When the http server gives an incomplete response to a smart-http
rpc call, it could lead to client waiting for a full response that
will never come. Teach the client side to notice this condition
and abort the transfer.
An improvement counterproposal has failed.
cf. <20161114194049.mktpsvgdhex2f4zv@sigill.intra.peff.net>
* dt/smart-http-detect-server-going-away:
upload-pack: optionally allow fetching any sha1
remote-curl: don't hang when a server dies before any output
Transport with dumb http can be fooled into following foreign URLs
that the end user does not intend to, especially with the server
side redirects and http-alternates mechanism, which can lead to
security issues. Tighten the redirection and make it more obvious
to the end user when it happens.
* jk/http-walker-limit-redirect-2.9:
http: treat http-alternates like redirects
http: make redirects more obvious
remote-curl: rename shadowed options variable
http: always update the base URL for redirects
http: simplify update_url_from_redirect
If a malicious server redirects the initial ref
advertisement, it may be able to leak sha1s from other,
unrelated servers that the client has access to. For
example, imagine that Alice is a git user, she has access to
a private repository on a server hosted by Bob, and Mallory
runs a malicious server and wants to find out about Bob's
private repository.
Mallory asks Alice to clone an unrelated repository from her
over HTTP. When Alice's client contacts Mallory's server for
the initial ref advertisement, the server issues an HTTP
redirect for Bob's server. Alice contacts Bob's server and
gets the ref advertisement for the private repository. If
there is anything to fetch, she then follows up by asking
the server for one or more sha1 objects. But who is the
server?
If it is still Mallory's server, then Alice will leak the
existence of those sha1s to her.
Since commit c93c92f30 (http: update base URLs when we see
redirects, 2013-09-28), the client usually rewrites the base
URL such that all further requests will go to Bob's server.
But this is done by textually matching the URL. If we were
originally looking for "http://mallory/repo.git/info/refs",
and we got pointed at "http://bob/other.git/info/refs", then
we know that the right root is "http://bob/other.git".
If the redirect appears to change more than just the root,
we punt and continue to use the original server. E.g.,
imagine the redirect adds a URL component that Bob's server
will ignore, like "http://bob/other.git/info/refs?dummy=1".
We can solve this by aborting in this case rather than
silently continuing to use Mallory's server. In addition to
protecting from sha1 leakage, it's arguably safer and more
sane to refuse a confusing redirect like that in general.
For example, part of the motivation in c93c92f30 is
avoiding accidentally sending credentials over clear http,
just to get a response that says "try again over https". So
even in a non-malicious case, we'd prefer to err on the side
of caution.
The downside is that it's possible this will break a
legitimate but complicated server-side redirection scheme.
The setup given in the newly added test does work, but it's
convoluted enough that we don't need to care about it. A
more plausible case would be a server which redirects a
request for "info/refs?service=git-upload-pack" to just
"info/refs" (because it does not do smart HTTP, and for some
reason really dislikes query parameters). Right now we
would transparently downgrade to dumb-http, but with this
patch, we'd complain (and the user would have to set
GIT_SMART_HTTP=0 to fetch).
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It seems a little silly to do a reachabilty check in the case where we
trust the user to access absolutely everything in the repository.
Also, it's racy in a distributed system -- perhaps one server
advertises a ref, but another has since had a force-push to that ref,
and perhaps the two HTTP requests end up directed to these different
servers.
Signed-off-by: David Turner <dturner@twosigma.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In the event that a HTTP server closes the connection after giving a
200 but before giving any packets, we don't want to hang forever
waiting for a response that will never come. Instead, we should die
immediately.
One case where this happens is when attempting to fetch a dangling
object by its object name. In this case, the server dies before
sending any data. Prior to this patch, fetch-pack would wait for
data from the server, and remote-curl would wait for fetch-pack,
causing a deadlock.
Despite this patch, there is other possible malformed input that could
cause the same deadlock (e.g. a half-finished pktline, or a pktline but
no trailing flush). There are a few possible solutions to this:
1. Allowing remote-curl to tell fetch-pack about the EOF (so that
fetch-pack could know that no more data is coming until it says
something else). This is tricky because an out-of-band signal would
be required, or the http response would have to be re-framed inside
another layer of pkt-line or something.
