Stop including config.h by default in cache.h. Instead only include
config.h in those files which require use of the config system.
Signed-off-by: Brandon Williams <bmwill@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow the http.postbuffer configuration variable to be set to a
size that can be expressed in size_t, which can be larger than
ulong on some platforms.
* dt/http-postbuffer-can-be-large:
http.postbuffer: allow full range of ssize_t values
Conversion from unsigned char [40] to struct object_id continues.
* bc/object-id:
Documentation: update and rename api-sha1-array.txt
Rename sha1_array to oid_array
Convert sha1_array_for_each_unique and for_each_abbrev to object_id
Convert sha1_array_lookup to take struct object_id
Convert remaining callers of sha1_array_lookup to object_id
Make sha1_array_append take a struct object_id *
sha1-array: convert internal storage for struct sha1_array to object_id
builtin/pull: convert to struct object_id
submodule: convert check_for_new_submodule_commits to object_id
sha1_name: convert disambiguate_hint_fn to take object_id
sha1_name: convert struct disambiguate_state to object_id
test-sha1-array: convert most code to struct object_id
parse-options-cb: convert sha1_array_append caller to struct object_id
fsck: convert init_skiplist to struct object_id
builtin/receive-pack: convert portions to struct object_id
builtin/pull: convert portions to struct object_id
builtin/diff: convert to struct object_id
Convert GIT_SHA1_RAWSZ used for allocation to GIT_MAX_RAWSZ
Convert GIT_SHA1_HEXSZ used for allocation to GIT_MAX_HEXSZ
Define new hash-size constants for allocating memory
Unfortunately, in order to push some large repos where a server does
not support chunked encoding, the http postbuffer must sometimes
exceed two gigabytes. On a 64-bit system, this is OK: we just malloc
a larger buffer.
This means that we need to use CURLOPT_POSTFIELDSIZE_LARGE to set the
buffer size.
Signed-off-by: David Turner <dturner@twosigma.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since this structure handles an array of object IDs, rename it to struct
oid_array. Also rename the accessor functions and the initialization
constant.
This commit was produced mechanically by providing non-Documentation
files to the following Perl one-liners:
perl -pi -E 's/struct sha1_array/struct oid_array/g'
perl -pi -E 's/\bsha1_array_/oid_array_/g'
perl -pi -E 's/SHA1_ARRAY_INIT/OID_ARRAY_INIT/g'
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make the internal storage for struct sha1_array use an array of struct
object_id internally. Update the users of this struct which inspect its
internals.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Teach remote-curl to understand push options and to be able to convey
them across HTTP.
Signed-off-by: Brandon Williams <bmwill@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
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
We instruct curl to always follow HTTP redirects. This is
convenient, but it creates opportunities for malicious
servers to create confusing situations. For instance,
imagine Alice is a git user with access to a private
repository on Bob's server. Mallory runs her own server and
wants to access objects from Bob's repository.
Mallory may try a few tricks that involve asking Alice to
clone from her, build on top, and then push the result:
1. Mallory may simply redirect all fetch requests to Bob's
server. Git will transparently follow those redirects
and fetch Bob's history, which Alice may believe she
got from Mallory. The subsequent push seems like it is
just feeding Mallory back her own objects, but is
actually leaking Bob's objects. There is nothing in
git's output to indicate that Bob's repository was
involved at all.
The downside (for Mallory) of this attack is that Alice
will have received Bob's entire repository, and is
likely to notice that when building on top of it.
2. If Mallory happens to know the sha1 of some object X in
Bob's repository, she can instead build her own history
that references that object. She then runs a dumb http
server, and Alice's client will fetch each object
individually. When it asks for X, Mallory redirects her
to Bob's server. The end result is that Alice obtains
objects from Bob, but they may be buried deep in
history. Alice is less likely to notice.
Both of these attacks are fairly hard to pull off. There's a
social component in getting Mallory to convince Alice to
work with her. Alice may be prompted for credentials in
accessing Bob's repository (but not always, if she is using
a credential helper that caches). Attack (1) requires a
certain amount of obliviousness on Alice's part while making
a new commit. Attack (2) requires that Mallory knows a sha1
in Bob's repository, that Bob's server supports dumb http,
and that the object in question is loose on Bob's server.
But we can probably make things a bit more obvious without
any loss of functionality. This patch does two things to
that end.
First, when we encounter a whole-repo redirect during the
initial ref discovery, we now inform the user on stderr,
making attack (1) much more obvious.
Second, the decision to follow redirects is now
configurable. The truly paranoid can set the new
http.followRedirects to false to avoid any redirection
entirely. But for a more practical default, we will disallow
redirects only after the initial ref discovery. This is
enough to thwart attacks similar to (2), while still
allowing the common use of redirects at the repository
level. Since c93c92f30 (http: update base URLs when we see
redirects, 2013-09-28) we re-root all further requests from
the redirect destination, which should generally mean that
no further redirection is necessary.
As an escape hatch, in case there really is a server that
needs to redirect individual requests, the user can set
http.followRedirects to "true" (and this can be done on a
per-server basis via http.*.followRedirects config).
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The discover_refs() function has a local "options" variable
to hold the http_get_options we pass to http_get_strbuf().
But this shadows the global "struct options" that holds our
program-level options, which cannot be accessed from this
function.
Let's give the local one a more descriptive name so we can
tell the two apart.
Signed-off-by: Jeff King <peff@peff.net>
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>
The existing "git fetch --depth=<n>" option was hard to use
correctly when making the history of an existing shallow clone
deeper. A new option, "--deepen=<n>", has been added to make this
easier to use. "git clone" also learned "--shallow-since=<date>"
and "--shallow-exclude=<tag>" options to make it easier to specify
"I am interested only in the recent N months worth of history" and
"Give me only the history since that version".
* nd/shallow-deepen: (27 commits)
fetch, upload-pack: --deepen=N extends shallow boundary by N commits
upload-pack: add get_reachable_list()
upload-pack: split check_unreachable() in two, prep for get_reachable_list()
t5500, t5539: tests for shallow depth excluding a ref
clone: define shallow clone boundary with --shallow-exclude
fetch: define shallow boundary with --shallow-exclude
upload-pack: support define shallow boundary by excluding revisions
refs: add expand_ref()
t5500, t5539: tests for shallow depth since a specific date
clone: define shallow clone boundary based on time with --shallow-since
fetch: define shallow boundary with --shallow-since
upload-pack: add deepen-since to cut shallow repos based on time
shallow.c: implement a generic shallow boundary finder based on rev-list
fetch-pack: use a separate flag for fetch in deepening mode
fetch-pack.c: mark strings for translating
fetch-pack: use a common function for verbose printing
fetch-pack: use skip_prefix() instead of starts_with()
upload-pack: move rev-list code out of check_non_tip()
upload-pack: make check_non_tip() clean things up on error
upload-pack: tighten number parsing at "deepen" lines
...
