Date:   Fri, 25 Nov 2022 11:59:46 +0100
From:   Alexandra Winter <wintera@...ux.ibm.com>
To:     Tony Lu <tonylu@...ux.alibaba.com>,
        Jan Karcher <jaka@...ux.ibm.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>,
        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] net/smc: Fix expected buffersizes and sync logic



On 25.11.22 08:05, Tony Lu wrote:
> On Fri, Nov 25, 2022 at 07:15:33AM +0100, Jan Karcher wrote:
>>
>>
>> On 24/11/2022 15:07, Alexandra Winter wrote:
>>>
>>>
>>> On 24.11.22 14:00, Alexandra Winter wrote:
>>>>
>>>>
>> [ ... ]>>>>> On Wed, Nov 23, 2022 at 11:49:07AM +0100, Jan Karcher wrote:
>>>>>>>> The fixed commit changed the expected behavior of buffersizes
>>>>>>>> set by the user using the setsockopt mechanism.
>>>>>>>> Before the fixed patch the logic for determining the buffersizes used
>>>>>>>> was the following:
>>>>>>>>
>>>>>>>> default  = net.ipv4.tcp_{w|r}mem[1]
>>> Jan, you explained to me: "the minima is 16Kib. This is enforced in smc_compress_bufsize
>>> which would move any value <= 16Kib into bucket 0 - which is 16KiB "
>>> net.ipv4.tcp_wmem[1] defaults to 8Kib. So in the default case (unchanged net.ipv4.tcp_wmem[1])
>>> the default for the send path is not net.ipv4.tcp_wmem[1]. Should be clarified here.
>>
>> The default value is still set to the net.ipv4.tcp_{w|r}mem[1]. This is a
>> *very* top level overview about what is happening and *not* a documentation.
>> I don't really want to explain the full code flow here.
>>
>> What we still should do - as Tony aggreed on - is documenting the SMC
>> behavior. This is a follow up on my list.
> 
> Hello Jan and Alexandra,
> 
> It looks like the misalignment of information is causing some trouble,
> which is introduced by my patch. Maybe we could have an off-maillist and
> online meeting to discussion?
> 
> We have some progress updates of scalability, and we are really like the
> extension of SMC-D. Also we have some ideas for SMC, in case of
> misalignment of information, we'd like to put them on the table and
> discuss them earlier. Maybe an online meeting is an efficient way. What
> do you think?
> 
> If possible, I would prepared the meetings and topics and send them to
> everyone first.
> 
> Cheers,
> Tony Lu
> 

Thanks a lot for your constructive proposals Tony. Yes, we should have a discussion off-mailinglist
about future topics.

My remaining concern for this fix is the default values (user does not use setsockopt, nor 
changes the new sysfs parameters, nor changes tcp defaults):
>>>> before 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock and make them tunable")
	    send: 16k recv: 64k
>>>> after net/smc: Fix expected buffersizes and sync logic   (this patch)
>>>>       send: 16k recv: 128k

@Jan, as this is the only patch you want to send to net, please change the default size of
the receive buffers back to 64k (I don't care how).


