git-fetch vs ipv6 routing issues

I just noticed that, given a remote URL with a hostname which has both A and AAAA RRs in the DNS, git-fetch will retry a git-protocol fetch using the v4 address if the v6 address is unreachable, but will not do so when the remote is an http URL.

(I'm currently running bdb87afb4b4 from last month but will be updating later today. I don't see any relevant changes in the git log between then and now.)

-JimC

On Sat, 31 May 2008, James Cloos wrote:

Isn't this simply because libcurl (used for http) has no retry functionality on this scenario while git itself has that for the git protocol?

Daniel> Isn't this simply because libcurl (used for http) has no retry Daniel> functionality on this scenario while git itself has that for the Daniel> git protocol?

Yes, that is true.

But git could be smarter about it. libcurl has CURLOPT_IPRESOLVE which can be set to any of CURL_IPRESOLVE_WHATEVER, CURL_IPRESOLVE_V4 or CURL_IPRESOLVE_V6. Git could at least allow setting that via a config option and/or an env var, just like it does for libcurl options like CURLOPT_LOW_SPEED_LIMIT.

Even better would be to set CURLOPT_CONNECT_ONLY and then try with CURL_IPRESOLVE_WHATEVER, CURL_IPRESOLVE_V6 and CURL_IPRESOLVE_V4 in turn until it gets a connection, and then use CURLINFO_LASTSOCKET and the full URL to do the GET.

-JimC

On Tue, 3 Jun 2008, James Cloos wrote:

Perhaps, but I thought the git protocol behavior was to dynamicly back off a failed connect attempt to try the next? That wouldn't be solvable with any preset option. It needs code added to/changed in libcurl.

Eeek. I really disagree with this suggestion. CURLOPT_CONNECT_ONLY means you only (TCP or over proxy) connect and nothing more. That would force git to implement a lot of HTTP details that libcurl already provides.

If you really really insist on letting git do it and not bring this feature to libcurl, then I'd suggest that you first resolve the host, verify it by any means you see fit, and then pass the IP to libcurl like in the URL as "HTTP://12.23.45.67/blabla" with the correct host name in a custom-provided host: header. It'd let you do the resolving magic and let libcurl do the HTTP magic. But I wouldn't recommend that either...

Daniel> Eeek. I really disagree with this suggestion. CURLOPT_CONNECT_ONLY Daniel> means you only (TCP or over proxy) connect and nothing more. That Daniel> would force git to implement a lot of HTTP details that libcurl Daniel> already provides.

OK.  Then I misread that section of the curl_easy_setopt(3) man page.

I do think git should be able to do this even when compiled against legacy versions of libcurl, though.

(Of course, even better would be the ability to aviod http altogether. But some projects I track are only available via http urls....)

-JimC