This should be part of every program, as otherwise users do
not get translated error messages. However, some external
commands forgot to do so (e.g., git-credential-store). This
fixes them, and eliminates the repeated code in programs
that did remember to use it.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Every program which links against libgit.a must call this
function, or risk hitting an assert() in system_path() that
checks whether we have configured argv0_path (though only
when RUNTIME_PREFIX is defined, so essentially only on
Windows).
Looking at the diff, you can see that putting it into the
common main() saves us having to do it individually in each
of the external commands. But what you can't see are the
cases where we _should_ have been doing so, but weren't
(e.g., git-credential-store, and all of the t/helper test
programs).
This has been an accident-waiting-to-happen for a long time,
but wasn't triggered until recently because it involves one
of those programs actually calling system_path(). That
happened with git-credential-store in v2.8.0 with ae5f677
(lazily load core.sharedrepository, 2016-03-11). The
program:
- takes a lock file, which...
- opens a tempfile, which...
- calls adjust_shared_perm to fix permissions, which...
- lazy-loads the config (as of ae5f677), which...
- calls system_path() to find the location of
/etc/gitconfig
On systems with RUNTIME_PREFIX, this means credential-store
reliably hits that assert() and cannot be used.
We never noticed in the test suite, because we set
GIT_CONFIG_NOSYSTEM there, which skips the system_path()
lookup entirely. But if we were to tweak git_config() to
find /etc/gitconfig even when we aren't going to open it,
then the test suite shows multiple failures (for
credential-store, and for some other test helpers). I didn't
include that tweak here because it's way too specific to
this particular call to be worth carrying around what is
essentially dead code.
The implementation is fairly straightforward, with one
exception: there is exactly one caller (git.c) that actually
cares about the result of the function, and not the
side-effect of setting up argv0_path. We can accommodate
that by simply replacing the value of argv[0] in the array
we hand down to cmd_main().
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are certain startup tasks that we expect every git
process to do. In some cases this is just to improve the
quality of the program (e.g., setting up gettext()). In
others it is a requirement for using certain functions in
libgit.a (e.g., system_path() expects that you have called
git_extract_argv0_path()).
Most commands are builtins and are covered by the git.c
version of main(). However, there are still a few external
commands that use their own main(). Each of these has to
remember to include the correct startup sequence, and we are
not always consistent.
Rather than just fix the inconsistencies, let's make this
harder to get wrong by providing a common main() that can
run this standard startup.
We basically have two options to do this:
- the compat/mingw.h file already does something like this by
adding a #define that replaces the definition of main with a
wrapper that calls mingw_startup().
The upside is that the code in each program doesn't need
to be changed at all; it's rewritten on the fly by the
preprocessor.
The downside is that it may make debugging of the startup
sequence a bit more confusing, as the preprocessor is
quietly inserting new code.
- the builtin functions are all of the form cmd_foo(),
and git.c's main() calls them.
This is much more explicit, which may make things more
obvious to somebody reading the code. It's also more
flexible (because of course we have to figure out _which_
cmd_foo() to call).
The downside is that each of the builtins must define
cmd_foo(), instead of just main().
This patch chooses the latter option, preferring the more
explicit approach, even though it is more invasive. We
introduce a new file common-main.c, with the "real" main. It
expects to call cmd_main() from whatever other objects it is
linked against.
We link common-main.o against anything that links against
libgit.a, since we know that such programs will need to do
this setup. Note that common-main.o can't actually go inside
libgit.a, as the linker would not pick up its main()
function automatically (it has no callers).
The rest of the patch is just adjusting all of the various
external programs (mostly in t/helper) to use cmd_main().
I've provided a global declaration for cmd_main(), which
means that all of the programs also need to match its
signature. In particular, many functions need to switch to
"const char **" instead of "char **" for argv. This effect
ripples out to a few other variables and functions, as well.
This makes the patch even more invasive, but the end result
is much better. We should be treating argv strings as const
anyway, and now all programs conform to the same signature
(which also matches the way builtins are defined).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In git-fetch, --depth argument is always relative with the latest
remote refs. This makes it a bit difficult to cover this use case,
where the user wants to make the shallow history, say 3 levels
deeper. It would work if remote refs have not moved yet, but nobody
can guarantee that, especially when that use case is performed a
couple months after the last clone or "git fetch --depth". Also,
modifying shallow boundary using --depth does not work well with
clones created by --since or --not.
This patch fixes that. A new argument --deepen=<N> will add <N> more (*)
parent commits to the current history regardless of where remote refs
are.
Have/Want negotiation is still respected. So if remote refs move, the
server will send two chunks: one between "have" and "want" and another
to extend shallow history. In theory, the client could send no "want"s
in order to get the second chunk only. But the protocol does not allow
that. Either you send no want lines, which means ls-remote; or you
have to send at least one want line that carries deep-relative to the
server..
The main work was done by Dongcan Jiang. I fixed it up here and there.
And of course all the bugs belong to me.
(*) We could even support --deepen=<N> where <N> is negative. In that
case we can cut some history from the shallow clone. This operation
(and --depth=<shorter depth>) does not require interaction with remote
side (and more complicated to implement as a result).
Helped-by: Duy Nguyen <pclouds@gmail.com>
Helped-by: Eric Sunshine <sunshine@sunshineco.com>
Helped-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Dongcan Jiang <dongcan.jiang@gmail.com>
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
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>
Update various codepaths to avoid manually-counted malloc().
* jk/tighten-alloc: (22 commits)
ewah: convert to REALLOC_ARRAY, etc
convert ewah/bitmap code to use xmalloc
diff_populate_gitlink: use a strbuf
transport_anonymize_url: use xstrfmt
git-compat-util: drop mempcpy compat code
sequencer: simplify memory allocation of get_message
test-path-utils: fix normalize_path_copy output buffer size
fetch-pack: simplify add_sought_entry
fast-import: simplify allocation in start_packfile
write_untracked_extension: use FLEX_ALLOC helper
prepare_{git,shell}_cmd: use argv_array
use st_add and st_mult for allocation size computation
convert trivial cases to FLEX_ARRAY macros
use xmallocz to avoid size arithmetic
convert trivial cases to ALLOC_ARRAY
convert manual allocations to argv_array
argv-array: add detach function
add helpers for allocating flex-array structs
harden REALLOC_ARRAY and xcalloc against size_t overflow
tree-diff: catch integer overflow in combine_diff_path allocation
...
