netdev - Re: [PATCH net-next] net/smc: Unbind smc control from tcp control

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <21f0ecf6-6154-78c0-7866-bfb4212ead99@linux.ibm.com>
Date:   Wed, 23 Nov 2022 14:36:56 +0100
From:   Jan Karcher <jaka@...ux.ibm.com>
To:     Tony Lu <tonylu@...ux.alibaba.com>
Cc:     David Miller <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>, netdev@...r.kernel.org,
        linux-s390@...r.kernel.org, Heiko Carstens <hca@...ux.ibm.com>,
        Alexandra Winter <wintera@...ux.ibm.com>,
        Wenjia Zhang <wenjia@...ux.ibm.com>,
        Thorsten Winkler <twinkler@...ux.ibm.com>,
        Stefan Raspl <raspl@...ux.ibm.com>,
        Karsten Graul <kgraul@...ux.ibm.com>
Subject: Re: [PATCH net-next] net/smc: Unbind smc control from tcp control



On 23/11/2022 13:07, Tony Lu wrote:
> On Wed, Nov 23, 2022 at 11:58:30AM +0100, Jan Karcher wrote:
>> In the past SMC used the values of tcp_{w|r}mem to create the send
>> buffer and RMB. We now have our own sysctl knobs to tune them without
>> influencing the TCP default.
>>
>> This patch removes the dependency on the TCP control by providing our
>> own initial values which aim for a low memory footprint.
>>
>> Signed-off-by: Jan Karcher <jaka@...ux.ibm.com>
>> Reviewed-by: Wenjia Zhang <wenjia@...ux.ibm.com>
>> ---
>>   Documentation/networking/smc-sysctl.rst |  4 ++--
>>   net/smc/smc_core.h                      |  6 ++++--
>>   net/smc/smc_sysctl.c                    | 10 ++++++----
>>   3 files changed, 12 insertions(+), 8 deletions(-)
>>
>> diff --git a/Documentation/networking/smc-sysctl.rst b/Documentation/networking/smc-sysctl.rst
>> index 6d8acdbe9be1..a1c634d3690a 100644
>> --- a/Documentation/networking/smc-sysctl.rst
>> +++ b/Documentation/networking/smc-sysctl.rst
>> @@ -44,7 +44,7 @@ smcr_testlink_time - INTEGER
>>   
>>   wmem - INTEGER
>>   	Initial size of send buffer used by SMC sockets.
>> -	The default value inherits from net.ipv4.tcp_wmem[1].
>> +	The default value aims for a small memory footprint and is set to 16KiB.
>>   
>>   	The minimum value is 16KiB and there is no hard limit for max value, but
>>   	only allowed 512KiB for SMC-R and 1MiB for SMC-D.
>> @@ -53,7 +53,7 @@ wmem - INTEGER
>>   
>>   rmem - INTEGER
>>   	Initial size of receive buffer (RMB) used by SMC sockets.
>> -	The default value inherits from net.ipv4.tcp_rmem[1].
>> +	The default value aims for a small memory footprint and is set to 64KiB.
>>   
>>   	The minimum value is 16KiB and there is no hard limit for max value, but
>>   	only allowed 512KiB for SMC-R and 1MiB for SMC-D.
>> diff --git a/net/smc/smc_core.h b/net/smc/smc_core.h
>> index 285f9bd8e232..67c3937f341d 100644
>> --- a/net/smc/smc_core.h
>> +++ b/net/smc/smc_core.h
>> @@ -206,8 +206,10 @@ struct smc_rtoken {				/* address/key of remote RMB */
>>   	u32			rkey;
>>   };
>>   
>> -#define SMC_BUF_MIN_SIZE	16384	/* minimum size of an RMB */
> 
> Hi Jan,
> 
> This patch inspired me that the min value of RMB and sndbuffer is 16KiB,
> it means that every connection costs 32KiB at least. It's still a large
> size for small environments, such as virtual machines or containers.
> 
> Also we have tested some cases with smaller buffer size (4KiB, with
> hacked code), it also shows good performance compared with larger buffer
> size.
> 
> So I am wondering that we could reduce the min value of RMB/send buffer,
> such as 4KiB.

That sounds interesting.
We did not think about reducing the minimum value.
One thing I'm wondering is if other OS like z/OS have own architectural 
limits or limits in general that we would have to consider in any way.
Let us look into it if we run into any trouble with lower memory.

- Jan

> 
> Cheers,
> Tony Lu
> 
>> -#define SMC_RMBE_SIZES		16	/* number of distinct RMBE sizes */
>> +#define SMC_SNDBUF_INIT_SIZE 16384 /* initial size of send buffer */
>> +#define SMC_RCVBUF_INIT_SIZE 65536 /* initial size of receive buffer */
>> +#define SMC_BUF_MIN_SIZE	 16384	/* minimum size of an RMB */
>> +#define SMC_RMBE_SIZES		 16	/* number of distinct RMBE sizes */
>>   /* theoretically, the RFC states that largest size would be 512K,
>>    * i.e. compressed 5 and thus 6 sizes (0..5), despite
>>    * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
>> diff --git a/net/smc/smc_sysctl.c b/net/smc/smc_sysctl.c
>> index b6f79fabb9d3..a63aa79d4856 100644
>> --- a/net/smc/smc_sysctl.c
>> +++ b/net/smc/smc_sysctl.c
>> @@ -19,8 +19,10 @@
>>   #include "smc_llc.h"
>>   #include "smc_sysctl.h"
>>   
>> -static int min_sndbuf = SMC_BUF_MIN_SIZE;
>> -static int min_rcvbuf = SMC_BUF_MIN_SIZE;
>> +static int initial_sndbuf	= SMC_SNDBUF_INIT_SIZE;
>> +static int initial_rcvbuf	= SMC_RCVBUF_INIT_SIZE;
>> +static int min_sndbuf		= SMC_BUF_MIN_SIZE;
>> +static int min_rcvbuf		= SMC_BUF_MIN_SIZE;
>>   
>>   static struct ctl_table smc_table[] = {
>>   	{
>> @@ -88,8 +90,8 @@ int __net_init smc_sysctl_net_init(struct net *net)
>>   	net->smc.sysctl_autocorking_size = SMC_AUTOCORKING_DEFAULT_SIZE;
>>   	net->smc.sysctl_smcr_buf_type = SMCR_PHYS_CONT_BUFS;
>>   	net->smc.sysctl_smcr_testlink_time = SMC_LLC_TESTLINK_DEFAULT_TIME;
>> -	WRITE_ONCE(net->smc.sysctl_wmem, READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]));
>> -	WRITE_ONCE(net->smc.sysctl_rmem, READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]));
>> +	WRITE_ONCE(net->smc.sysctl_wmem, initial_sndbuf);
>> +	WRITE_ONCE(net->smc.sysctl_rmem, initial_rcvbuf);
>>   
>>   	return 0;
>>   
>> -- 
>> 2.34.1