| Message ID | pull.1681.v2.git.git.1708653536115.gitgitgadget@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] Add unix domain socket support to HTTP transport | expand |
"Leslie Cheng via GitGitGadget" <gitgitgadget@gmail.com> writes: > Subject: Re: [PATCH v2] Add unix domain socket support to HTTP transport Perhaps Subject: [PATCH] http: enable proxying via unix-domain socket to follow the usual "<area>: <description>" format? > From: Leslie Cheng <leslie.cheng5@gmail.com> > > This changeset introduces an `http.unixSocket` option so that users can "This changeset introduces" -> "Introduce". There may be other gotchas that might use help from Documentation/SubmittingPatches, but I didn't read too carefully. Besides, it is a single patch, not a set of changes ;-). `http.unixSocket` is a configuration variable. It may be confusing to use the word "option". Speaking of options, shouldn't there be a command line option that overrides the configured value? We should honor the usual http.<url>.VARIABLE convention where http.<url>.VARIABLE that is destination-specific overrides a more generic http.VARIABLE configuration variable. > proxy their git over HTTP remotes to a unix domain socket. In terms of > why, since UDS are local and git already has a local protocol: some > corporate environments use a UDS to proxy requests to internal resources > (ie. source control), so this change would support those use-cases. This "ie." -> "i.e.,"? > proxy can occasionally be necessary to attach MFA tokens or client > certificates for CLI tools. > > The implementation leverages `--unix-socket` option [0] via the > `CURLOPT_UNIX_SOCKET_PATH` flag available with libcurl [1]. There is a feature in libcURL library, that is enabled by setting the CURLOPT_UNIX_SOCKET_PATH option via the curl_easy_setopt() call, and their command line utility. You do the same to implement this feature. But when you are not adding "--unix-socket" option to any of our commands, mention of that option name makes it more confusing than necessary. The usual way to compose a log message of this project is to - Give an observation on how the current system work in the present tense (so no need to say "Currently X is Y", just "X is Y"), and discuss what you perceive as a problem in it. - Propose a solution (optional---often, problem description trivially leads to an obvious solution in reader's minds). - Give commands to the codebase to "become like so". in this order. How about following that convention, perhaps like: In some corporate environments, the proxy server listens to a local unix domain socket for requests, instead of listening to a network port. Even though we have http.proxy (and more destination specific http.<url>.proxy) configuration variables to specify the network address/port of a proxy, that would not help if your proxy does not listen to the network. Introduce an `http.unixSocket` (and `http.<url>.unixSocket`) configuration variables that specify the path to a unix domain socket for such a proxy. Recent versions of libcURL library added CURLOPT_UNIX_SOCKET_PATH to support "curl --unix-socket <path>"---use the same mechanism to implement it. > `GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH` and `NO_UNIX_SOCKETS` were kept > separate so that we can spit out better error messages for users if git > was compiled with `NO_UNIX_SOCKETS`. Unlike NO_UNIX_SOCKETS, GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH is entirely internal to your implementation and not surfaced to neither the end-users or the binary packagers. Because of that, I suspect that any description that has to use that name probably falls on the other side of "too much implementation details" to be useful to help future developers.. Besides, I suspect that GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH might not be the optimum approach. See below. > diff --git a/Documentation/config/http.txt b/Documentation/config/http.txt > index 2d4e0c9b869..bf48cbd599a 100644 > --- a/Documentation/config/http.txt > +++ b/Documentation/config/http.txt > @@ -277,6 +277,11 @@ http.followRedirects:: > the base for the follow-up requests, this is generally > sufficient. The default is `initial`. > > +http.unixSocket:: > + Connect through this Unix domain socket via HTTP, instead of using the > + network. If set, this config takes precendence over `http.proxy` and > + is incompatible with the proxy options (see `curl(1)`). Talking about precedence between this and http.proxy is good thing, but one very important piece of information is missing. What value does it take? The absolute path of a unix-domain socket to pass the HTTP traffic over, instead of using the network. or something, perhaps? > diff --git a/git-curl-compat.h b/git-curl-compat.h > index fd96b3cdffd..f0f3bec0e17 100644 > --- a/git-curl-compat.h > +++ b/git-curl-compat.h > @@ -74,6 +74,13 @@ > #define GIT_CURL_HAVE_CURLE_SSL_PINNEDPUBKEYNOTMATCH 1 > #endif > > +/** > + * CURLOPT_UNIX_SOCKET_PATH was added in 7.40.0, released in January 2015. > + */ > +#if LIBCURL_VERSION_NUM >= 0x074000 > +#define GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH 1 > +#endif The "HAVE" part in GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH is a statement of a fact. If the version of cURL library we have is certain value, we have it. OK. > diff --git a/http.c b/http.c > index e73b136e589..8cfdcaeac82 100644 > --- a/http.c > +++ b/http.c > @@ -79,6 +79,9 @@ static const char *http_proxy_ssl_ca_info; > static struct credential proxy_cert_auth = CREDENTIAL_INIT; > static int proxy_ssl_cert_password_required; It might make the code easier to follow if you did: #if !defined(NO_CURLOPT_UNIX_SOCKET_PATH) && !defined(NO_UNIX_SOCKETS) #if defined(GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH) #define USE_CURLOPT_UNIX_SOCKET_PATH #endif #endif The points are (1) the users can decline to use CURLOPT_UNIX_SOCKET_PATH while still using unix domain socket for other purposes, and (2) you do not have to care if you HAVE it or not most of time; what matters more often is if the user told you to USE it. Hmm? > +#if defined(GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH) && !defined(NO_UNIX_SOCKETS) > +static const char *curl_unix_socket_path; > +#endif The guard here would become "#ifdef USE_CURLOPT_UNIX_SOCKET_PATH" if we wanted this to be conditional, but I think it is easier to make the variable unconditionally available; see below. > @@ -455,6 +458,20 @@ static int http_options(const char *var, const char *value, > return 0; > } > > + if (!strcmp("http.unixsocket", var)) { > +#ifdef GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH > +#ifndef NO_UNIX_SOCKETS > + return git_config_string(&curl_unix_socket_path, var, value); > +#else > + warning(_("Unix socket support unavailable in this build of Git")); > + return 0; > +#endif > +#else > + warning(_("Unix socket support is not supported with cURL < 7.40.0")); > + return 0; > +#endif > + } In general, it is inadvisable to issue a warning in the codepath that parses configuration variables, as the values we read may not be necessarily used. We could instead accept the given path into a variable unconditionally, and complain only before it gets used, near the call to curl_easy_setopt(). > if (!strcmp("http.cookiefile", var)) > return git_config_pathname(&curl_cookie_file, var, value); > if (!strcmp("http.savecookies", var)) { > @@ -1203,6 +1220,12 @@ static CURL *get_curl_handle(void) > } > init_curl_proxy_auth(result); > > +#if defined(GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH) && !defined(NO_UNIX_SOCKETS) > + if (curl_unix_socket_path) { > + curl_easy_setopt(result, CURLOPT_UNIX_SOCKET_PATH, curl_unix_socket_path); > + } > +#endif Here, the guard may become more like if (curl_unix_socket_path) { #ifdef USE_CURLOPT_UNIX_SOCKET_PATH curl_easy_setopt(...); #elif defined(NO_CURLOPT_UNIX_SOCKET_PATH) || defined(NO_UNIX_SOCKETS) warning(_("this build disables the unix-domain-socket feature")); #elif warning(_("your cURL library is too old")); #endif }
Junio C Hamano <gitster@pobox.com> writes: > "Leslie Cheng via GitGitGadget" <gitgitgadget@gmail.com> writes: > >> Subject: Re: [PATCH v2] Add unix domain socket support to HTTP transport > > Perhaps > > Subject: [PATCH] http: enable proxying via unix-domain socket > > to follow the usual "<area>: <description>" format? > >> From: Leslie Cheng <leslie.cheng5@gmail.com> >> >> This changeset introduces an `http.unixSocket` option so that users can > > "This changeset introduces" -> "Introduce". There may be other > gotchas that might use help from Documentation/SubmittingPatches, > but I didn't read too carefully. > > Besides, it is a single patch, not a set of changes ;-). > > `http.unixSocket` is a configuration variable. It may be confusing > to use the word "option". Speaking of options, shouldn't there be a > command line option that overrides the configured value? > > We should honor the usual http.<url>.VARIABLE convention where > http.<url>.VARIABLE that is destination-specific overrides a more > generic http.VARIABLE configuration variable. Clarification. I know the above is automatically achieved, given the way we have laid urlmatch foundation to allow easy parsing for configuration variables structured this way. I did not mean that you'd need to do anything special; rather, I meant that we should advertise that we do in the commit log message.
