Re: [squid-users] Ongoing Running out of filedescriptors

On Wed, Feb 10, 2010 at 8:50 AM, Luis Daniel Lucio Quiroz
<luis.daniel.lucio_at_gmail.com> wrote:
> Le Mardi 9 F�vrier 2010 19:34:13, Amos Jeffries a �crit :
>> On Tue, 9 Feb 2010 17:39:37 -0600, Luis Daniel Lucio Quiroz
>>
>> <luis.daniel.lucio_at_gmail.com> wrote:
>> > Le Mardi 9 Février 2010 17:29:23, Landy Landy a écrit :
>> >> I don't know what to do with my current squid, I even upgraded to
>> >> 3.0.STABLE21 but, the problem persist every three days:
>> >>
>> >> /usr/local/squid/sbin/squid -v
>> >> Squid Cache: Version 3.0.STABLE21
>> >> configure options: �'--prefix=/usr/local/squid'
>>
>> '--sysconfdir=/etc/squid'
>>
>> >> '--enable-delay-pools' '--enable-kill-parent-hack' '--disable-htcp'
>> >> '--enable-default-err-language=Spanish' '--enable-linux-netfilter'
>> >> '--disable-ident-lookups' '--localstatedir=/var/log/squid3.1'
>> >> '--enable-stacktraces' '--with-default-user=proxy' '--with-large-files'
>> >> '--enable-icap-client' '--enable-async-io' '--enable-storeio=aufs'
>> >> '--enable-removal-policies=heap,lru' '--with-maxfd=32768'
>> >>
>> >> I built with --with-maxfd=32768 option but, when squid is started it
>>
>> says
>>
>> >> is working with only 1024 filedescriptor.
>> >>
>> >> I even added the following to the squid.conf:
>> >>
>> >> max_open_disk_fds 0
>> >>
>> >> But it hasn't resolve anything. I'm using squid on Debian Lenny. I
>>
>> don't
>>
>> >> know what to do. Here's part of cache.log:
>> <snip logs>
>>
>> > You got a bug! that behaivor happens when a coredump occurs in squid,
>> > please
>> > file a ticket with gdb output, rice debug at maximum if you can.
>>
>> WTF are you talking about Luis? None of the above problems have anything
>> to do with crashing Squid.
>>
>> They are in order:
>>
>> "WARNING! Your cache is running out of filedescriptors"
>> �* either the system limits being set too low during run-time operation.
>> �* or the system limits were too small during the configure and build
>> process.
>> � �-> Squid may drop new client connections to maintain lower than desired
>> traffic levels.
>>
>> � NP: patching the kernel headers to artificially trick squid into
>> believing the kernel supports more by default than it does is not a good
>> solution. The ulimit utility exists for that purpose instead.
>> <snip kernel patch>
>>
>>
>> "Unsupported method attempted by 172.16.100.83"
>> �* The machine at 172.16.100.83 is pushing non-HTTP data into Squid.
>> � -> Squid will drop these connections.
>>
>> "clientNatLookup: NF getsockopt(SO_ORIGINAL_DST) failed: (2) No such file
>> or directory"
>> �* NAT interception is failing to locate the NAT table entries for some
>> client connection.
>> �* usually due to configuring the same port with "transparent" option and
>> regular traffic.
>> �-> for now Squid will treat these connections as if the directly
>> connecting box was the real client. This WILL change in some near future
>> release.
>>
>>
>> As you can see in none of those handling operations does squid crash or
>> core dump.
>>
>>
>> Amos
>
>
> Amos, that is the exactly behaivor I did have with a bug, dont you remember
> the DIGEST bug that makes squid restart internaly? HNO did help me, but the
> fact is that is a symptom of a coredump internal restart because he complains
> his sq is already compilled with more than 1024.
>
> After retarting, I did have 1024 descriptors, no matters i did compile with
> 64k FDs.

As I said -

The running configure / make / compile environment has to be set to
64k file descriptors. The build environment's max file descriptors
are an overriding limit on the actual usable FDs, no matter what you
set the configure maxfd value to. If ulimit -n = 1024 at configure
time, that's what you're stuck at.

# ulimit -HSn 32768 (or 64k) ; ./configure (options...) ; make

-- 
-george william herbert
george.herbert_at_gmail.com