"git fetch" and friends that make network connections can now be
told to only use ipv4 (or ipv6).
* ew/force-ipv4:
connect & http: support -4 and -6 switches for remote operations
Each of these cases can be converted to use ALLOC_ARRAY or
REALLOC_ARRAY, which has two advantages:
1. It automatically checks the array-size multiplication
for overflow.
2. It always uses sizeof(*array) for the element-size,
so that it can never go out of sync with the declared
type of the array.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are many manual argv allocations that predate the
argv_array API. Switching to that API brings a few
advantages:
1. We no longer have to manually compute the correct final
array size (so it's one less thing we can screw up).
2. In many cases we had to make a separate pass to count,
then allocate, then fill in the array. Now we can do it
in one pass, making the code shorter and easier to
follow.
3. argv_array handles memory ownership for us, making it
more obvious when things should be free()d and and when
not.
Most of these cases are pretty straightforward. In some, we
switch from "run_command_v" to "run_command" which lets us
directly use the argv_array embedded in "struct
child_process".
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
For curl error 35 (CURLE_SSL_CONNECT_ERROR) users need the
additional text stored in CURLOPT_ERRORBUFFER to debug why
the connection did not start. This is curl_errorstr inside
of http.c, so include that in the message if it is non-empty.
Sometimes HTTP response codes aren't yet available, such as
when the SSL setup fails. Don't include HTTP 0 in the message.
Signed-off-by: Shawn Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Sometimes it is necessary to force IPv4-only or IPv6-only operation
on networks where name lookups may return a non-routable address and
stall remote operations.
The ssh(1) command has an equivalent switches which we may pass when
we run them. There may be old ssh(1) implementations out there
which do not support these switches; they should report the
appropriate error in that case.
rsync support is untouched for now since it is deprecated and
scheduled to be removed.
Signed-off-by: Eric Wong <normalperson@yhbt.net>
Reviewed-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The strbuf_getline() interface allows a byte other than LF or NUL as
the line terminator, but this is only because I wrote these
codepaths anticipating that there might be a value other than NUL
and LF that could be useful when I introduced line_termination long
time ago. No useful caller that uses other value has emerged.
By now, it is clear that the interface is overly broad without a
good reason. Many codepaths have hardcoded preference to read
either LF terminated or NUL terminated records from their input, and
then call strbuf_getline() with LF or NUL as the third parameter.
This step introduces two thin wrappers around strbuf_getline(),
namely, strbuf_getline_lf() and strbuf_getline_nul(), and
mechanically rewrites these call sites to call either one of
them. The changes contained in this patch are:
* introduction of these two functions in strbuf.[ch]
* mechanical conversion of all callers to strbuf_getline() with
either '\n' or '\0' as the third parameter to instead call the
respective thin wrapper.
After this step, output from "git grep 'strbuf_getline('" would
become a lot smaller. An interim goal of this series is to make
this an empty set, so that we can have strbuf_getline_crlf() take
over the shorter name strbuf_getline().
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert the parse_fetch function to use struct object_id. Remove the
strlen check as get_oid_hex will fail safely on receiving a too-short
NUL-terminated string.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Jeff King <peff@peff.net>
Use struct object_id in three fields in struct ref and convert all the
necessary places that use it.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Jeff King <peff@peff.net>
This saves us some manual computation, and eliminates a call
to strcpy.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a new flag --sign=true (or --sign=false), which means the same
thing as the original --signed (or --no-signed). Give it a third
value --sign=if-asked to tell push and send-pack to send a push
certificate if and only if the server advertised a push cert nonce.
If not, warn the user that their push may not be as secure as they
thought.
Signed-off-by: Dave Borowitz <dborowitz@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Clear the git_zstream variable at the start of git_deflate_init() etc.
so that callers don't have to do that.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Using environment variable LANGUAGE and friends on the client side,
HTTP-based transports now send Accept-Language when making requests.
* ye/http-accept-language:
http: add Accept-Language header if possible
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>
Add an Accept-Language header which indicates the user's preferred
languages defined by $LANGUAGE, $LC_ALL, $LC_MESSAGES and $LANG.
Examples:
LANGUAGE= -> ""
LANGUAGE=ko:en -> "Accept-Language: ko, en;q=0.9, *;q=0.1"
LANGUAGE=ko LANG=en_US.UTF-8 -> "Accept-Language: ko, *;q=0.1"
LANGUAGE= LANG=en_US.UTF-8 -> "Accept-Language: en-US, *;q=0.1"
This gives git servers a chance to display remote error messages in
the user's preferred language.
Limit the number of languages to 1,000 because q-value must not be
smaller than 0.001, and limit the length of Accept-Language header to
4,000 bytes for some HTTP servers which cannot accept such long header.
Signed-off-by: Yi EungJun <eungjun.yi@navercorp.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow "git push" request to be signed, so that it can be verified and
audited, using the GPG signature of the person who pushed, that the
tips of branches at a public repository really point the commits
the pusher wanted to, without having to "trust" the server.
* jc/push-cert: (24 commits)
receive-pack::hmac_sha1(): copy the entire SHA-1 hash out
signed push: allow stale nonce in stateless mode
signed push: teach smart-HTTP to pass "git push --signed" around
signed push: fortify against replay attacks
signed push: add "pushee" header to push certificate
signed push: remove duplicated protocol info
send-pack: send feature request on push-cert packet
receive-pack: GPG-validate push certificates
push: the beginning of "git push --signed"
pack-protocol doc: typofix for PKT-LINE
gpg-interface: move parse_signature() to where it should be
gpg-interface: move parse_gpg_output() to where it should be
send-pack: clarify that cmds_sent is a boolean
send-pack: refactor inspecting and resetting status and sending commands
send-pack: rename "new_refs" to "need_pack_data"
receive-pack: factor out capability string generation
send-pack: factor out capability string generation
send-pack: always send capabilities
send-pack: refactor decision to send update per ref
send-pack: move REF_STATUS_REJECT_NODELETE logic a bit higher
...
The number of refs that can be pushed at once over smart HTTP was
limited by the command line length. The limitation has been lifted
by passing these refs from the standard input of send-pack.
* jk/send-pack-many-refspecs:
send-pack: take refspecs over stdin
The "--signed" option received by "git push" is first passed to the
transport layer, which the native transport directly uses to notice
that a push certificate needs to be sent. When the transport-helper
is involved, however, the option needs to be told to the helper with
set_helper_option(), and the helper needs to take necessary action.
For the smart-HTTP helper, the "necessary action" involves spawning
the "git send-pack" subprocess with the "--signed" option.
