#826 connect race with multi and socks proxy

closed-fixed
libcurl (354)
5
2009-05-11
2009-04-30
No

Symptom: connections through SOCKS proxy fail when using multi interface - especially SOCKS4A.

Problem: socket is being used by Curl_SOCKS4 before it is connected

Solution: block on connect

Seen on Windows.

Discussion

1 2 > >> (Page 1 of 2)
  • See attached patch. Not all connections fail - some succeed.

     
  • Your patch makes the connect phase blocking. I don't see why that is necessary. libcurl should simply proceed as usual until connected, why is using a (socks) proxy any different that any other proxy or connection?

     
  • Better description of problem: I was seeing 75% of my Socks4a connections failing on setup with socket not connected. Now I see 0%.

    The flow I was seeing (before the patch) is the following:

    ConnectPlease calls Curl_connecthere which returns immediately in case of curl_multi, not waiting for connet to finish

    ConnectPlease calls Curl_SOCKS4
    which (in the case of SOCKS4a) calls Curl_write_plain
    which calls send() which gets back socket is not connected

    The eventual solution is to make Curl_SOCKS support multi directly and not require blocking, but that will take a lot more time. In the meantime, there is bug on the known bugs list that says that curl blocks during proxy connect and multi, but it actually doesn't on Windows.

     
  • Ah, yes thanks for clarifying!

    But instead of taking the route of making the behavior even more blocking, I would instead suggest that we make the socks proxy approach more similar to the http proxy one.

    It uses the asynch connection approch, so for the case when we use the multi interface and Curl_connecthost() doesn't complete the connect, we could instead make the socks negotiation magic get called when we reach the CURLM_STATE_PROTOCONNECT state in lib/multi.c. Wouldn't that work (better) you think?

     
  • sorry, it should be at the CURLM_STATE_WAITCONNECT position, which happens at TCP connect

     
  • Agree that not blocking would be the best solution but until that gets written, I suggest this patch as a stopgap. Since many SOCKS proxies are in the LAN i.e. low delay, blocking may not be that bad for the app. The current behavior is bad.

    Here are the states I see for SOCKS:
    - waiting for SOCKS connect
    - Send DNS request for final hostname (not 4a or 5_HOSTNAME)
    - waiting for DNS response
    - sending SOCKS request
    - waiting for SOCKS reply

    I suppose those states can get assigned states at the top level state machine or SOCKS could
    maintain its own substate under CURLM_STATE_WAITCONNECT.

     
  • my take at doing the socks connect magic after the tcp connect

     
  • Right, but my suggested approach is only a stop-gap take anyway, just less crude than yours. See my uploaded patch attempt for an explanation on what I mean.

    This approach waits for the TCP connect to be verified before it calls the socks-functions, and thus I hope it'll survive better.

    What do you think about this?

     
  • I like your approach better. I will test it.

     
1 2 > >> (Page 1 of 2)