1 - restore the previous sigaction struct as soon as we are about to shut
off our timeout
2 - restore the previous alarm() timeout, in case an application or similar
had it running before we "borrowed" it for a while.

No, this does not fix the multi-thread problem you get with alarm(). This
patch should correct bug report #478780:
//sourceforge.net/tracker/?func=detail&atid=100976&aid=478780&group_id=976

If not, please post details!

of real integers for versions, the defines are added to curl.h

should work a lot better!

HTTP server

for each *_perform(). It makes things a lot easier, as the first one that
detects the error get to write the final error reason...

even under normal unixes.

that is dead, because it can only find one entry anyway and if that is dead
there won't be any other entry that matches

information as well. Improves debug outputs etc.

rely on the new go-ahead-and-try mechanism that I just added to Transfer()

new logic can retry the same operation on a new connection!

- which now is made using a macro named FTPSENDF. I turned it all caps just
to make it more visible that it is in fact a macro.