Once the above all gets wired in, the smart-HTTP transport now can
use the push certificate mechanism to authenticate its pushes.
Add a test that is modeled after tests for the native transport in
t5534-push-signed.sh to t5541-http-push-smart.sh. Update the test
Apache configuration to pass GNUPGHOME environment variable through.
As PassEnv would trigger warnings for an environment variable that
is not set, export it from test-lib.sh set to a harmless value when
GnuPG is not being used in the tests.
Note that the added test is deliberately loose and does not check
the nonce in this step. This is because the stateless RPC mode is
inevitably flaky and a nonce that comes back in the actual push
processing is one issued by a different process; if the two
interactions with the server crossed a second boundary, the nonces
will not match and such a check will fail. A later patch in the
series will work around this shortcoming.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Pushing a large number of refs works over most transports,
because we implement send-pack as an internal function.
However, it can sometimes fail when pushing over http,
because we have to spawn "git send-pack --stateless-rpc" to
do the heavy lifting, and we pass each refspec on the
command line. This can cause us to overflow the OS limits on
the size of the command line for a large push.
We can solve this by giving send-pack a --stdin option and
using it from remote-curl. We already dealt with this on
the fetch-pack side in 078b895 (fetch-pack: new --stdin
option to read refs from stdin, 2012-04-02). The stdin
option (and in particular, its use of packet-lines for
stateless-rpc input) is modeled after that solution.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Most struct child_process variables are cleared using memset first after
declaration. Provide a macro, CHILD_PROCESS_INIT, that can be used to
initialize them statically instead. That's shorter, doesn't require a
function call and is slightly more readable (especially given that we
already have STRBUF_INIT, ARGV_ARRAY_INIT etc.).
Helped-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* jk/remote-curl-squelch-extra-errors:
remote-curl: mark helper-protocol errors more clearly
remote-curl: use error instead of fprintf(stderr)
remote-curl: do not complain on EOF from parent git
When we encounter an error in remote-curl, we generally just
report it to stderr. There is no need for the user to care
that the "could not connect to server" error was generated
by git-remote-https rather than a function in the parent
git-fetch process.
However, when the error is in the protocol between git and
the helper, it makes sense to clearly identify which side is
complaining. These cases shouldn't ever happen, but when
they do, we can make them less confusing by being more
verbose.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We usually prefix our error messages with "error: ", but
many error messages from remote-curl are simply printed with
fprintf. This can make the output a little harder to read
(especially because such message may be intermingled with
errors from the parent git process).
There is no reason to avoid error(), as we are already
calling it many places (in addition to libgit.a functions
which use it).
While we're adjusting the messages, we can also drop the
capitalization which makes them unlike other git error
messages.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The parent git process is supposed to send us an empty line
to indicate that the conversation is over. However, the
parent process may die() if there is a problem with the
operation (e.g., we try to fetch a ref that does not exist).
In this case, it produces a useful message, but then
remote-curl _also_ produces an unhelpful message:
$ git pull origin matser
fatal: couldn't find remote ref matser
Unexpected end of command stream
The "right" way to fix this is to teach the parent git to
always cleanly close the connection to the helper, letting
it know that we are done. Implementing that is rather
clunky, though, as it would involve either replacing die()
operations with returning errors up the stack (until we
disconnect the transport), or adding an atexit handler to
clean up any transport helpers left open.
It's much simpler to just suppress the EOF message in
remote-curl. It was not added to address any real-world
situation in the first place, but rather a "we should
probably report unexpected things" suggestion[1].
It is the parent git which drives the operation, and whose
exit value actually matters. If the parent dies, then the
helper has no need to complain (except as a debugging aid).
In the off chance that the pipe is closed without the parent
dying, it can still notice the non-zero exit code.
[1] http://article.gmane.org/gmane.comp.version-control.git/176036
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It's a common idiom to match a prefix and then skip past it
with strlen, like:
if (starts_with(foo, "bar"))
foo += strlen("bar");
This avoids magic numbers, but means we have to repeat the
string (and there is no compiler check that we didn't make a
typo in one of the strings).
We can use skip_prefix to handle this case without repeating
ourselves.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We currently recognize an error message with a content-type
"text/plain; charset=utf-16" as text, but we ignore the
charset parameter entirely. Let's encode it to
log_output_encoding, which is presumably something the
user's terminal can handle.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we get a content-type from curl, we get the whole
header line, including any parameters, and without any
normalization (like downcasing or whitespace) applied.
If we later try to match it with strcmp() or even
strcasecmp(), we may get false negatives.
This could cause two visible behaviors:
1. We might fail to recognize a smart-http server by its
content-type.
2. We might fail to relay text/plain error messages to
users (especially if they contain a charset parameter).
This patch teaches the http code to extract and normalize
just the type/subtype portion of the string. This is
technically passing out less information to the callers, who
can no longer see the parameters. But none of the current
callers cares, and a future patch will add back an
easier-to-use method for accessing those parameters.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We currently don't reuse http connections when fetching via
the smart-http protocol. This is bad because the TCP
handshake introduces latency, and especially because SSL
connection setup may be non-trivial.
We can fix it by consistently using curl's "multi"
interface. The reason is rather complicated:
Our http code has two ways of being used: queuing many
"slots" to be fetched in parallel, or fetching a single
request in a blocking manner. The parallel code is built on
curl's "multi" interface. Most of the single-request code
uses http_request, which is built on top of the parallel
code (we just feed it one slot, and wait until it finishes).
However, one could also accomplish the single-request scheme
by avoiding curl's multi interface entirely and just using
curl_easy_perform. This is simpler, and is used by post_rpc
in the smart-http protocol.
It does work to use the same curl handle in both contexts,
as long as it is not at the same time. However, internally
curl may not share all of the cached resources between both
contexts. In particular, a connection formed using the
"multi" code will go into a reuse pool connected to the
"multi" object. Further requests using the "easy" interface
will not be able to reuse that connection.
The smart http protocol does ref discovery via http_request,
which uses the "multi" interface, and then follows up with
the "easy" interface for its rpc calls. As a result, we make
two HTTP connections rather than reusing a single one.
We could teach the ref discovery to use the "easy"
interface. But it is only once we have done this discovery
that we know whether the protocol will be smart or dumb. If
it is dumb, then our further requests, which want to fetch
objects in parallel, will not be able to reuse the same
connection.
Instead, this patch switches post_rpc to build on the
parallel interface, which means that we use it consistently
everywhere. It's a little more complicated to use, but since
we have the infrastructure already, it doesn't add any code;
we can just factor out the relevant bits from http_request.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fetching from a shallow-cloned repository used to be forbidden,
primarily because the codepaths involved were not carefully vetted
and we did not bother supporting such usage. This attempts to allow
object transfer out of a shallow-cloned repository in a controlled
way (i.e. the receiver become a shallow repository with truncated
history).
