On 14 May 2013, at 17:18, Daniel Stenberg <daniel_at_haxx.se> wrote:

>> We've been getting a problem with our unit tests where FTP uploads using curl_multi were stalling.
>
> Multi or easy makes no difference internally these days!

Good point.

Our layer on top (https://github.com/karelia/ConnectionKit incidentally) has an option to do things either synchronously or asynchronously, which for FTP essentially boils down to using curl_easy or curl_multi.

If we use the synchronous one, we don't get the stall. But then we're not doing things in parallel in that situation, so it's not really a fair comparison.

>> The test in question starts off three uploads and three downloads to the same server in parallel. Unless we're somehow missing an error condition somewhere, it looks like all the uploads hang.
>
> Are you using active FTP? Does the problem require these three uploads/downloads to trigger? Do all the uploads hang at the same spot in the transfer, and if so where exactly is that?

Passive I think. I'll have to go back and verify the answers to the other questions, but I think that they are no, and yes respectively!

There's a report of the stalls in some discussion here:

https://github.com/karelia/ConnectionKit/commit/a3b69c2160a205312259c6b338ae305fdf8383d5#commitcomment-3210475

and in that case it seems to be:

2013-05-14 19:34:26.223 Orbitink[73606:5c1b] TYPE I
2013-05-14 19:34:26.223 Orbitink[73606:5c1b] FTP 0x10087bd10 (line 3594) state change from STOP to STOR_TYPE
2013-05-14 19:34:26.296 Orbitink[73606:5c1b] 200 Switching to Binary mode.
2013-05-14 19:34:26.296 Orbitink[73606:5c1b] STOR index.html
2013-05-14 19:34:26.296 Orbitink[73606:5c1b] FTP 0x10087bd10 (line 1701) state change from STOR_TYPE to STOR

which does look like the right sort of place.

However, that thread is actually contradicting my claim that reverting your commit fixed it - so it may be a red herring, or there may be other things going on.

>> I seem to have tracked this down to commit 57aeabcc1a2
>
> I don't understand how this waiting for an additional socket (that the commit introduced) can cause a hang like that! Can you?

I have to confess that I've not traced the code through at the libcurl level yet.

I just did a kind of manual git bisect to find the point where the stall started happening, and it came down to that commit.

-------------------------------------------------------------------
List admin: http://cool.haxx.se/list/listinfo/curl-library
Etiquette: http://curl.haxx.se/mail/etiquette.html
Received on 2013-05-14