diff mbox series

[v2] Added Curl Option to Override Request Method v2

Message ID pull.1225.v2.git.1650988659280.gitgitgadget@gmail.com (mailing list archive)
State New, archived
Headers show
Series [v2] Added Curl Option to Override Request Method v2 | expand

Commit Message

Drew Green April 26, 2022, 3:57 p.m. UTC
From: agreenbhm <agreenbhm@gmail.com>

Added support for environment variable "CURLOPT_CUSTOMREQUEST"
and config option "http.customrequest" to allow setting the Curl
option to override the default request method used by HTTP Git
operations.  Primary reason for this is to allow support for
cloning repositories where only GET requests
are allowed by a local web proxy but not POSTs.  When cloning
a repo first a GET is made to the server and then a
POST is made to the "git-upload-pack" endpoint.  In some
corporate environments with strong controls
only GET requests are allowed to known repository hosts (such
as GitHub) through a web proxy to prevent data leakage.  Using this
new setting, a user can set the "CURLOPT_CUSTOMREQUEST=GET" env at runtime
or "http.customrequest = GET" in their config file which will
change the second request from a POST to a GET, bypassing
web proxy restrictions on the type of requests allowed.
Tested with GitHub, changing the request from POST to GET still
results in the expected behavior of the repo successfully being cloned.

This is v2 of this patch, which refactored the placement of the env
and added the ability to set the config file option.

Signed-off-by: agreenbhm <agreenbhm@gmail.com>
---
    Added Curl Option to Override Request Method
    
    Added support for environment variable "CURLOPT_CUSTOMREQUEST" to allow
    setting the curl option to override the default request method used by
    HTTP Git operations. Primary reason for this is to allow support for
    cloning repositories where only GET requests are allowed but not POSTs.
    When cloning a repo first a GET is made to the server and then a POST is
    made to the "git-upload-pack" endpoint. In some corporate environments
    with strong controls only GET requests are allowed to known repository
    hosts (such as GitHub) to prevent data leakage by sending data. Using
    this new environmental variable, a user can set
    "CURLOPT_CUSTOMREQUEST=GET" which will change the second request from a
    POST to a GET, bypassing web proxy restrictions on the type of requests
    allowed. Tested with GitHub, changing the request from POST to GET still
    results in the expected behavior of the repo successfully being cloned.
    
    Signed-off-by: agreenbhm agreenbhm@gmail.com

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1225%2Fagreenbhm%2Fmaster-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1225/agreenbhm/master-v2
Pull-Request: https://github.com/gitgitgadget/git/pull/1225

Range-diff vs v1:

 1:  8bf14c61c2a < -:  ----------- Added Curl Option to Override Request Method
 -:  ----------- > 1:  8734bf28344 Added Curl Option to Override Request Method v2


 http.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)


base-commit: 6cd33dceed60949e2dbc32e3f0f5e67c4c882e1e

Comments

brian m. carlson April 26, 2022, 9:28 p.m. UTC | #1
On 2022-04-26 at 15:57:39, Drew Green via GitGitGadget wrote:
> From: agreenbhm <agreenbhm@gmail.com>
> 
> Added support for environment variable "CURLOPT_CUSTOMREQUEST"
> and config option "http.customrequest" to allow setting the Curl
> option to override the default request method used by HTTP Git
> operations.  Primary reason for this is to allow support for
> cloning repositories where only GET requests
> are allowed by a local web proxy but not POSTs.  When cloning
> a repo first a GET is made to the server and then a
> POST is made to the "git-upload-pack" endpoint.  In some
> corporate environments with strong controls
> only GET requests are allowed to known repository hosts (such
> as GitHub) through a web proxy to prevent data leakage.  Using this
> new setting, a user can set the "CURLOPT_CUSTOMREQUEST=GET" env at runtime
> or "http.customrequest = GET" in their config file which will
> change the second request from a POST to a GET, bypassing
> web proxy restrictions on the type of requests allowed.
> Tested with GitHub, changing the request from POST to GET still
> results in the expected behavior of the repo successfully being cloned.

I don't think this is a good idea.  It may happen that GitHub or other
servers happen to accept a GET request here, but that is a bug and
should be fixed.  It is definitely not something we should depend on or
rely on, and it isn't a documented part of the protocol.

