewer than 1G
    records, and therefore the int array will fit in 4G.

Since the limit is arbitrary anyway, I chose to go under a
gigabyte, to leave a safety margin (e.g., we would not want
to overflow by allocating "(records + 1) * sizeof(int)" or
similar.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

When we call into xdiff to perform a diff, we generally lose
the return code completely. Typically by ignoring the return
of our xdi_diff wrapper, but sometimes we even propagate
that return value up and then ignore it later.  This can
lead to us silently producing incorrect diffs (e.g., "git
log" might produce no output at all, not even a diff header,
for a content-level diff).

In practice this does not happen very often, because the
typical reason for xdiff to report failure is that it
malloc() failed (it uses straight malloc, and not our
xmalloc wrapper).  But it could also happen when xdiff
triggers one our callbacks, which returns an error (e.g.,
outf() in builtin/rerere.c tries to report a write failure
in this way). And the next patch also plans to add more
failure modes.

Let's notice an error return from xdiff and react
appropriately. In most of the diff.c code, we can simply
die(), which matches the surrounding code (e.g., that is
what we do if we fail to load a file for diffing in the
first place). This is not that elegant, but we are probably
better off dying to let the user know there was a problem,
rather than simply generating bogus output.

We could also just die() directly in xdi_diff, but the
callers typically have a bit more context, and can provide a
better message (and if we do later decide to pass errors up,
we're one step closer to doing so).

There is one interesting case, which is in diff_grep(). Here
if we cannot generate the diff, there is nothing to match,
and we silently return "no hits". This is actually what the
existing code does already, but we make it a little more
explicit.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

By default, libcurl will follow circular http redirects
forever. Let's put a cap on this so that somebody who can
trigger an automated fetch of an arbitrary repository (e.g.,
for CI) cannot convince git to loop infinitely.

The value chosen is 20, which is the same default that
Firefox uses.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Previously, libcurl would follow redirection to any protocol
it was compiled for support with. This is desirable to allow
redirection from HTTP to HTTPS. However, it would even
successfully allow redirection from HTTP to SFTP, a protocol
that git does not otherwise support at all. Furthermore
git's new protocol-whitelisting could be bypassed by
following a redirect within the remote helper, as it was
only enforced at transport selection time.

This patch limits redirects within libcurl to HTTP, HTTPS,
FTP and FTPS. If there is a protocol-whitelist present, this
list is limited to those also allowed by the whitelist. As
redirection happens from within libcurl, it is impossible
for an HTTP redirect to a protocol implemented within
another remote helper.

When the curl version git was compiled with is too old to
support restrictions on protocol redirection, we warn the
user if GIT_ALLOW_PROTOCOL restrictions were requested. This
is a little inaccurate, as even without that variable in the
environment, we would still restrict SFTP, etc, and we do
not warn in that case. But anything else means we would
literally warn every time git accesses an http remote.

This commit includes a test, but it is not as robust as we
would hope. It redirects an http request to ftp, and checks
that curl complained about the protocol, which means that we
are relying on curl's specific error message to know what
happened. Ideally we would redirect to a working ftp server
and confirm that we can clone without protocol restrictions,
and not with them. But we do not have a portable way of
providing an ftp server, nor any other protocol that curl
supports (https is the closest, but we would have to deal
with certificates).

[jk: added test and version warning]
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

The current callers only want to die when their transport is
prohibited. But future callers want to query the mechanism
without dying.

Let's break out a few query functions, and also save the
results in a static list so we don't have to re-parse for
each query.
Based-on-a-patch-by: NBlake Burkhart <bburky@bburky.com>
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Some protocols (like git-remote-ext) can execute arbitrary
code found in the URL. The URLs that submodules use may come
from arbitrary sources (e.g., .gitmodules files in a remote
repository). Let's restrict submodules to fetching from a
known-good subset of protocols.

Note that we apply this restriction to all submodule
commands, whether the URL comes from .gitmodules or not.
This is more restrictive than we need to be; for example, in
the tests we run:

  git submodule add ext::...

which should be trusted, as the URL comes directly from the
command line provided by the user. But doing it this way is
simpler, and makes it much less likely that we would miss a
case. And since such protocols should be an exception
(especially because nobody who clones from them will be able
to update the submodules!), it's not likely to inconvenience
anyone in practice.
Reported-by: NBlake Burkhart <bburky@bburky.com>
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

If we are cloning an untrusted remote repository into a
sandbox, we may also want to fetch remote submodules in
order to get the complete view as intended by the other
side. However, that opens us up to attacks where a malicious
user gets us to clone something they would not otherwise
have access to (this is not necessarily a problem by itself,
but we may then act on the cloned contents in a way that
exposes them to the attacker).

Ideally such a setup would sandbox git entirely away from
high-value items, but this is not always practical or easy
to set up (e.g., OS network controls may block multiple
protocols, and we would want to enable some but not others).

