| Ticket UUID: | 824251 | ||
| Title: | async client sockets block for DNS | ||
| Type: | Bug | Created on: | 2003-10-15 16:47:07 |
| Submitter: | nobody | Assigned to: | aku |
| Subsystem: | 27. Channel Types | Severity: | Minor |
| Priority: | 7 High | Last modified: | 2022-09-08 16:47:51 |
| Status: | Open | Closed by: | nobody |
| Resolution: | Later | Closed on: | |
| Version: | obsolete: 8.3.2 | ||
| Description: | ||||
The geturl documentation claims that, geturl will return immediatly with a token if the -command option is invoked. If geturl can not make the initial socket connection to a url, it blocks until the timeout expires and calls the -command proc before geturl has returned a token. This causes at least the following two problems: 1. The block causes any following geturl calls to hang before processing. This destroys the asynchronious operation. 2. It invokes the -command proc with a token that is as yet unknown to the general processing loop. [email protected] | ||||
| User Comments: | ||||
dkf added on 2003-10-16 03:49:01:
Logged In: YES user_id=79902 Problem (still in 8.5a0) is that [socket -async] still does the hostname -> IP address synchronously. Which I've known about for ages, but it is really hard to fix. The problem is that the OS interface for doing the mapping is itself purely synchronous and a completely monolithic mess. Working around that requires a helper process or thread, and hence the whole thing is rather complex to fix. davygrvy added on 2003-11-13 16:07:52:
Logged In: YES user_id=7549 I can name that extension on windows in 45 minutes.. Does BSD have anyway to do an async gethostbyname? Does this fall into OS specific for sol/irix/linux/etc.. ? dkf added on 2003-11-13 21:42:32:
Logged In: YES user_id=79902 The problem is the sheer complexity of stuff hidden under the covers. If it was just DNS (or if there was a sensible way of parsing the config files across all platforms), that'd be so much easier. dkf added on 2005-04-27 04:44:48:
Logged In: YES user_id=79902 Joe English suggests using getaddrinfo() for name lookups instead of gethostbyname(). In terms of portability, it seems to be available identically on Linux, Solaris and IRIX for sure (no access to other Unix platforms!) and it is also the subject of standardization (see POSIX 1003.1g and RFC 2553). IOW, perfect! And if there are platforms which don't support it, we could detect that in configure and fall back to the bad old gethostbyname() for [socket -async] processing. davygrvy added on 2005-04-27 05:46:37:
Logged In: YES user_id=7549 getaddrinfo() works on windows, too. But where's the spec for timeout behavior? Just for fun, I've been experimenting with an external helper app to be used for async name lookups. If we launch it in an async pipe, problem solved. http://cvs.sf.net/viewcvs.py/iocpsock/iocpsock/tclname.c?rev=HEAD&view=auto dkf added on 2005-04-27 06:28:40:
Logged In: YES user_id=79902 Doesn't seem to have any defined or controllable timeout behaviour. You might (on POSIX systems) be able to use a signal, but the result of doing so is undefined. dkf added on 2012-07-03 14:54:03:
Could also use a subprocess to do the resolution specifically for HTTP. dkf added on 2012-08-27 22:47:50:
Worth looking at doing something about this (well, at least for the http package) after TIP #409 is done. kjnash added on 2022-09-08 16:47:51:
I have added a workaround to 8.7 [608f2fbab4]. The ::socket command is evaluated asynchronously in another thread while the script waits in the event loop. Because the Thread package may not be present, the new facility is optional and is switched off by default. Its use is controlled by the new http::config option -threadlevel. The new code runs in an idletask coroutine. This uses a trace to detect command completion, hence avoiding the stacking issues of vwait. The socket replacement command http::socket is short, and if desired it can easily be replaced with one that uses a different technique. | ||||
