Re: listen(2) backlog changes in or around Linux 3.1?

[Venkat Venkatsubra <venkat.x.venkatsubra@oracle.com>]

On 10/15/2012 12:26 PM, enh wrote:
> On Mon, Oct 15, 2012 at 10:12 AM, Venkat Venkatsubra
> <venkat.x.venkatsubra@oracle.com>  wrote:
>> On 10/12/2012 6:40 PM, enh wrote:
>>> i used to use the following hack to unit test connect timeouts: i'd
>>> call listen(2) on a socket and then deliberately connect (backlog + 3)
>>> sockets without accept(2)ing any of the connections. (why 3? because
>>> Stevens told me so, and experiment backed him up. see figure 4.10 in
>>> his UNIX Network Programming.)
>>>
>>> with "old" kernels, 2.6.35-ish to 3.0-ish, this worked great. my next
>>> connect(2) to the same loopback port would hang indefinitely. i could
>>> even unblock the connect by calling accept(2) in another thread. this
>>> was awesome for testing.
>>>
>>> in 3.1 on ARM, 3.2 on x86 (Ubuntu desktop), and 3.4 on ARM, this no
>>> longer works. it doesn't seem to be as simple as "the constant is no
>>> longer 3". my tests are now flaky. sometimes they work like they used
>>> to, and sometimes an extra connect(2) will succeed. (or, if i'm in
>>> non-blocking mode, my poll(2) will return with the non-blocking socket
>>> that's trying to connect now ready.)
>>>
>>> i'm guessing if this changed in 3.1 and is still changed in 3.4,
>>> whatever's changed wasn't an accident. but i haven't been able to find
>>> the right search terms to RTFM. i also finally got around to grepping
>>> the kernel for the "+ 3", but wasn't able to find that. (so i'd be
>>> interested to know where the old behavior came from too.)
>>>
>>> my least worst workaround at the moment is to use one of RFC5737's
>>> test networks, but that requires that the device have a network
>>> connection, otherwise my connect(2)s fail immediately with
>>> ENETUNREACH, which is no use to me. also, unlike my old trick, i've
>>> got no way to suddenly "unblock" a slow connect(2) (this is useful for
>>> unit testing the code that does the poll(2) part of the usual
>>> connect-with-timeout implementation).
>>> https://android-review.googlesource.com/#/c/44563/
>>>
>>> hopefully someone here can shed some light on this? ideally someone
>>> will have a workaround as good as my old trick. i realize i was
>>> relying on undocumented behavior, and i'm happy to have to check
>>> /proc/version and behave appropriately, but i'd really like a way to
>>> keep my unit tests!
>>>
>>> thanks,
>>>    elliott
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe netdev" in
>>> the body of a message to majordomo@vger.kernel.org
>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>> Hi Elliott,
>>
>> In BSD I think the backlog used to be reset to 3/2 times that passed by the
>> user. So, 2 becomes 3.
>> Probably the 1/2 times increase was to accommodate the ones in
>> partial/incomplete queue.
>> In Linux is it possible you were getting the same behavior before the below
>> commit ?
>> Since the check used to be "backlog+1" a 2 will behave as 3 ?
> i don't think so, because with<= 3.0 kernels i used to have a backlog
> of 1 and be able to make _4_ connections before my next connect would
> hang. but this>  to>= change is at least something for me to
> investigate...
>
>> commit 8488df894d05d6fa41c2bd298c335f944bb0e401
>> Author: Wei Dong<weid@np.css.fujitsu.com>
>> Date:   Fri Mar 2 12:37:26 2007 -0800
>>
>>      [NET]: Fix bugs in "Whether sock accept queue is full" checking
>>
>>          when I use linux TCP socket, and find there is a bug in function
>> sk_acceptq_is_full().
>>
>>          When a new SYN comes, TCP module first checks its validation. If
>> valid,
>>      send SYN,ACK to the client and add the sock to the syn hash table. Next
>>      time if received the valid ACK for SYN,ACK from the client. server will
>>      accept this connection and increase the sk->sk_ack_backlog -- which is
>>      done in function tcp_check_req().We check wether acceptq is full in
>>      function tcp_v4_syn_recv_sock().
>>
>>      Consider an example:
>>
>>       After listen(sockfd, 1) system call, sk->sk_max_ack_backlog is set to
>>      1. As we know, sk->sk_ack_backlog is initialized to 0. Assuming accept()
>>      system call is not invoked now.
>>
>>      1. 1st connection comes. invoke sk_acceptq_is_full().
>>       sk->sk_ack_backlog=0 sk->sk_max_ack_backlog=1, function return 0 accept
>> this connection.
>>       Increase the sk->sk_ack_backlog
>>      2. 2nd connection comes. invoke sk_acceptq_is_full().
>>       sk->sk_ack_backlog=1 sk->sk_max_ack_backlog=1, function return 0 accept
>> this connection.
>>       Increase the sk->sk_ack_backlog
>>      3. 3rd connection comes. invoke sk_acceptq_is_full().
>>       sk->sk_ack_backlog=2 sk->sk_max_ack_backlog=1, function return 1.
>> Refuse this connection.
>>
>>      I think it has bugs. after listen system call. sk->sk_max_ack_backlog=1
>>      but now it can accept 2 connections.
>>
>>      Signed-off-by: Wei Dong<weid@np.css.fujitsu.com>
>>      Signed-off-by: David S. Miller<davem@davemloft.net>
>>
>> Venkat
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
Ignore my sk_acceptq_is_full() > and >= changes.
That commit was reverted back by this one :
commit 64a146513f8f12ba204b7bf5cb7e9505594ead42
Author: David S. Miller <davem@sunset.davemloft.net>
Date:   Tue Mar 6 11:21:05 2007 -0800

     [NET]: Revert incorrect accept queue backlog changes.

     This reverts two changes:

     8488df894d05d6fa41c2bd298c335f944bb0e401
     248f06726e866942b3d8ca8f411f9067713b7ff8

     A backlog value of N really does mean allow "N + 1" connections
     to queue to a listening socket.  This allows one to specify
     "0" as the backlog and still get 1 connection.

     Noticed by Gerrit Renker and Rick Jones.

     Signed-off-by: David S. Miller <davem@davemloft.net>

Venkat