On 2024-02-23 12:37 a.m., Junio C Hamano wrote: > How about following that convention, perhaps like: > > In some corporate environments, the proxy server listens to a > local unix domain socket for requests, instead of listening to a > network port. Even though we have http.proxy (and more > destination specific http.<url>.proxy) configuration variables > to specify the network address/port of a proxy, that would not > help if your proxy does not listen to the network. > > Introduce an `http.unixSocket` (and `http.<url>.unixSocket`) > configuration variables that specify the path to a unix domain > socket for such a proxy. Recent versions of libcURL library > added CURLOPT_UNIX_SOCKET_PATH to support "curl --unix-socket > <path>"---use the same mechanism to implement it. This is excellent, thanks for the guidance (and all the other suggestions prior)! I'll update in the next patch. > Unlike NO_UNIX_SOCKETS, GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH is > entirely internal to your implementation and not surfaced to neither > the end-users or the binary packagers. Because of that, I suspect > that any description that has to use that name probably falls on the > other side of "too much implementation details" to be useful to help > future developers.. That's reasonable, I figured it would fit as a cover letter detail but I agree it's not relevant as a commit message that lives in the history of the project. I'll also update this in the next patch. > Talking about precedence between this and http.proxy is good thing, > but one very important piece of information is missing. What value > does it take? > > The absolute path of a unix-domain socket to pass the HTTP > traffic over, instead of using the network. > > or something, perhaps? I like that wording, I'll update in the next patch. > It might make the code easier to follow if you did: > > #if !defined(NO_CURLOPT_UNIX_SOCKET_PATH) && !defined(NO_UNIX_SOCKETS) > #if defined(GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH) > #define USE_CURLOPT_UNIX_SOCKET_PATH > #endif > #endif > > The points are > > (1) the users can decline to use CURLOPT_UNIX_SOCKET_PATH while > still using unix domain socket for other purposes, and > > (2) you do not have to care if you HAVE it or not most of time; > what matters more often is if the user told you to USE it. > > Hmm? Do you think this functionality is worth adding another macro to conditionally include it in the build? It felt out-of-the-way enough that we could just use the same `NO_UNIX_SOCKETS` macro to control for environments that don't support unix domain sockets. >> +#if defined(GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH) && !defined(NO_UNIX_SOCKETS) >> +static const char *curl_unix_socket_path; >> +#endif > > The guard here would become "#ifdef USE_CURLOPT_UNIX_SOCKET_PATH" if > we wanted this to be conditional, but I think it is easier to make > the variable unconditionally available; see below. Agreed in general, I was looking to other patterns for conditional variables in file, e.g. https://github.com/git/git/blob/3c2a3fdc388747b9eaf4a4a4f2035c1c9ddb26d0/http.c#L66-L68 But, revisiting, this looks like an exception rather than the norm. > In general, it is inadvisable to issue a warning in the codepath > that parses configuration variables, as the values we read may not > be necessarily used. We could instead accept the given path into a > variable unconditionally, and complain only before it gets used, > near the call to curl_easy_setopt(). Similar to above, I followed what was already done for certain configuration variables (e.g. https://github.com/git/git/blob/3c2a3fdc388747b9eaf4a4a4f2035c1c9ddb26d0/http.c#L485-L501), but I agree with your feedback since this would result in constant warnings. To summarize, I'll do the following in the next patch: - change the wording of the commit message - move the conditional variables and parsing to a check at `curl_easy_setopt()` time I'm still undecided on whether I should introduce another macro specifically for this functionality, and I'd like to hear your thoughts on why it might make sense.