If your corporate environment doesn't allow POST requests, you may wish
to use SSH for Git operations instead, or you may need to explain to
your company why you cannot do your job with their proxy in place.
Drew Green April 26, 2022, 11:34 p.m. UTC | #2
On Tue, Apr 26, 2022 at 7:30 PM brian m. carlson
<sandals@crustytoothpaste.net> wrote:
>
> On 2022-04-26 at 15:57:39, Drew Green via GitGitGadget wrote:
> > From: agreenbhm <agreenbhm@gmail.com>
> >
> > Added support for environment variable "CURLOPT_CUSTOMREQUEST"
> > and config option "http.customrequest" to allow setting the Curl
> > option to override the default request method used by HTTP Git
> > operations.  Primary reason for this is to allow support for
> > cloning repositories where only GET requests
> > are allowed by a local web proxy but not POSTs.  When cloning
> > a repo first a GET is made to the server and then a
> > POST is made to the "git-upload-pack" endpoint.  In some
> > corporate environments with strong controls
> > only GET requests are allowed to known repository hosts (such
> > as GitHub) through a web proxy to prevent data leakage.  Using this
> > new setting, a user can set the "CURLOPT_CUSTOMREQUEST=GET" env at runtime
> > or "http.customrequest = GET" in their config file which will
> > change the second request from a POST to a GET, bypassing
> > web proxy restrictions on the type of requests allowed.
> > Tested with GitHub, changing the request from POST to GET still
> > results in the expected behavior of the repo successfully being cloned.
>
> I don't think this is a good idea.  It may happen that GitHub or other
> servers happen to accept a GET request here, but that is a bug and
> should be fixed.  It is definitely not something we should depend on or
> rely on, and it isn't a documented part of the protocol.
>
> If your corporate environment doesn't allow POST requests, you may wish
> to use SSH for Git operations instead, or you may need to explain to
> your company why you cannot do your job with their proxy in place.
> --
> brian m. carlson (he/him or they/them)
> Toronto, Ontario, CA

Brian - I understand what you're saying, however I don't think adding
this feature is detrimental in any way. It is simply leveraging a
feature of curl without any promises of resulting behavior. Why not
allow users to take advantage of a library feature if it can help?
brian m. carlson April 27, 2022, 12:49 a.m. UTC | #3
On 2022-04-26 at 23:34:32, Drew Green wrote:
> Brian - I understand what you're saying, however I don't think adding
> this feature is detrimental in any way. It is simply leveraging a
> feature of curl without any promises of resulting behavior. Why not
> allow users to take advantage of a library feature if it can help?

Because the protocol is defined to use POST.  We might well define GET
to mean something different in the future.  HTTP verbs are not
interchangeable, and GET requests can be subject to caching while POST
requests typically cannot.  Similarly, GET requests are idempotent,
while POST requests are not, and what you're proposing would allow
creating non-idempotent GET requests, thereby violating the HTTP
specification.

Also, people will expect it to work and complain when it doesn't. People
will complain that GitHub no longer supports it if we fix the bug[0]
because Git has an option to use GET.  People will also expect this
support in Git LFS as well even though it cannot possibly work there,
and be angry that we won't support it.

I'm very much opposed to this.  There are all sorts of broken proxies
which break Git in various ways, and this is just another situation in
which they do that.  The solution is to use SSH, if that's available, or
fix the proxy.  A proxy that doesn't handle POST properly doesn't even
speak HTTP/1.1 properly, and that was defined in 1999.  There's really
no excuse to have such software on the Internet today.  Such a proxy
wouldn't even be usable for a web browser.

[0] It is my intention to fix this bug in GitHub's Git service in short
order.
diff mbox series

Patch

diff --git a/http.c b/http.c
index 229da4d1488..eaf269fc5a7 100644
--- a/http.c
+++ b/http.c
@@ -75,6 +75,7 @@  static const char *http_proxy_ssl_key;
 static const char *http_proxy_ssl_ca_info;
 static struct credential proxy_cert_auth = CREDENTIAL_INIT;
 static int proxy_ssl_cert_password_required;
+static const char *http_custom_request;
 
 static struct {
 	const char *name;
@@ -403,6 +404,9 @@  static int http_options(const char *var, const char *value, void *cb)
 		return 0;
 	}
 
+	if(!strcmp("http.customrequest", var))
+		return git_config_string(&http_custom_request, var, value);
+
 	/* Fall back on the default ones */
 	return git_default_config(var, value, cb);
 }
@@ -1099,6 +1103,7 @@  void http_init(struct remote *remote, const char *url, int proactive_auth)
 		    starts_with(url, "https://"))
 			ssl_cert_password_required = 1;
 	}
+	set_from_env(&http_custom_request, "CURLOPT_CUSTOMREQUEST");
 
 	curl_default = get_curl_handle();
 }
@@ -1212,7 +1217,7 @@  struct active_request_slot *get_active_slot(void)
 		curl_easy_setopt(slot->curl, CURLOPT_COOKIEJAR, curl_cookie_file);
 	curl_easy_setopt(slot->curl, CURLOPT_HTTPHEADER, pragma_header);
 	curl_easy_setopt(slot->curl, CURLOPT_ERRORBUFFER, curl_errorstr);
-	curl_easy_setopt(slot->curl, CURLOPT_CUSTOMREQUEST, NULL);
+	curl_easy_setopt(slot->curl, CURLOPT_CUSTOMREQUEST, http_custom_request);
 	curl_easy_setopt(slot->curl, CURLOPT_READFUNCTION, NULL);
 	curl_easy_setopt(slot->curl, CURLOPT_WRITEFUNCTION, NULL);
 	curl_easy_setopt(slot->curl, CURLOPT_POSTFIELDS, NULL);