When http-backend spawns "upload-pack" to do ref
negotiation, it streams the http request body to
upload-pack, who then streams the http response back to the
client as it reads. In theory, git can go full-duplex; the
client can consume our response while it is still sending
the request. In practice, however, HTTP is a half-duplex
protocol. Even if our client is ready to read and write
simultaneously, we may have other HTTP infrastructure in the
way, including the webserver that spawns our CGI, or any
intermediate proxies.
In at least one documented case[1], this leads to deadlock
when trying a fetch over http. What happens is basically:
1. Apache proxies the request to the CGI, http-backend.
2. http-backend gzip-inflates the data and sends
the result to upload-pack.
3. upload-pack acts on the data and generates output over
the pipe back to Apache. Apache isn't reading because
it's busy writing (step 1).
This works fine most of the time, because the upload-pack
output ends up in a system pipe buffer, and Apache reads
it as soon as it finishes writing. But if both the request
and the response exceed the system pipe buffer size, then we
deadlock (Apache blocks writing to http-backend,
http-backend blocks writing to upload-pack, and upload-pack
blocks writing to Apache).
We need to break the deadlock by spooling either the input
or the output. In this case, it's ideal to spool the input,
because Apache does not start reading either stdout _or_
stderr until we have consumed all of the input. So until we
do so, we cannot even get an error message out to the
client.
The solution is fairly straight-forward: we read the request
body into an in-memory buffer in http-backend, freeing up
Apache, and then feed the data ourselves to upload-pack. But
there are a few important things to note:
1. We limit the in-memory buffer to prevent an obvious
denial-of-service attack. This is a new hard limit on
requests, but it's unlikely to come into play. The
default value is 10MB, which covers even the ridiculous
100,000-ref negotation in the included test (that
actually caps out just over 5MB). But it's configurable
on the off chance that you don't mind spending some
extra memory to make even ridiculous requests work.
2. We must take care only to buffer when we have to. For
pushes, the incoming packfile may be of arbitrary
size, and we should connect the input directly to
receive-pack. There's no deadlock problem here, though,
because we do not produce any output until the whole
packfile has been read.
For upload-pack's initial ref advertisement, we
similarly do not need to buffer. Even though we may
generate a lot of output, there is no request body at
all (i.e., it is a GET, not a POST).
[1] http://article.gmane.org/gmane.comp.version-control.git/269020
Test-adapted-from: Dennis Kaarsemaker <dennis@kaarsemaker.net>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
One of our tests in t5551 creates a large number of tags,
and jumps through some hoops to do it efficiently. Let's
factor that out into a function so we can make other similar
tests.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we die() in http-backend, we call a custom handler that
writes an HTTP 500 response to stdout, then reports the
error to stderr. Our routines for writing out the HTTP
response may themselves die, leading to us entering die()
again.
When it was originally written, that was OK; our custom
handler keeps a variable to notice this and does not
recurse. However, since cd163d4 (usage.c: detect recursion
in die routines and bail out immediately, 2012-11-14), the
main die() implementation detects recursion before we even
get to our custom handler, and bails without printing
anything useful.
We can handle this case by doing two things:
1. Installing a custom die_is_recursing handler that
allows us to enter up to one level of recursion. Only
the first call to our custom handler will try to write
out the error response. So if we die again, that is OK.
If we end up dying more than that, it is a sign that we
are in an infinite recursion.
2. Reporting the error to stderr before trying to write
out the HTTP response. In the current code, if we do
die() trying to write out the response, we'll exit
immediately from this second die(), and never get a
chance to output the original error (which is almost
certainly the more interesting one; the second die is
just going to be along the lines of "I tried to write
to stdout but it was closed").
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* jk/read-packed-refs-without-path-max:
read_packed_refs: use skip_prefix instead of static array
read_packed_refs: pass strbuf to parse_ref_line
read_packed_refs: use a strbuf for reading lines
Our is_hfs_dotgit function relies on the hackily-implemented
next_hfs_char to give us the next character that an HFS+
filename comparison would look at. It's hacky because it
doesn't implement the full case-folding table of HFS+; it
gives us just enough to see if the path matches ".git".
At the end of next_hfs_char, we use tolower() to convert our
32-bit code point to lowercase. Our tolower() implementation
only takes an 8-bit char, though; it throws away the upper
24 bits. This means we can't have any false negatives for
is_hfs_dotgit. We only care about matching 7-bit ASCII
characters in ".git", and we will correctly process 'G' or
'g'.
However, we _can_ have false positives. Because we throw
away the upper bits, code point \u{0147} (for example) will
look like 'G' and get downcased to 'g'. It's not known
whether a sequence of code points whose truncation ends up
as ".git" is meaningful in any language, but it does not
hurt to be more accurate here. We can just pass out the full
32-bit code point, and compare it manually to the upper and
lowercase characters we care about.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The build procedure did not bother fixing perl and python scripts
when NO_PERL and NO_PYTHON build-time configuration changed.