diff --git a/Documentation/config/http.txt b/Documentation/config/http.txt index 2d4e0c9b869..bf48cbd599a 100644 --- a/Documentation/config/http.txt +++ b/Documentation/config/http.txt @@ -277,6 +277,11 @@ http.followRedirects:: the base for the follow-up requests, this is generally sufficient. The default is `initial`. +http.unixSocket:: + Connect through this Unix domain socket via HTTP, instead of using the + network. If set, this config takes precendence over `http.proxy` and + is incompatible with the proxy options (see `curl(1)`). + http.<url>.*:: Any of the http.* options above can be applied selectively to some URLs. For a config key to match a URL, each element of the config key is diff --git a/git-curl-compat.h b/git-curl-compat.h index fd96b3cdffd..f0f3bec0e17 100644 --- a/git-curl-compat.h +++ b/git-curl-compat.h @@ -74,6 +74,13 @@ #define GIT_CURL_HAVE_CURLE_SSL_PINNEDPUBKEYNOTMATCH 1 #endif +/** + * CURLOPT_UNIX_SOCKET_PATH was added in 7.40.0, released in January 2015. + */ +#if LIBCURL_VERSION_NUM >= 0x074000 +#define GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH 1 +#endif + /** * CURL_HTTP_VERSION_2 was added in 7.43.0, released in June 2015. * diff --git a/http.c b/http.c index e73b136e589..8cfdcaeac82 100644 --- a/http.c +++ b/http.c @@ -79,6 +79,9 @@ static const char *http_proxy_ssl_ca_info; static struct credential proxy_cert_auth = CREDENTIAL_INIT; static int proxy_ssl_cert_password_required; +#if defined(GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH) && !defined(NO_UNIX_SOCKETS) +static const char *curl_unix_socket_path; +#endif static struct { const char *name; long curlauth_param; @@ -455,6 +458,20 @@ static int http_options(const char *var, const char *value, return 0; } + if (!strcmp("http.unixsocket", var)) { +#ifdef GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH +#ifndef NO_UNIX_SOCKETS + return git_config_string(&curl_unix_socket_path, var, value); +#else + warning(_("Unix socket support unavailable in this build of Git")); + return 0; +#endif +#else + warning(_("Unix socket support is not supported with cURL < 7.40.0")); + return 0; +#endif + } + if (!strcmp("http.cookiefile", var)) return git_config_pathname(&curl_cookie_file, var, value); if (!strcmp("http.savecookies", var)) { @@ -1203,6 +1220,12 @@ static CURL *get_curl_handle(void) } init_curl_proxy_auth(result); +#if defined(GIT_CURL_HAVE_CURLOPT_UNIX_SOCKET_PATH) && !defined(NO_UNIX_SOCKETS) + if (curl_unix_socket_path) { + curl_easy_setopt(result, CURLOPT_UNIX_SOCKET_PATH, curl_unix_socket_path); + } +#endif + set_curl_keepalive(result); return result; diff --git a/t/t5565-http-unix-domain-socket.sh b/t/t5565-http-unix-domain-socket.sh new file mode 100755 index 00000000000..2f9c53ab14f --- /dev/null +++ b/t/t5565-http-unix-domain-socket.sh @@ -0,0 +1,109 @@ +#!