[squid-users] [PATCH] Squid 3.5.19 SMP under OpenBSD - setsockopt for UDS

[Silamael]

On 27.06.2016 13:19, Amos Jeffries wrote:
> On 27/06/2016 9:16 p.m., Silamael wrote:
>> Hi,
>>
>> I'm playing around with the SMP feature on OpenBSD 5.9 and noticed that
>> Squid does not run due to hard coded limits for the receive and send
>> buffer sizes of Unix Domain Sockets. In contrary to other OSes these
>> limits cannot be adjusted by a sysctl.
> 
> Is that a new regression in 5.9?
> Can you provide a reference please?

Haven't tested it with a prior version than 5.9.
But as the buffer limits haven't changed for a long time I don't think
that's a regression in 5.9.

> 
> We have had Squid working just fine with SMP on OpenBSD before.

That's interesting. Without the patch with no disk cache at all
configured and 4 workers, Squid seem to run a short time until the
children terminate due to a registration timeout.

> 
>> The attached patch adds some setsockopt() calls to comm.cc which sets
>> the buffer sizes to 256kb.
> 
> Please submit patchs to the squid-dev mailing list.
> 
> Please also include an example of the errors that are visible to users
> with your squid-dev submission so we can refer people complaining about
> specific error messages to the appropriate fix.

The error I got was a write failed on the UDS socket. Error code was
EMSGSIZE. Squid tried to write 4320 bytes on the socket.

> 
> And detail of where the 2K/4K limit you mention is located. So we can
> verify if you are looking at the right limitation and what else it affects.
>  AFAIK, the UDS limitations are about the size of objects which are used
> over UDS sockets. Reading 256KB into a 4KB object does not have pretty
> results. Likewise a blocking queue of up to 64 time-critical SMP actions
> awaiting the first one getting its synchronous UDS response does not
> have good side effects.

The limits can be found in /usr/src/sys/kern/uipc_usrreq.c:
u_long unpdg_sendspace = 2*1024;
u_long unpdg_recvspace = 4*1024;

Quick check against OpenBSD 5.2 showed no difference.
The value of 256kb is just taken from the Squid SMP FAQ page.

> 
> We no longer accept #ifdef construction in squid coding guidelines. Use
> #if instead. If you can please also identify what other *BSD are needing
> this and add them to the #if condition.
>  Though if this is accepted the change may be relevant to all OS and not
> needing a wrapper at all.

Ok, I can change the #ifdef into an #if. But I don't have other BSDs for
testing.

> 
>>
>> PS: Also i noticed that if squid is called with -z for creating the
>> cache directories, that ${process_number} macro is not expanded but
>> always set to 0. squid -z only creates one cache dir although in the
>> configuration
>> cache_dir ufs /var/squid/cache/cache-foo-${process_number} 700 8 16
>> is set.
> 
> That sounds more like the behaviour of "-N -z".

Ah, that explains a lot. Our wrapper scripts are using -N when creating
the cache directories...