* nd/shallow-clone: (31 commits)
t5537: fix incorrect expectation in test case 10
shallow: remove unused code
send-pack.c: mark a file-local function static
git-clone.txt: remove shallow clone limitations
prune: clean .git/shallow after pruning objects
clone: use git protocol for cloning shallow repo locally
send-pack: support pushing from a shallow clone via http
receive-pack: support pushing to a shallow clone via http
smart-http: support shallow fetch/clone
remote-curl: pass ref SHA-1 to fetch-pack as well
send-pack: support pushing to a shallow clone
receive-pack: allow pushes that update .git/shallow
connected.c: add new variant that runs with --shallow-file
add GIT_SHALLOW_FILE to propagate --shallow-file to subprocesses
receive/send-pack: support pushing from a shallow clone
receive-pack: reorder some code in unpack()
fetch: add --update-shallow to accept refs that update .git/shallow
upload-pack: make sure deepening preserves shallow roots
fetch: support fetching from a shallow repository
clone: support remote shallow repository
...
Remove a few duplicate implementations of prefix/suffix comparison
functions, and rename them to starts_with and ends_with.
* cc/starts-n-ends-with:
replace {pre,suf}fixcmp() with {starts,ends}_with()
strbuf: introduce starts_with() and ends_with()
builtin/remote: remove postfixcmp() and use suffixcmp() instead
environment: normalize use of prefixcmp() by removing " != 0"
No callers pass a non-empty pointer as shallow_points at this
stage. As a result, all clients still refuse to talk to shallow
repository on the other end.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Leaving only the function definitions and declarations so that any
new topic in flight can still make use of the old functions, replace
existing uses of the prefixcmp() and suffixcmp() with new API
functions.
The change can be recreated by mechanically applying this:
$ git grep -l -e prefixcmp -e suffixcmp -- \*.c |
grep -v strbuf\\.c |
xargs perl -pi -e '
s|!prefixcmp\(|starts_with\(|g;
s|prefixcmp\(|!starts_with\(|g;
s|!suffixcmp\(|ends_with\(|g;
s|suffixcmp\(|!ends_with\(|g;
'
on the result of preparatory changes in this series.
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Issue "100 Continue" responses to help use of GSS-Negotiate
authentication scheme over HTTP transport when needed.
* bc/http-100-continue:
remote-curl: fix large pushes with GSSAPI
remote-curl: pass curl slot_results back through run_slot
http: return curl's AUTHAVAIL via slot_results
Due to an interaction between the way libcurl handles GSSAPI
authentication over HTTP and the way git uses libcurl, large
pushes (those over http.postBuffer bytes) would fail due to
an authentication failure requiring a rewind of the curl
buffer. Such a rewind was not possible because the data did
not fit into the entire buffer.
Enable the use of the Expect: 100-continue header for large
requests where the server offers GSSAPI authentication to
avoid this issue, since the request would otherwise fail.
This allows git to get the authentication data right before
sending the pack contents. Existing cases where pushes
would succeed, including small requests using GSSAPI, still
disable the use of 100 Continue, as it causes problems for
some remote HTTP implementations (servers and proxies).
Signed-off-by: Brian M. Carlson <sandals@crustytoothpaste.net>
Signed-off-by: Jeff King <peff@peff.net>
Some callers may want to know more than just the integer
error code we return. Let them optionally pass a
slot_results struct to fill in (or NULL if they do not
care). In either case we continue to return the integer
code.
We can also give probe_rpc the same treatment (since it
builds directly on run_slot).
Signed-off-by: Jeff King <peff@peff.net>
Handle the case where http transport gets redirected during the
authorization request better.
* jk/http-auth-redirects:
http.c: Spell the null pointer as NULL
remote-curl: rewrite base url from info/refs redirects
remote-curl: store url as a strbuf
remote-curl: make refs_url a strbuf
http: update base URLs when we see redirects
http: provide effective url to callers
http: hoist credential request out of handle_curl_result
http: refactor options to http_get_*
http_request: factor out curlinfo_strbuf
http_get_file: style fixes
For efficiency and security reasons, an earlier commit in
this series taught http_get_* to re-write the base url based
on redirections we saw while making a specific request.
This commit wires that option into the info/refs request,
meaning that a redirect from
http://example.com/foo.git/info/refs
to
https://example.com/bar.git/info/refs
will behave as if "https://example.com/bar.git" had been
provided to git in the first place.
The tests bear some explanation. We introduce two new
hierearchies into the httpd test config:
1. Requests to /smart-redir-limited will work only for the
initial info/refs request, but not any subsequent
requests. As a result, we can confirm whether the
client is re-rooting its requests after the initial
contact, since otherwise it will fail (it will ask for
"repo.git/git-upload-pack", which is not redirected).
2. Requests to smart-redir-auth will redirect, and require
auth after the redirection. Since we are using the
redirected base for further requests, we also update
the credential struct, in order not to mislead the user
(or credential helpers) about which credential is
needed. We can therefore check the GIT_ASKPASS prompts
to make sure we are prompting for the new location.
Because we have neither multiple servers nor https
support in our test setup, we can only redirect between
paths, meaning we need to turn on
credential.useHttpPath to see the difference.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
We use a strbuf to generate the string containing the remote
URL, but then detach it to a bare pointer. This makes it
harder to later manipulate the URL, as we have forgotten the
length (and the allocation semantics are not as clear).
Let's instead keep the strbuf around. As a bonus, this
eliminates a confusing double-use of the "buf" strbuf in
main(). Prior to this, it was used both for constructing the
url, and for reading commands from stdin.
The downside is that we have to update each call site to
refer to "url.buf" rather than just "url" when they want the
C string.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
In the discover_refs function, we use a strbuf named
"buffer" for multiple purposes. First we build the info/refs
URL in it, and then detach that to a bare pointer. Then, we
use the same strbuf to store the result of fetching the
refs.
Let's instead keep a separate refs_url strbuf. This is less
confusing, as the "buffer" strbuf is now used for only one
thing.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
When we are handling a curl response code in http_request or
in the remote-curl RPC code, we use the handle_curl_result
helper to translate curl's response into an easy-to-use
code. When we see an HTTP 401, we do one of two things:
1. If we already had a filled-in credential, we mark it as
rejected, and then return HTTP_NOAUTH to indicate to
the caller that we failed.
2. If we didn't, then we ask for a new credential and tell
the caller HTTP_REAUTH to indicate that they may want
to try again.
