I guess this should be adapted to #5199 as well. I will wait for @jay's PR to be merged first.

I'm in the process of redoing schannel_send to be non-blocking, hopefully it will help with #5177. The socket writable check there is likely going be removed. It calls select which will block on a non-blocking socket. AFAICS if the timeout is 0 and that means no timeout and the non-blocking socket is not ready then 0 will be returned. If that is the case then returning CURLE_OPERATION_TIMEDOUT is wrong. In other words, if the user has not given a timeout, and therefore the timeout is 0, then we shouldn't timeout if the socket is not ready.

Okay, I will wait for your enhancements to schannel_send then. I guess gopher.c can then be adjusted in a similar way.

Maybe Curl_timeleft should not return 0 when there is no timeout set? Return TIMEDIFF_T_MAX?

Maybe Curl_timeleft should not return 0 when there is no timeout set? Return TIMEDIFF_T_MAX?

Just checked Curl_timeleft("unspecifed")=> TIMEDIFF_T_MAX with schannel problem and it's works fine.
But personally i would prefer solution when sequentially pushed buffer to nonblock socket will not
use select-write check at all:

• select-write for non-blocked socket was always used to detect "is-connected" state.
  When we trying to push buffer, socket already passed connection-success state - no reason to check again.
• write-send method for nonblock socket will return E_AGAIN for retry or some fatal error. So there is no reason to make two syscalls insead of one.
  (like it's done here )

Maybe Curl_timeleft should not return 0 when there is no timeout set? Return TIMEDIFF_T_MAX?

I think that sounds like a pretty clever way to enhance the function and remove the need to special-case when it returns zero...

• select-write for non-blocked socket was always used to detect "is-connected" state.
  When we trying to push buffer, socket already passed connection-success state - no reason to check again.

I don't think this is correct, because:

If the socket is not processing a connect call, writability means a send, sendto, or WSASendto are guaranteed to succeed.

Source: https://docs.microsoft.com/en-us/windows/win32/api/winsock2/nf-winsock2-select#remarks

• write-send method for nonblock socket will return E_AGAIN for retry or some fatal error. So there is no reason to make two syscalls insead of one.
  (like it's done here )

I don't know if this is a good idea, because this is basically busy-looping and putting pressure on the socket write buffers. But I am not sure, maybe it is an alternative solution? cc @bagder @jay?

I don't think I understand the question. Should we avoid select() ? If we would do it entirely correctly, it should loop at all and then it shouldn't do select()...

But you're right, we should not busy-loop. That's the worst.

I don't think I understand the question. Should we avoid select() ?

I was referring to @ajax16384 points above and the specific reference to "(like it's done here)" there Curl_write_plain is called in a for-loop without waiting or select'ing, therefore busy-looping.

If we would do it entirely correctly, it should loop at all and then it shouldn't do select()...

Is that even possible inside schannel_send, due to the following?

And I am not sure if that would also apply to gopher_do, which this PR is actually about.

• write-send method for nonblock socket will return E_AGAIN for retry or some fatal error. So there is no reason to make two syscalls insead of one.
  (like it's done here )

...

I don't know if this is a good idea, because this is basically busy-looping and putting pressure on the socket write buffers. But I am not sure, maybe it is an alternative solution? cc @bagder @jay?

Yes you're right it's a busy loop if the socket isn't available. Though we shouldn't block, if we must then we should block waiting for the socket rather than call the underlying write repeatedly and eat CPU. That code in security.c needs to be improved, even if it can't be made non-blocking. I've added GSSAPI for FTP (afaict this is what security.c does) as a blocking operation to the TODO list. #5216

Source: https://docs.microsoft.com/en-us/windows/win32/api/winsock2/nf-winsock2-select#remarks

@mback2k well noticed: "If TIMEVAL is initialized to {0, 0}, select will return immediately; this is used to poll the state of the selected sockets.".
It shows that underlying socket is ready for send-write (without error or E_AGAIN) but almost the same way as calling to send-write (no cpu time release).
And that's why "SOCKET_WRITABLE(sockfd, 0)" useless to avoid buzy looping and i proposed that it might be removed.
But i did not realized (my bad) that You also deal with cpu hogging in THIS PR.
so "Curl_timeleft(unspecifed)=> TIMEDIFF_T_MAX - really good idea: fix buzy loop, fix schannel send timeout.
(Sorry for missleading proposal about removing SOCKET_WRITABLE before send-write)

Maybe Curl_timeleft should not return 0 when there is no timeout set? Return TIMEDIFF_T_MAX?

I think that sounds like a pretty clever way to enhance the function and remove the need to special-case when it returns zero...

I'm having second thoughts. I've since I looked through the codebase and it is depended on elsewhere that Curl_timeleft uses 0 to signal there is no timeout, so it may not be as easy as I thought. The crux of the issue is it is not immediately clear to contributors that Curl_timeleft returns 0 to signal there is no timeout set, though there is some comment at the end of the function that implies it. However most of the code seems to have been written with an understanding of that. Example:

It would be easier to update the documentation and then address those cases individually where 0 is wrongly treated as time remaining. For example in this PR

    timeleft = Curl_timeleft(conn->data, NULL, FALSE);
    if(timeleft < 0) {
      result = CURLE_OPERATION_TIMEDOUT;
      break;
    }
    if(!timeleft || timeleft > TIME_T_MAX)
      timeleft = TIME_T_MAX;

Otherwise we'd have to trace every Curl_timeleft to find which code depends on 0 to signal no timeout and then walk through that logic, and change Curl_timeleft to return 0 to mean no time left. Possibly renaming the function Curl_calculate_timeout or something would help to make it clear that 0 may have some special value?

@jay thanks, I updated this PR accordingly.

I found a similar issue in socks #5220 and I renamed timeleft to timeout_ms, since I think that may be easier to understand.

Thanks, looks good! I will also rename the variable here, because that makes a lot more sense.

Thanks