Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.unix.programmer > #1319 > unrolled thread

Non-blocking socket operations

Back to article view | Back to comp.unix.programmer

Contents

  Non-blocking socket operations ittium <ittium@gmail.com> - 2011-09-26 10:39 +0530
    Re: Non-blocking socket operations Nobody <nobody@nowhere.com> - 2011-09-26 13:02 +0100
    Re: Non-blocking socket operations Rainer Weikusat <rweikusat@mssgmbh.com> - 2011-09-26 19:10 +0100
    Re: Non-blocking socket operations William Ahern <william@wilbur.25thandClement.com> - 2011-09-26 13:44 -0700

#1319 — Non-blocking socket operations

Group,
One of my software was running in to performance problems randomly. 
Debugging of the problem revealed, that blocking socket calls ( blocking 
the worker thread) were reason for these performance problems.

I have recently modified the code so that it now uses the non-blocking 
operations. Given below are the changes that I have made.

connect
-------
connect on non-blocking socket may either return immediately or with 
EINPROGRESS. In the second case we start a timer and on timer expiry, 
using select, we find out whether connect was successful. fd is marked 
as connected if

	- socket fd is writable
	- writting zero bytes on socket is successful. If not

Issue -  Is this logic portable. Currently I can test the system on 
linux only but I would like to keep the system portable.

Accept
------
	- Accept is attempted on a listening non-blocking socket socket, till 
you get  EAGAIN.
	
Following errors are ignored (ie call accept again)
	
	-ECONNABORTED
	-EPROTO
Issue -  Is this logic portable. Currently I can test the system on 
linux only but I would like to keep the system portable.

getaddrinfo
-----------
	getaddrinfo is executed in a thread result is returned in a call back 
and then this thread exit.

Issue- When software is doing lot of DNS lookups, lots of such threads 
will be needed. Is there a better way of doing it.

thanks
Ittium

[toc] | [next] | [standalone]

#1322

On Mon, 26 Sep 2011 10:39:11 +0530, ittium wrote:

> getaddrinfo
> -----------
> 	getaddrinfo is executed in a thread result is returned in a call back 
> and then this thread exit.
> 
> Issue- When software is doing lot of DNS lookups, lots of such threads 
> will be needed. Is there a better way of doing it.

You could use the resolver functions directly, but it's non-trivial to get
it right (e.g. re-issuing requests using TCP if a UDP query results in a
truncated response), and it will only handle lookups via DNS, not
/etc/hosts, NIS, LDAP, etc.

On Linux, you can use getaddrinfo_a(), but it isn't portable. If your
program is GPL, you can use the GNU adns library.

[toc] | [prev] | [next] | [standalone]

#1324

ittium <ittium@gmail.com> writes:

[...]

> connect
> -------
> connect on non-blocking socket may either return immediately or with
> EINPROGRESS. In the second case we start a timer and on timer expiry,
> using select, we find out whether connect was successful. fd is marked
> as connected if
>
> 	- socket fd is writable
> 	- writting zero bytes on socket is successful. If not
>
> Issue -  Is this logic portable.

According to the current UNIX(*) standard an attempt to write zero
bytes 'may detect and return errors'. Using getsockopt querying the
option SO_ERROR at level SOL_SOCKET will do so. It is unclear what the
purpose of the timer is supposed to be. The networking stack already
contains logic for that and just waiting until the socket becomes
writable an inspection a possibly available error code should be
sufficient.

[...]

> getaddrinfo
> -----------
> 	getaddrinfo is executed in a thread result is returned in a
> call back and then this thread exit.
>
> Issue- When software is doing lot of DNS lookups, lots of such threads
> will be needed. Is there a better way of doing it.

The simple way would be to use a fixed number of address lookup threads
and some kind of (userspace) queue to distribute lookup requests among
them.

[toc] | [prev] | [next] | [standalone]

#1326

ittium <ittium@gmail.com> wrote:
<snip>
> getaddrinfo
> -----------
>         getaddrinfo is executed in a thread result is returned in a call back 
> and then this thread exit.

> Issue- When software is doing lot of DNS lookups, lots of such threads 
> will be needed. Is there a better way of doing it.

I generally don't like pimping my warez, but IMNSHO you might want to try
this:

	http://25thandclement.com/~william/projects/dns.c.html

See resolve_addrinfo() at the bottom of the source file for an example of
how to use the struct addrinfo-based API.

I've used c-ares, ADNS, etc for years. They were all poorly designed, IMHO.
They worked sort-of-okay for simple A lookups, but were a PITA for anything
more sophisticated. Plus, I hated being forced to use callbacks.

[toc] | [prev] | [standalone]

Back to top | Article view | comp.unix.programmer

csiph-web