* jk/rebuild-perl-scripts-with-no-perl-seting-change:
Makefile: have python scripts depend on NO_PYTHON setting
Makefile: simplify by using SCRIPT_{PERL,SH}_GEN macros
Makefile: have perl scripts depend on NO_PERL setting
Some tests that depend on perl lacked PERL prerequisite to protect
them, breaking build with NO_PERL configuration.
* jk/no-perl-tests:
t960[34]: mark cvsimport tests as requiring perl
t0090: mark add-interactive test with PERL prerequisite
Git 2.0 was supposed to make the "simple" mode for the default of
"git push", but it didn't.
* jk/push-simple:
push: truly use "simple" as default, not "upstream"
"git init" (hence "git clone") initialized the per-repository
configuration file .git/config with x-bit by mistake.
* mh/config-flip-xbit-back-after-checking:
create_default_files(): don't set u+x bit on $GIT_DIR/config
"gitweb" used to depend on a behaviour that was deprecated by recent
CGI.pm.
* jk/gitweb-with-newer-cgi-multi-param:
gitweb: hack around CGI's list-context param() handling
open() emulated on Windows platforms did not give EISDIR upon an
attempt to open a directory for writing.
* js/windows-open-eisdir-error:
Windows: correct detection of EISDIR in mingw_open()
"git config --get-color" did not parse its command line arguments
carefully.
* jk/colors-fix:
t4026: test "normal" color
config: fix parsing of "git config --get-color some.key -1"
docs: describe ANSI 256-color mode
"git checkout $treeish $path", when $path in the index and the
working tree already matched what is in $treeish at the $path,
still overwrote the $path unnecessarily.
* jk/checkout-from-tree:
checkout $tree: do not throw away unchanged index entries
Now that the index can block pathnames that can be mistaken
to mean ".git" on NTFS and FAT32, it would be helpful for
fsck to notice such problematic paths. This lets servers
which use receive.fsckObjects block them before the damage
spreads.
Note that the fsck check is always on, even for systems
without core.protectNTFS set. This is technically more
restrictive than we need to be, as a set of users on ext4
could happily use these odd filenames without caring about
NTFS.
However, on balance, it's helpful for all servers to block
these (because the paths can be used for mischief, and
servers which bother to fsck would want to stop the spread
whether they are on NTFS themselves or not), and hardly
anybody will be affected (because the blocked names are
variants of .git or git~1, meaning mischief is almost
certainly what the tree author had in mind).
Ideally these would be controlled by a separate
"fsck.protectNTFS" flag. However, it would be much nicer to
be able to enable/disable _any_ fsck flag individually, and
any scheme we choose should match such a system. Given the
likelihood of anybody using such a path in practice, it is
not unreasonable to wait until such a system materializes.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The point of disallowing ".git" in the index is that we
would never want to accidentally overwrite files in the
repository directory. But this means we need to respect the
filesystem's idea of when two paths are equal. The prior
commit added a helper to make such a comparison for NTFS
and FAT32; let's use it in verify_path().
We make this check optional for two reasons:
1. It restricts the set of allowable filenames, which is
unnecessary for people who are not on NTFS nor FAT32.
In practice this probably doesn't matter, though, as
the restricted names are rather obscure and almost
certainly would never come up in practice.
2. It has a minor performance penalty for every path we
insert into the index.
This patch ties the check to the core.protectNTFS config
option. Though this is expected to be most useful on Windows,
we allow it to be set everywhere, as NTFS may be mounted on
other platforms. The variable does default to on for Windows,
though.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We do not allow paths with a ".git" component to be added to
the index, as that would mean repository contents could
overwrite our repository files. However, asking "is this
path the same as .git" is not as simple as strcmp() on some
filesystems.
On NTFS (and FAT32), there exist so-called "short names" for
backwards-compatibility: 8.3 compliant names that refer to the same files
as their long names. As ".git" is not an 8.3 compliant name, a short name
is generated automatically, typically "git~1".
Depending on the Windows version, any combination of trailing spaces and
periods are ignored, too, so that both "git~1." and ".git." still refer
to the Git directory. The reason is that 8.3 stores file names shorter
than 8 characters with trailing spaces. So literally, it does not matter
for the short name whether it is padded with spaces or whether it is
shorter than 8 characters, it is considered to be the exact same.
