Re: [Question]: reqsk table size limited to 16?

[Eric Dumazet <eric.dumazet@gmail.com>]

Gerrit Renker a écrit :
> Can someone please have a look, it may be that I am missing something?
> 
> It seems that in the following the maximum number of table entries is set
> to always 16, despite sysctl_max_syn_backlog (tcp_max_syn_backlog), 
> overriding the 'backlog' parameter to listen(2).

False alarm ;)

> 
> net/core/request_sock.c
> -----------------------
> 
> int reqsk_queue_alloc(struct request_sock_queue *queue,
>                       unsigned int nr_table_entries)
> {
>         size_t lopt_size = sizeof(struct listen_sock);
>         struct listen_sock *lopt;
> 
> 	nr_table_entries = min_t(u32, nr_table_entries, sysctl_max_syn_backlog);

Here we take the _minimum_ value.
If you have  nr_table_entries=4096 and sysctl_max_syn_backlog=1024,
result is 1024

>         nr_table_entries = max_t(u32, nr_table_entries, 8);

Here we take the _maximum_ value of nr_table_entries and 8

-> 1024

Deal is : We want at least 8 slots, even if users called listen(fd, 1);

(Later, user can change its mind and call listen(fd, 1024).

We dont resize hashtable yet, so we guarantee at least 8 slots fot pathological cases.

>         nr_table_entries = roundup_pow_of_two(nr_table_entries + 1);
> 
> 	//...
> 	for (lopt->max_qlen_log = 3;
>              (1 << lopt->max_qlen_log) < nr_table_entries;
>              lopt->max_qlen_log++);
> 
>  	//...
> 	lopt->nr_table_entries = nr_table_entries;
> 	
> 	//...
> 	return 0
> }
> 
> The function is called with an argument 'nr_table_entries', which is then clamped as
> 
>    sysctl_max_syn_backlog <= nr_table_entries <= 8
> 
> If nr_table_entries = 8, then round_pow_of_two(8 + 1) = 16.
> 
> The sysctl value is set to a much higher value (default 128 or 1024, net/ipv4/tcp.c).
> 
> The reqsk_queue_alloc() gets 'nr_table_entries' passed directly from inet_csk_listen_start(),
> which in turn gets its 'nr_table_entries' as the 'backlog' argument to listen(2) via
>  * net/dccp/proto.c   (dccp_listen_start) or
>  * net/ipv4/af_inet.c (inet_listen).