/bin/sh + +test_description="test fetching through http via unix domain socket" + +. ./test-lib.sh +. "$TEST_DIRECTORY"/lib-httpd.sh + +test -z "$NO_UNIX_SOCKETS" || { + skip_all='skipping http-unix-socket tests, unix sockets not available' + test_done +} + +if ! test_have_prereq PERL +then + skip_all='skipping http-unix-socket tests; perl not available' + test_done +fi + +SOCKET_PROXY_PIDFILE="$(pwd)/proxy.pid" +UDS_SOCKET="$(pwd)/uds.sock" +UNRESOLVABLE_ENDPOINT=http://unresolved + +start_proxy_unix_to_tcp() { + test_atexit 'stop_proxy_unix_to_tcp' + + perl -Mstrict -MIO::Select -MIO::Socket::INET -MIO::Socket::UNIX -e ' + my $uds_path = $ARGV[0]; + my $host = $ARGV[1]; + my $port = $ARGV[2]; + my $pidfile = $ARGV[3]; + + open(my $fh, ">", $pidfile) or die "failed to create pidfile"; + print $fh "$$"; + close($fh); + + my $uds = IO::Socket::UNIX->new( + Local => $uds_path, + Type => SOCK_STREAM, + Listen => 5, + ) or die "failed to create unix domain socket"; + + while (my $conn = $uds->accept()) { + my $tcp_client = IO::Socket::INET->new( + PeerAddr => $host, + PeerPort => $port, + Proto => "tcp", + ) or die "failed to create TCP socket"; + + my $sel = IO::Select->new($conn, $tcp_client); + + while (my @ready = $sel->can_read(10)) { + foreach my $socket (@ready) { + my $other = ($socket == $conn) ? $tcp_client : $conn; + my $data; + my $bytes = $socket->sysread($data, 4096); + + if ($bytes) { + $other->syswrite($data, $bytes); + } else { + $socket->close(); + } + } + } + } + ' "$UDS_SOCKET" "127.0.0.1" "$LIB_HTTPD_PORT" "$SOCKET_PROXY_PIDFILE" & + SOCKET_PROXY_PID=$! +} + +stop_proxy_unix_to_tcp() { + kill -9 "$(cat "$SOCKET_PROXY_PIDFILE")" + rm -f "$SOCKET_PROXY_PIDFILE" + rm -f "$UDS_SOCKET" +} + +start_httpd +start_proxy_unix_to_tcp + +test_expect_success 'setup repository' ' + test_commit foo && + git init --bare "$HTTPD_DOCUMENT_ROOT_PATH/repo.git" && + git push --mirror "$HTTPD_DOCUMENT_ROOT_PATH/repo.git" +' + +# sanity check that we can't clone normally +test_expect_success 'cloning without UDS fails' ' + test_must_fail git clone "$UNRESOLVABLE_ENDPOINT/smart/repo.git" clone +' + +test_expect_success 'cloning with UDS succeeds' ' + test_when_finished "rm -rf clone" && + test_config_global http.unixsocket "$UDS_SOCKET" && + git clone "$UNRESOLVABLE_ENDPOINT/smart/repo.git" clone +' + +test_expect_success 'cloning with a non-existent http proxy fails' ' + git clone $HTTPD_URL/smart/repo.git clone && + rm -rf clone && + test_config_global http.proxy 127.0.0.1:0 && + test_must_fail git clone $HTTPD_URL/smart/repo.git clone +' + +test_expect_success 'UDS socket takes precedence over http proxy' ' + test_when_finished "rm -rf clone" && + test_config_global http.proxy 127.0.0.1:0 && + test_config_global http.unixsocket "$UDS_SOCKET" && + git clone $HTTPD_URL/smart/repo.git clone +' + +test_done