We can help this case by providing a way to restrict
particular protocols. We use a whitelist in the environment.
This is more annoying to set up than a blacklist, but
defaults to safety if the set of protocols git supports
grows). If no whitelist is specified, we continue to default
to allowing all protocols (this is an "unsafe" default, but
since the minority of users will want this sandboxing
effect, it is the only sensible one).

A note on the tests: ideally these would all be in a single
test file, but the git-daemon and httpd test infrastructure
is an all-or-nothing proposition rather than a test-by-test
prerequisite. By putting them all together, we would be
unable to test the file-local code on machines without
apache.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Signed-off-by: NJunio C Hamano <gitster@pobox.com>

When we show "branch@{0}", we format into a fixed-size
buffer using sprintf. This can overflow if you have long
branch names. We can fix it by using a temporary strbuf.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

This function assumes that the relative_base path passed
into it is no larger than PATH_MAX, and writes into a
fixed-size buffer. However, this path may not have actually
come from the filesystem; for example, add_submodule_odb
generates a path using a strbuf and passes it in. This is
hard to trigger in practice, though, because the long
submodule directory would have to exist on disk before we
would try to open its info/alternates file.

We can easily avoid the bug, though, by simply creating the
filename on the heap.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

When we are loading a notes tree into our internal hash
table, we also collect any files that are clearly non-notes.
We format the name of the file into a PATH_MAX buffer, but
unlike true notes (which cannot be larger than a fanned-out
sha1 hash), these tree entries can be arbitrarily long,
overflowing our buffer.

We can fix this by switching to a strbuf. It doesn't even
cost us an extra allocation, as we can simply hand ownership
of the buffer over to the non-note struct.

This is of moderate security interest, as you might fetch
notes trees from an untrusted remote. However, we do not do
so by default, so you would have to manually fetch into the
notes namespace.
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

When unpack-trees wants to know whether a path will
overwrite anything in the working tree, we use lstat() to
see if there is anything there. But if we are going to write
"foo/bar", we can't just lstat("foo/bar"); we need to look
for leading prefixes (e.g., "foo"). So we use the lstat cache
to find the length of the leading prefix, and copy the
filename up to that length into a temporary buffer (since
the original name is const, we cannot just stick a NUL in
it).

The copy we make goes into a PATH_MAX-sized buffer, which
will overflow if the prefix is longer than PATH_MAX. How
this happens is a little tricky, since in theory PATH_MAX is
the biggest path we will have read from the filesystem. But
this can happen if:

  - the compiled-in PATH_MAX does not accurately reflect
    what the filesystem is capable of

  - the leading prefix is not _quite_ what is on disk; it
    contains the next element from the name we are checking.
    So if we want to write "aaa/bbb/ccc/ddd" and "aaa/bbb"
    exists, the prefix of interest is "aaa/bbb/ccc". If
    "aaa/bbb" approaches PATH_MAX, then "ccc" can overflow
    it.

So this can be triggered, but it's hard to do. In
particular, you cannot just "git clone" a bogus repo. The
verify_absent checks happen before unpack-trees writes
anything to the filesystem, so there are never any leading
prefixes during the initial checkout, and the bug doesn't
trigger. And by definition, these files are larger than
PATH_MAX, so writing them will fail, and clone will
complain (though it may write a partial path, which will
cause a subsequent "git checkout" to hit the bug).

We can fix it by creating the temporary path on the heap.
The extra malloc overhead is not important, as we are
already making at least one stat() call (and probably more
for the prefix discovery).
Signed-off-by: NJeff King <peff@peff.net>
Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Signed-off-by: NJunio C Hamano <gitster@pobox.com>

Abandoning an already applied change in "git rebase -i" with
"--continue" left CHERRY_PICK_HEAD and confused later steps.

* js/rebase-i-clean-up-upon-continue-to-skip:
  rebase -i: do not leave a CHERRY_PICK_HEAD file behind
  t3404: demonstrate CHERRY_PICK_HEAD bug

Code simplification.

* ss/clone-guess-dir-name-simplify:
  clone: simplify string handling in guess_dir_name()

* sg/completion-commit-cleanup:
  completion: teach 'scissors' mode to 'git commit --cleanup='

Various fixes around "git am" that applies a patch to a history
that is not there yet.

* pt/am-abort-fix:
  am --abort: keep unrelated commits on unborn branch
  am --abort: support aborting to unborn branch
  am --abort: revert changes introduced by failed 3way merge
  am --skip: support skipping while on unborn branch
  am -3: support 3way merge on unborn branch
  am --skip: revert changes introduced by failed 3way merge

"git for-each-ref" reported "missing object" for 0{40} when it
encounters a broken ref.  The lack of object whose name is 0{40} is
not the problem; the ref being broken is.

* mh/reporting-broken-refs-from-for-each-ref:
  read_loose_refs(): treat NULL_SHA1 loose references as broken
  read_loose_refs(): simplify function logic
  for-each-ref: report broken references correctly
  t6301: new tests of for-each-ref error handling