On Apr 30, 2014, at 6:54 AM, Yann Ylavic wrote:
> -- Forwarded message --
> From: Yann Ylavic
> Date: Wed, Apr 30, 2014 at 11:59 AM
> Subject: Re: apr_socket_opt_set always sets TCP_DEFER_ACCEPT to 1
> To: APR Developer List
>
> On Tue, Apr 29, 2014 at 5:41 PM, Jim Jagielski wrote:
>>
>> On Apr 22, 2014, at 9:43 AM, Brian J. France wrote:
>>>
>>> For some reason I completely missed that. APR is fine, but httpd needs
>>> updated as it is hard coded to 1:
>>>
>>> rv = apr_socket_opt_set(s, APR_TCP_DEFER_ACCEPT, 1);
>>>
>>> Will move to the httpd list.
>>
>>
>> Did I miss this on dev@httpd?
>
> I seems that httpd trunk and 2.4.x use "apr_socket_opt_set(s,
> APR_TCP_DEFER_ACCEPT, 30);" whereas 2.2.x uses a value of 1.
>
> This come from PR
> http://issues.apache.org/bugzilla/show_bug.cgi?id=41270, commit
> http://svn.apache.org/viewvc?view=revision&revision=501364, which
> since then has been forked to 2.4.x, and never merged in 2.2.x.
>
> Linux (code and mostly comment) is quite clear on the setsockopt() value.
> In master/net/ipv4/tcp.c::do_tcp_setsockopt() :
>case TCP_DEFER_ACCEPT:
>/* Translate value in seconds to number of retransmits */
>icsk->icsk_accept_queue.rskq_defer_accept =
>secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
>TCP_RTO_MAX / HZ);
>break;
>
> One second is too low imo (cf. PR above).
Yes, 1 is way to low, even 60 doesn't help us with the RHEL 6 kernel bug.
I am trying to get time to work on a patch that would allow something like this:
AcceptFilter data:60
for 2.2.x. That is the only way I can see to allow tweaking the time and
staying binary compatible with 2.2.x.
Original thought was to change AcceptFilter to a AP_INIT_TAKE23 and add a
accf_opt_map table, but that would change the conf struct size and break binary
compatibility for 2.2.x.
Brian