2. Make remote-curl understand some of the protocol. It turns out
that in addition to understanding pkt-line, it would need to watch for
ack/nak. This is somewhat fragile, as information about the protocol
would end up in two places. Also, pkt-lines which are already at the
length limit would need special handling.
Both of these solutions would require a fair amount of work, whereas
this hack is easy and solves at least some of the problem.
Still to do: it would be good to give a better error message
than "fatal: The remote end hung up unexpectedly".
Signed-off-by: David Turner <dturner@twosigma.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use the new GIT_TRACE_CURL environment variable instead
of the deprecated GIT_CURL_VERBOSE.
Signed-off-by: Elia Pinto <gitter.spiros@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To support this developer's use case of allowing build agents token-based
access to private repositories, we introduced the http.extraheader
feature, allowing extra HTTP headers to be sent along with every HTTP
request.
This patch verifies that we can configure these extra HTTP headers via the
command-line for use with `git submodule update`, too. Example: git -c
http.extraheader="Secret: Sauce" submodule update --init
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To test that extra HTTP headers are passed correctly, t5551 verifies that
a fetch succeeds when two required headers are passed, and that the fetch
does not succeed when those headers are not passed.
However, this test would also succeed if the configuration required only
one header. As Apache's configuration is notoriously tricky (this
developer frequently requires StackOverflow's help to understand Apache's
documentation), especially when still supporting the 2.2 line, let's just
really make sure that the test verifies what we want it to verify.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We introduce a way to send custom HTTP headers with all requests.
This allows us, for example, to send an extra token from build agents
for temporary access to private repositories. (This is the use case that
triggered this patch.)
This feature can be used like this:
git -c http.extraheader='Secret: sssh!' fetch $URL $REF
Note that `curl_easy_setopt(..., CURLOPT_HTTPHEADER, ...)` takes only
a single list, overriding any previous call. This means we have to
collect _all_ of the headers we want to use into a single list, and
feed it to cURL in one shot. Since we already unconditionally set a
"pragma" header when initializing the curl handles, we can add our new
headers to that list.
For callers which override the default header list (like probe_rpc),
we provide `http_copy_default_headers()` so they can do the same
trick.
Big thanks to Jeff King and Junio Hamano for their outstanding help and
patient reviews.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Communication between the HTTP server and http_backend process can
lead to a dead-lock when relaying a large ref negotiation request.
Diagnose the situation better, and mitigate it by reading such a
request first into core (to a reasonable limit).
* jk/http-backend-deadlock:
http-backend: spool ref negotiation requests to buffer
t5551: factor out tag creation
http-backend: fix die recursion with custom handler
* jk/http-backend-deadlock-2.3:
http-backend: spool ref negotiation requests to buffer
t5551: factor out tag creation
http-backend: fix die recursion with custom handler
* jk/http-backend-deadlock-2.2:
http-backend: spool ref negotiation requests to buffer
t5551: factor out tag creation
http-backend: fix die recursion with custom handler
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>
Test clean-up.
* jk/skip-http-tests-under-no-curl:
tests: skip dav http-push tests under NO_EXPAT=NoThanks
t/lib-httpd.sh: skip tests if NO_CURL is defined
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>
If we built git without curl, we can't actually test against
an http server. In fact, all of the test scripts which
include lib-httpd.sh already perform this check, with one
exception: t5540. For those scripts, this is a noop, and for
t5540, this is a bugfix (it used to fail when built with
NO_CURL, though it could go unnoticed if you had a stale
git-remote-https in your build directory).
Noticed-by: Junio C Hamano <junio@pobox.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
People often forget to chain the commands in their test together
with &&, leaving a failure from an earlier command in the test go
unnoticed. The new GIT_TEST_CHAIN_LINT mechanism allows you to
catch such a mistake more easily.