Rejecting in the first case makes sense; it is the natural
result of the request we just made. However, prompting for
more credentials in the second step does not always make
sense. We do not know for sure that the caller is going to
make a second request, and nor are we sure that it will be
to the same URL. Logically, the prompt belongs not to the
request we just finished, but to the request we are (maybe)
about to make.
In practice, it is very hard to trigger any bad behavior.
Currently, if we make a second request, it will always be to
the same URL (even in the face of redirects, because curl
handles the redirects internally). And we almost always
retry on HTTP_REAUTH these days. The one exception is if we
are streaming a large RPC request to the server (e.g., a
pushed packfile), in which case we cannot restart. It's
extremely unlikely to see a 401 response at this stage,
though, as we would typically have seen it when we sent a
probe request, before streaming the data.
This patch drops the automatic prompt out of case 2, and
instead requires the caller to do it. This is a few extra
lines of code, and the bug it fixes is unlikely to come up
in practice. But it is conceptually cleaner, and paves the
way for better handling of credentials across redirects.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Over time, the http_get_strbuf function has grown several
optional parameters. We now have a bitfield with multiple
boolean options, as well as an optional strbuf for returning
the content-type of the response. And a future patch in this
series is going to add another strbuf option.
Treating these as separate arguments has a few downsides:
1. Most call sites need to add extra NULLs and 0s for the
options they aren't interested in.
2. The http_get_* functions are actually wrappers around
2 layers of low-level implementation functions. We have
to pass these options through individually.
3. The http_get_strbuf wrapper learned these options, but
nobody bothered to do so for http_get_file, even though
it is backed by the same function that does understand
the options.
Let's consolidate the options into a single struct. For the
common case of the default options, we'll allow callers to
simply pass a NULL for the options struct.
The resulting code is often a few lines longer, but it ends
up being easier to read (and to change as we add new
options, since we do not need to update each call site).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Allow a safer "rewind of the remote tip" push than blind "--force",
by requiring that the overwritten remote ref to be unchanged since
the new history to replace it was prepared.
The machinery is more or less ready. The "--force" option is again
the big red button to override any safety, thanks to J6t's sanity
(the original round allowed --lockref to defeat --force).
The logic to choose the default implemented here is fragile
(e.g. "git fetch" after seeing a failure will update the
remote-tracking branch and will make the next "push" pass,
defeating the safety pretty easily). It is suitable only for the
simplest workflows, and it may hurt users more than it helps them.
* jc/push-cas:
push: teach --force-with-lease to smart-http transport
send-pack: fix parsing of --force-with-lease option
t5540/5541: smart-http does not support "--force-with-lease"
t5533: test "push --force-with-lease"
push --force-with-lease: tie it all together
push --force-with-lease: implement logic to populate old_sha1_expect[]
remote.c: add command line option parser for "--force-with-lease"
builtin/push.c: use OPT_BOOL, not OPT_BOOLEAN
cache.h: move remote/connect API out of it
We have been passing enough information to enable the
compare-and-swap logic down to the transport layer, but the
transport helper was not passing it to smart-http transport.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is an extension of c6807a4 (clone: open a shortcut for
connectivity check - 2013-05-26) to reduce the cost of connectivity
check at clone time, this time with smart http protocol.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead of using a hand-managed argument array, use argv-array API
to manage dynamically formulated command line.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we encounter an unknown http error (e.g., a 403), we
hand the error code to http_error, which then prints it with
error(). After that we die with the redundant message "HTTP
request failed".
Instead, let's just drop http_error entirely, which does
nothing but pass arguments to error(), and instead die
directly with a useful message.
So before:
$ git clone https://example.com/repo.git
Cloning into 'repo'...
error: unable to access 'https://example.com/repo.git': The requested URL returned error: 403 Forbidden
fatal: HTTP request failed
and after:
$ git clone https://example.com/repo.git
Cloning into 'repo'...
fatal: unable to access 'https://example.com/repo.git': The requested URL returned error: 403 Forbidden
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This helper function should really be a one-liner that
prints an error message, but it has ended up unnecessarily
complicated:
1. We call error() directly when we fail to start the curl
request, so we must later avoid printing a duplicate
error in http_error().
It would be much simpler in this case to just stuff the
error message into our usual curl_errorstr buffer
rather than printing it ourselves. This means that
http_error does not even have to care about curl's exit
value (the interesting part is in the errorstr buffer
already).
2. We return the "ret" value passed in to us, but none of
the callers actually cares about our return value. We
can just drop this entirely.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we report http errors in fetching the initial ref
advertisement, we show the full URL we attempted to use,
including "info/refs?service=git-upload-pack". While this
may be useful for debugging a broken server, it is
unnecessarily verbose and confusing for most cases, in which
the client user is not even the same person as the owner of
the repository.
Let's just show the repository URL; debugging can happen
with GIT_CURL_VERBOSE, which shows way more useful
information, anyway.
At the same time, let's also make sure to mention the
repository URL when we report failed authentication
(previously we said only "Authentication failed"). Knowing
the URL can help the user realize why authentication failed
(e.g., they meant to push to remote A, not remote B).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we get an http 404 trying to get the initial list of
refs from the server, we try to be helpful and remind the
user that update-server-info may need to be run. This looks
like:
$ git clone https://github.com/non/existent
Cloning into 'existent'...
fatal: https://github.com/non/existent/info/refs?service=git-upload-pack not found: did you run git update-server-info on the server?
Suggesting update-server-info may be a good suggestion for
users who are in control of the server repo and who are
planning to set up dumb http. But for users of smart http,
and especially users who are not in control of the server
repo, the advice is useless and confusing.
Since most people are expected to use smart http these days,
it does not make sense to keep the update-server-info hint.
We not only drop the mention of update-server-info, but also
show only the main repo URL, not the full "info/refs" and
service parameter. These elements may be useful for
debugging a broken server configuration, but in the majority
of cases, users are not fetching from their own
repositories, but rather from other people's repositories;
they have neither the power nor interest to fix a broken
configuration, and the extra components just make the
message more confusing. Users who do want to debug can and
should use GIT_CURL_VERBOSE to get more complete information
on the actual URLs visited.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we get an http 404 trying to get the initial list of
refs from the server, we try to be helpful and remind the
user that update-server-info may need to be run. This looks
like:
$ git clone https://github.com/non/existent
Cloning into 'existent'...
fatal: https://github.com/non/existent/info/refs?service=git-upload-pack not found: did you run git update-server-info on the server?
Suggesting update-server-info may be a good suggestion for
users who are in control of the server repo and who are
planning to set up dumb http. But for users of smart http,
and especially users who are not in control of the server
repo, the advice is useless and confusing.
