On Mon, 5 May 2008, tetetest tetetest wrote:

> I have checked the usage of Curl_write(), and have come to conclusion that
> if we modify it "to do right", the existing code will break. Currently,
> libcurl does not handle EAGAIN at all - it simply retries to send the data
> until it all goes out or a failure occurs. By the way, this means that on
> slow lines, libcurl can take up lots of CPU time continuously calling send()
> in a cycle (am I right? Or did I miss some waits on the socket?).

Where do you see this happening?

> Handling EAGAIN from Curl_write() will require modifications of the "core"
> libcurl code (e.g., lib/transfer.c around line 1537). Not that I am afraid
> of hacking the sources, but isn't it too much for such a simple feature?
>
> So for the time being I simply check that Curl_write() returns CURLE_OK with
> zero bytes sent, and assume that it's the symptom of EAGAIN.

Yes, I believe it is and we can leave it like that for now.

> Please find attached the patch (5 existing files modified), 2 man pages
> (going to docs/libcurl), and one example of usage (going to docs/examples).

Ah nice, bonus points for the sample code!

I'll just await Michal Marek's checks on the 'const void *' front first and
see where that leads, then I hope to be able to commit your patch and we can
have some early testing and see how it all behaves!

Thanks for all your work and patience!

-- 
  Commercial curl and libcurl Technical Support: http://haxx.se/curl.html