* jk/test-chain-lint: (36 commits)
t9001: drop save_confirm helper
t0020: use test_* helpers instead of hand-rolled messages
t: simplify loop exit-code status variables
t: fix some trivial cases of ignored exit codes in loops
t7701: fix ignored exit code inside loop
t3305: fix ignored exit code inside loop
t0020: fix ignored exit code inside loops
perf-lib: fix ignored exit code inside loop
t6039: fix broken && chain
t9158, t9161: fix broken &&-chain in git-svn tests
t9104: fix test for following larger parents
t4104: drop hand-rolled error reporting
t0005: fix broken &&-chains
t7004: fix embedded single-quotes
t0050: appease --chain-lint
t9001: use test_when_finished
t4117: use modern test_* helpers
t6034: use modern test_* helpers
t1301: use modern test_* helpers
t0020: use modern test_* helpers
...
Test fixes.
* jk/test-annoyances:
t5551: make EXPENSIVE test cheaper
t5541: move run_with_cmdline_limit to test-lib.sh
t: pass GIT_TRACE through Apache
t: redirect stderr GIT_TRACE to descriptor 4
t: translate SIGINT to an exit
These are tests which are missing a link in their &&-chain,
but during a setup phase. We may fail to notice failure in
commands that build the test environment, but these are
typically not expected to fail at all (but it's still good
to double-check that our test environment is what we
expect).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These are tests which are missing a link in their &&-chain,
but in a way that probably does not effect the outcome of
the test. Most of these are of the form:
some_cmd >actual
test_cmp expect actual
The main point of the test is to verify the output, and a
failure in some_cmd would probably be noticed by bogus
output. But it is good for the tests to also confirm that
"some_cmd" does not die unexpectedly after producing its
output.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
These are tests which are missing a link in their &&-chain,
in a location which causes a significant portion of the test
to be missed (e.g., the test effectively does nothing, or
consists of a long string of actions and output comparisons,
and we throw away the exit code of at least one part of the
string).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We create 50,000 tags to check that we don't overflow the
command-line of fetch-pack. But by using run_with_cmdline_limit,
we can get the same effect with a much smaller number of
tags. This makes the test fast enough that we can drop the
EXPENSIVE prereq, which means people will actually run it.
It was not documented to do so, but this test was also the
only test of a clone-over-http that requires multiple POSTs
during the conversation. We can continue to test that by
dropping http.postbuffer to its minimum size, and checking
that we get two POSTs.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When upload-pack advertises the refs (either for a normal,
non-stateless request, or for the initial contact in a
stateless one), we call for_each_ref with the send_ref
function as its callback. send_ref, in turn, calls
mark_our_ref, which checks whether the ref is hidden, and
sets OUR_REF or HIDDEN_REF on the object as appropriate. If
it is hidden, mark_our_ref also returns "1" to signal
send_ref that the ref should not be advertised.
If we are not advertising refs, (i.e., the follow-up
invocation by an http client to send its "want" lines), we
use mark_our_ref directly as a callback to for_each_ref. Its
marking does the right thing, but when it then returns "1"
to for_each_ref, the latter interprets this as an error and
stops iterating. As a result, we skip marking all of the
refs that come lexicographically after it. Any "want" lines
from the client asking for those objects will fail, as they
were not properly marked with OUR_REF.
To solve this, we introduce a wrapper callback around
mark_our_ref which always returns 0 (even if the ref is
hidden, we want to keep iterating). We also tweak the
signature of mark_our_ref to exclude unnecessary parameters
that were present only to conform to the callback interface.
This should make it less likely for somebody to accidentally
use it as a callback in the future.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* 'jc/test-lazy-prereq' (early part):
t3419: drop unnecessary NOT_EXPENSIVE pseudo-prerequisite
t3302: drop unnecessary NOT_EXPENSIVE pseudo-prerequisite
t3302: do not chdir around in the primary test process
t3302: coding style updates
test: turn USR_BIN_TIME into a lazy prerequisite
test: turn EXPENSIVE into a lazy prerequisite
Make clear which one is for dumb protocol, which one is for smart from
their file name.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Junio C Hamano
Jeff King
David Turner
Elia Pinto
Johannes Schindelin
Nguyễn Thái Ngọc Duy