The previous patch taught remote-curl to show custom advice
from the server when it is available. When we have shown
messages from the server, we can also drop our custom
advice; what the server has to say is likely to be more
accurate and helpful.
We not only drop the mention of update-server-info, but also
show only the main repo URL, not the full "info/refs" and
service parameter. These elements may be useful for
debugging a broken server configuration, but again, anything
the server has provided is likely to be more useful (and one
can still use GIT_CURL_VERBOSE to get much more complete
debugging information).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If an http request to a remote git server fails, we show
only the http response code, or sometimes a custom message
for particular codes. This gives the server no opportunity
to offer a more detailed explanation of the reason for the
failure, or to give extra advice.
This patch teaches remote-curl to record and display the
body content of a failed http response. We only display such
responses when the content-type is advertised as text/plain,
as it is the most likely to look presentable on the user's
terminal (and it is hoped to be a good indication that the
message is intended for git clients, and not for a web
browser).
Each line of the new output is prepended with "remote:".
Example output may look like this (assuming the server is
configured to display such a helpful message):
$ GIT_SMART_HTTP=0 git clone https://example.com/some/repo.git
Cloning into 'repo'...
remote: Sorry, fetching via dumb http is forbidden.
remote: Please upgrade your git client to v1.6.6 or greater
remote: and make sure that smart-http is enabled.
error: The requested URL returned error: 403 while accessing http://localhost:5001/some/repo.git/info/refs
fatal: HTTP request failed
For the sake of simplicity, we only record and display these
errors during the initial fetch of the ref list, as that is
the initial contact with the server and where the most
common, interesting errors happen (and there is already
precedent, as that is the only place we currently massage
http error codes into more helpful messages).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When remote-curl receives a list of refs from a server, it
keeps the whole buffer intact. When we get a "list" command,
we feed the result to get_remote_heads, and when we get a
"fetch" or "push" command, we feed it to fetch-pack or
send-pack, respectively.
If the HTTP response from the server is truncated for any
reason, we will get an incomplete ref advertisement. If we
then feed this incomplete list to fetch-pack, one of a few
things may happen:
1. If the truncation is in a packet header, fetch-pack
will notice the bogus line and complain.
2. If the truncation is inside a packet, fetch-pack will
keep waiting for us to send the rest of the packet,
which we never will.
3. If the truncation is at a packet boundary, fetch-pack
will keep waiting for us to send the next packet, which
we never will.
As a result, fetch-pack hangs, waiting for input. However,
remote-curl believes it has sent all of the advertisement,
and therefore waits for fetch-pack to speak. The two
processes end up in a deadlock.
We do notice the broken ref list if we feed it to
get_remote_heads. So if git asks the helper to do a "list"
followed by a "fetch", we are safe; we'll abort during the
list operation, which parses the refs.
This patch teaches remote-curl to always parse and save the
incoming ref list when we read the ref advertisement from a
server. That means that we will always verify and abort
before even running fetch-pack (or send-pack) when reading a
corrupted list, even if we do not run the "list" command
explicitly.
Since we save the result, in the common case of running
"list" then "fetch", we do not do any extra parsing at all.
In the case of just a "fetch", we do an extra round of
parsing, but only once.
Note also that the "fetch" case will now also initialize
server_capabilities from the remote (in remote-curl; we
already would do so inside fetch-pack). Doing "list+fetch"
already does this. It doesn't actually matter now, but the
new behavior is arguably more correct, should remote-curl
ever start caring about the server's capability list.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The ref-parsing functions are static. Let's move them up in
the file to be available to more functions, which will help
us with later refactoring.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Until recently, get_remote_heads only knew how to read refs
from a file descriptor. To hack around this, we spawned a
thread (or forked a process) to write the buffer back to us.
Now that we can just pass it our buffer directly, we don't
have to use this hack anymore.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that we can read packet data from memory as easily as a
descriptor, get_remote_heads can take either one as a
source. This will allow further refactoring in remote-curl.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The packet_read function reads from a descriptor. The
packet_get_line function is similar, but reads from an
in-memory buffer, and uses a completely separate
implementation. This patch teaches the generic packet_read
function to accept either source, and we can do away with
packet_get_line's implementation.
There are two other differences to account for between the
old and new functions. The first is that we used to read
into a strbuf, but now read into a fixed size buffer. The
only two callers are fine with that, and in fact it
simplifies their code, since they can use the same
static-buffer interface as the rest of the packet_read_line
callers (and we provide a similar convenience wrapper for
reading from a buffer rather than a descriptor).
This is technically an externally-visible behavior change in
that we used to accept arbitrary sized packets up to 65532
bytes, and now cap out at LARGE_PACKET_MAX, 65520. In
practice this doesn't matter, as we use it only for parsing
smart-http headers (of which there is exactly one defined,
and it is small and fixed-size). And any extension headers
would be breaking the protocol to go over LARGE_PACKET_MAX
anyway.
The other difference is that packet_get_line would return
on error rather than dying. However, both callers of
packet_get_line are actually improved by dying.
The first caller does its own error checking, but we can
drop that; as a result, we'll actually get more specific
reporting about protocol breakage when packet_read dies
internally. The only downside is that packet_read will not
print the smart-http URL that failed, but that's not a big
deal; anybody not debugging can already see the remote's URL
already, and anybody debugging would want to run with
GIT_CURL_VERBOSE anyway to see way more information.
The second caller, which is just trying to skip past any
extra smart-http headers (of which there are none defined,
but which we allow to keep room for future expansion), did
not error check at all. As a result, it would treat an error
just like a flush packet. The resulting mess would generally
cause an error later in get_remote_heads, but now we get
error reporting much closer to the source of the problem.
Brown-paper-bag-fixes-by: Ramsay Jones <ramsay@ramsay1.demon.co.uk>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The packets sent during ref negotiation are all terminated
by newline; even though the code to chomp these newlines is
short, we end up doing it in a lot of places.
This patch teaches packet_read_line to auto-chomp the
trailing newline; this lets us get rid of a lot of inline
chomping code.
As a result, some call-sites which are not reading
line-oriented data (e.g., when reading chunks of packfiles
alongside sideband) transition away from packet_read_line to
the generic packet_read interface. This patch converts all
of the existing callsites.
Since the function signature of packet_read_line does not
change (but its behavior does), there is a possibility of
new callsites being introduced in later commits, silently
introducing an incompatibility. However, since a later
patch in this series will change the signature, such a
commit would have to be merged directly into this commit,
not to the tip of the series; we can therefore ignore the
issue.
