Just to inform you, that I am almost sure that the problem is in the
underlining socket subsystem.

I've extracted example code again and compiled it exactly as it had been
done in the initial case - together with the libcurl sources, and using
the same configuration files (except strlcat which is absent on my linux
- but it doesn't play a role - I've verified it later). I've replaced
all marmalade-specific calls to the ncurses calls.

I've been running an example under valgrind for a hour, and then stopped
it. No any problems in the libcurl was found. No any problems was found
in the example itself - except some 'still reachable' memory in
finalization code of the example and ncurses library.

On the other side. We've tried to work around the problem, using direct
addresses instead of host names. Finally we've found that this
workaround doesn't work in the application. Then we locked connection
phase, allowing only one simultaneous connection phase (from connect()
call up to first byte received) and leaving the data transfer
concurrent. It works!

All of it makes me sure that the problem happens somewhere in the
underlining socket subsystem - when two simultaneous connections have
connection success notifications ready to be reported to the upper level.

Regards,
Vsevolod

-------------------------------------------------------------------
List admin: http://cool.haxx.se/list/listinfo/curl-library
Etiquette: http://curl.haxx.se/mail/etiquette.html
Received on 2011-09-14