The period is the separator between file name and file extension, and
again, an empty extension consists just of spaces in 8.3 format. So
technically, we would need only take care of the equivalent of this
regex:
(\.git {0,4}|git~1 {0,3})\. {0,3}
However, there are indications that at least some Windows versions might
be more lenient and accept arbitrary combinations of trailing spaces and
periods and strip them out. So we're playing it real safe here. Besides,
there can be little doubt about the intention behind using file names
matching even the more lenient pattern specified above, therefore we
should be fine with disallowing such patterns.
Extra care is taken to catch names such as '.\\.git\\booh' because the
backslash is marked as a directory separator only on Windows, and we want
to use this new helper function also in fsck on other platforms.
A big thank you goes to Ed Thomson and an unnamed Microsoft engineer for
the detailed analysis performed to come up with the corresponding fixes
for libgit2.
This commit adds a function to detect whether a given file name can refer
to the Git directory by mistake.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that the index can block pathnames that case-fold to
".git" on HFS+, it would be helpful for fsck to notice such
problematic paths. This lets servers which use
receive.fsckObjects block them before the damage spreads.
Note that the fsck check is always on, even for systems
without core.protectHFS set. This is technically more
restrictive than we need to be, as a set of users on ext4
could happily use these odd filenames without caring about
HFS+.
However, on balance, it's helpful for all servers to block
these (because the paths can be used for mischief, and
servers which bother to fsck would want to stop the spread
whether they are on HFS+ themselves or not), and hardly
anybody will be affected (because the blocked names are
variants of .git with invisible Unicode code-points mixed
in, meaning mischief is almost certainly what the tree
author had in mind).
Ideally these would be controlled by a separate
"fsck.protectHFS" flag. However, it would be much nicer to
be able to enable/disable _any_ fsck flag individually, and
any scheme we choose should match such a system. Given the
likelihood of anybody using such a path in practice, it is
not unreasonable to wait until such a system materializes.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The point of disallowing ".git" in the index is that we
would never want to accidentally overwrite files in the
repository directory. But this means we need to respect the
filesystem's idea of when two paths are equal. The prior
commit added a helper to make such a comparison for HFS+;
let's use it in verify_path.
We make this check optional for two reasons:
1. It restricts the set of allowable filenames, which is
unnecessary for people who are not on HFS+. In practice
this probably doesn't matter, though, as the restricted
names are rather obscure and almost certainly would
never come up in practice.
2. It has a minor performance penalty for every path we
insert into the index.
This patch ties the check to the core.protectHFS config
option. Though this is expected to be most useful on OS X,
we allow it to be set everywhere, as HFS+ may be mounted on
other platforms. The variable does default to on for OS X,
though.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We do not allow paths with a ".git" component to be added to
the index, as that would mean repository contents could
overwrite our repository files. However, asking "is this
path the same as .git" is not as simple as strcmp() on some
filesystems.
HFS+'s case-folding does more than just fold uppercase into
lowercase (which we already handle with strcasecmp). It may
also skip past certain "ignored" Unicode code points, so
that (for example) ".gi\u200ct" is mapped ot ".git".
The full list of folds can be found in the tables at:
https://www.opensource.apple.com/source/xnu/xnu-1504.15.3/bsd/hfs/hfscommon/Unicode/UCStringCompareData.h
Implementing a full "is this path the same as that path"
comparison would require us importing the whole set of
tables. However, what we want to do is much simpler: we
only care about checking ".git". We know that 'G' is the
only thing that folds to 'g', and so on, so we really only
need to deal with the set of ignored code points, which is
much smaller.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We complain about ".git" in a tree because it cannot be
loaded into the index or checked out. Since we now also
reject ".GIT" case-insensitively, fsck should notice the
same, so that errors do not propagate.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We check that fsck notices and complains about confusing
paths in trees. However, there are a few shortcomings:
1. We check only for these paths as file entries, not as
intermediate paths (so ".git" and not ".git/foo").
2. We check "." and ".." together, so it is possible that
we notice only one and not the other.
3. We repeat a lot of boilerplate.
Let's use some loops to be more thorough in our testing, and
still end up with shorter code.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We do not allow ".git" to enter into the index as a path
component, because checking out the result to the working
tree may causes confusion for subsequent git commands.
However, on case-insensitive file systems, ".Git" or ".GIT"
is the same. We should catch and prevent those, too.
Note that technically we could allow this for repos on
case-sensitive filesystems. But there's not much point. It's
unlikely that anybody cares, and it creates a repository
that is unexpectedly non-portable to other systems.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We should prevent nonsense paths from entering the index in
the first place, as they can cause confusing results if they
are ever checked out into the working tree. We already do
so, but we never tested it.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When unpack_trees tries to write an entry to the index,
add_index_entry may report an error to stderr, but we ignore
its return value. This leads to us returning a successful
exit code for an operation that partially failed. Let's make
sure to propagate this code.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