This is an internal cleanup and should produce no change of
behavior in the normal case. However, there is one corner
case to note. Callers of packet_read_line have never been
able to tell the difference between a flush packet ("0000")
and an empty packet ("0004"), as both cause packet_read_line
to return a length of 0. Readers treat them identically,
even though Documentation/technical/protocol-common.txt says
we must not; it also says that implementations should not
send an empty pkt-line.
By stripping out the newline before the result gets to the
caller, we will now treat the newline-only packet ("0005\n")
the same as an empty packet, which in turn gets treated like
a flush packet. In practice this doesn't matter, as neither
empty nor newline-only packets are part of git's protocols
(at least not for the line-oriented bits, and readers who
are not expecting line-oriented packets will be calling
packet_read directly, anyway). But even if we do decide to
care about the distinction later, it is orthogonal to this
patch. The right place to tighten would be to stop treating
empty packets as flush packets, and this change does not
make doing so any harder.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is just write_or_die by another name. The one
distinction is that write_or_die will treat EPIPE specially
by suppressing error messages. That's fine, as we die by
SIGPIPE anyway (and in the off chance that it is disabled,
write_or_die will simulate it).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Before parsing a suspected smart-HTTP response verify the returned
Content-Type matches the standard. This protects a client from
attempting to process a payload that smells like a smart-HTTP
server response.
JGit has been doing this check on all responses since the dawn of
time. I mistakenly failed to include it in git-core when smart HTTP
was introduced. At the time I didn't know how to get the Content-Type
from libcurl. I punted, meant to circle back and fix this, and just
plain forgot about it.
Signed-off-by: Shawn Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In particular, gcc issues an "'gzip_size' might be used uninitialized"
warning (-Wuninitialized). However, this warning is a false positive,
since the 'gzip_size' variable would not, in fact, be used uninitialized.
In order to suppress the warning, we simply initialise the variable to
zero in it's declaration.
Signed-off-by: Ramsay Jones <ramsay@ramsay1.demon.co.uk>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fixes fetch from servers that ask for auth only during the actual
packing phase. This is not really a recommended configuration, but it
cleans up the code at the same time.
* jk/maint-http-half-auth-fetch:
remote-curl: retry failed requests for auth even with gzip
remote-curl: hoist gzip buffer size to top of post_rpc
Commit b81401c taught the post_rpc function to retry the
http request after prompting for credentials. However, it
did not handle two cases:
1. If we have a large request, we do not retry. That's OK,
since we would have sent a probe (with retry) already.
2. If we are gzipping the request, we do not retry. That
was considered OK, because the intended use was for
push (e.g., listing refs is OK, but actually pushing
objects is not), and we never gzip on push.
This patch teaches post_rpc to retry even a gzipped request.
This has two advantages:
1. It is possible to configure a "half-auth" state for
fetching, where the set of refs and their sha1s are
advertised, but one cannot actually fetch objects.
This is not a recommended configuration, as it leaks
some information about what is in the repository (e.g.,
an attacker can try brute-forcing possible content in
your repository and checking whether it matches your
branch sha1). However, it can be slightly more
convenient, since a no-op fetch will not require a
password at all.
2. It future-proofs us should we decide to ever gzip more
requests.
Signed-off-by: Jeff King <peff@peff.net>
When we gzip the post data for a smart-http rpc request, we
compute the gzip body and its size inside the "use_gzip"
conditional. We keep track of the body after the conditional
ends, but not the size. Let's remember both, which will
enable us to retry failed gzip requests in a future patch.
Signed-off-by: Jeff King <peff@peff.net>
Further clean-up to the http codepath that picks up results after
cURL library is done with one request slot.
* jk/maint-http-init-not-in-result-handler:
http: do not set up curl auth after a 401
remote-curl: do not call run_slot repeatedly
Fixes a regression in maint-1.7.11 (v1.7.11.7), maint (v1.7.12.1)
and master (v1.8.0-rc0).
* jk/maint-http-half-auth-push:
http: fix segfault in handle_curl_result
When we get an http 401, we prompt for credentials and put
them in our global credential struct. We also feed them to
the curl handle that produced the 401, with the intent that
they will be used on a retry.
When the code was originally introduced in commit 42653c0,
this was a necessary step. However, since dfa1725, we always
feed our global credential into every curl handle when we
initialize the slot with get_active_slot. So every further
request already feeds the credential to curl.
Moreover, accessing the slot here is somewhat dubious. After
the slot has produced a response, we don't actually control
it any more. If we are using curl_multi, it may even have
been re-initialized to handle a different request.
It just so happens that we will reuse the curl handle within
the slot in such a case, and that because we only keep one
global credential, it will be the one we want. So the
current code is not buggy, but it is misleading.
By cleaning it up, we can remove the slot argument entirely
from handle_curl_result, making it much more obvious that
slots should not be accessed after they are marked as
finished.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit b81401c (http: prompt for credentials on failed POST)
taught post_rpc to call run_slot in a loop in order to retry
a request after asking the user for credentials. However,
after a call to run_slot we will have called
finish_active_slot. This means we have released the slot,
and we should no longer look at it.
As it happens, this does not cause any bugs in the current
code, since we know that we are not using curl_multi in this
code path, and therefore nobody will have taken over our
slot in the meantime. However, it is good form to actually
call get_active_slot again. It also future proofs us against
changes in the http code.
We can do this by jumping back to a retry label at the top
of our function. We just need to reorder a few setup lines
that should not be repeated; everything else within the loop
is either idempotent, needs to be repeated, or in a path we
do not follow (e.g., we do not even try when large_request
is set, because we don't know how much data we might have
streamed from our helper program).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we create an http active_request_slot, we can set its
"results" pointer back to local storage. The http code will
fill in the details of how the request went, and we can
access those details even after the slot has been cleaned
up.
Commit 8809703 (http: factor out http error code handling)
switched us from accessing our local results struct directly
to accessing it via the "results" pointer of the slot. That
means we're accessing the slot after it has been marked as
finished, defeating the whole purpose of keeping the results
storage separate.
Most of the time this doesn't matter, as finishing the slot
does not actually clean up the pointer. However, when using
curl's multi interface with the dumb-http revision walker,
we might actually start a new request before handing control
back to the original caller. In that case, we may reuse the
slot, zeroing its results pointer, and leading the original
caller to segfault while looking for its results inside the
slot.
Instead, we need to pass a pointer to our local results
storage to the handle_curl_result function, rather than
relying on the pointer in the slot struct. This matches what
the original code did before the refactoring (which did not
use a separate function, and therefore just accessed the
results struct directly).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allows users to turn off smart-http when talking to dumb-only
servers.
* jk/smart-http-switch:
remote-curl: let users turn off smart http
remote-curl: rename is_http variable