>>
>>>>>>>> sockopt  = the setsockopt mechanism
>>>>>>>> val      = the value assigned in default or via setsockopt
>>>>>>>> sk_buf   = short for sk_{snd|rcv}buf
>>>>>>>> real_buf = the real size of the buffer (sk_buf_size in __smc_buf_create)
>>>>>>>>
>>>>>>>>     exposed   | net/core/sock.c  |    af_smc.c    |  smc_core.c
>>>>>>>>               |                  |                |
>>>>>>>> +---------+ |                  | +------------+ | +-------------------+
>>>>>>>> | default |----------------------| sk_buf=val |---| real_buf=sk_buf/2 |
>>>>>>>> +---------+ |                  | +------------+ | +-------------------+
>>>>>>>>               |                  |                |    ^
>>>>>>>>               |                  |                |    |
>>>>>>>> +---------+ | +--------------+ |                |    |
>>>>>>>> | sockopt |---| sk_buf=val*2 |-----------------------|
>>>>>>>> +---------+ | +--------------+ |                |
>>>>>>>>               |                  |                |
>>>>>>>>
>>>>>>>> The fixed patch introduced a dedicated sysctl for smc
>>>>>>>> and removed the /2 in smc_core.c resulting in the following flow:
>>>>>>>>
>>>>>>>> default  = net.smc.{w|r}mem (which defaults to net.ipv4.tcp_{w|r}mem[1])
>>>>>>>> sockopt  = the setsockopt mechanism
>>>>>>>> val      = the value assigned in default or via setsockopt
>>>>>>>> sk_buf   = short for sk_{snd|rcv}buf
>>>>>>>> real_buf = the real size of the buffer (sk_buf_size in __smc_buf_create)
>>>>>>>>
>>>>>>>>     exposed   | net/core/sock.c  |    af_smc.c    |  smc_core.c
>>>>>>>>               |                  |                |
>>>>>>>> +---------+ |                  | +------------+ | +-----------------+
>>>>>>>> | default |----------------------| sk_buf=val |---| real_buf=sk_buf |
>>>>>>>> +---------+ |                  | +------------+ | +-----------------+
>>>>>>>>               |                  |                |    ^
>>>>>>>>               |                  |                |    |
>>>>>>>> +---------+ | +--------------+ |                |    |
>>>>>>>> | sockopt |---| sk_buf=val*2 |-----------------------|
>>>>>>>> +---------+ | +--------------+ |                |
>>>>>>>>               |                  |                |
>>>>>>>>
>>>>>>>> This would result in double of memory used for existing configurations
>>>>>>>> that are using setsockopt.
>>>>>>>
>>>>>>> Firstly, thanks for your detailed diagrams :-)
>>>>>>>
>>>>>>> And the original decision to use user-provided values rather than
>>>>>>> value/2 to follow the instructions of the socket manual [1].
>>>>>>>
>>>>>>>     SO_RCVBUF
>>>>>>>            Sets or gets the maximum socket receive buffer in bytes.
>>>>>>>            The kernel doubles this value (to allow space for
>>>>>>>            bookkeeping overhead) when it is set using setsockopt(2),
>>>>>>>            and this doubled value is returned by getsockopt(2).  The
>>>>>>>            default value is set by the
>>>>>>>            /proc/sys/net/core/rmem_default file, and the maximum
>>>>>>>            allowed value is set by the /proc/sys/net/core/rmem_max
>>>>>>>            file.  The minimum (doubled) value for this option is 256.
>>>>>>>
>>>>>>> [1] https://man7.org/linux/man-pages/man7/socket.7.html
>>>>>>>
>>>>>>> The user of SMC should know that setsockopt() with SO_{RCV|SND}BUF will
>>>>>>
>>>>>> I totally agree that an educated user of SMC should know about that behavior
>>>>>> if they decide to use it.
>>>>>> We do provide our users preload libraries where they can pass preferred
>>>>>> buffersizes via arguments and we handle the Sockopts for them.
>>>>>>
>>>>>>> double the values in kernel, and getsockopt() will return the doubled
>>>>>>> values. So that they should use half of the values which are passed to
>>>>>>> setsockopt(). The original patch tries to make things easier in SMC and
>>>>>>> let user-space to handle them following the socket manual.
>>>>>>>
>>>>>>>> SMC historically decided to use the explicit value given by the user
>>>>>>>> to allocate the memory. This is why we used the /2 in smc_core.c.
>>>>>>>> That logic was not applied to the default value.
>>>>>>>
>>>>>>> Yep, let back to the patch which introduced smc_{w|r}mem knobs, it's a
>>>>>>> trade-off to follow original logic of SMC, or follow the socket manual.
>>>>>>> We decides to follow the instruction of manuals in the end.
>>>>>>
>>>>>> I understand the point. I spend a lot of time trying to decide what to do.
>>>>>>
>>>>>> Since it was an intentional decision to not follow the general socket
>>>>>> option, and we do not have anyone complaining we do not really have a reason
>>>>>> to change it.
>>>>>> Changing it means that users with existing configurations would have to
>>>>>> change their configs on an update or suddenly expect double the memory
>>>>>> consumption.
>>>>>> That's why we in the end preffered to stay with the current logic.
>>>>>
>>>>> I can't agree with you more with the points to follow the historic logic
>>>>> and not break the user-space applications.
>>>>>
>>>>>> I'm thinking that maybe - if we stay with the historic logic - we should
>>>>>> document that desicion somewhere. So that in the future, if a user that
>>>>>> expects the man page behavior, has a way to understand what SMC is doing.
>>>>>> What do oyu think?
>>>>>
>>>>> Yep, we _really_ need to document it if we change the convention.
>>>>> Actually, I spent a lot of time to find the history about the logic of
>>>>> buffer (/2 and *2) in SMC. So I'm really in favor of adding
>>>>> documentation, at least code comments to help others to understand them.
>>>>>
>>>>> Cheers,
>>>>> Tony Lu
>>>> Iiuc you are changing the default values in this a patch and your other patch:
>>>> Default values for real_buf for send and receive:
>>>>
>>>> before 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock and make them tunable")
>>>>      real_buf=net.ipv4.tcp_{w|r}mem[1]/2   send: 8k  recv: 64k
>>>        see above: 			    send: 16k recv: 64k
>>>> after 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock and make them tunable")
>>>> real_buf=net.ipv4.tcp_{w|r}mem[1]   send: 16k (16*1024) recv: 128k (131072)
>>>>
>>>> after net/smc: Fix expected buffersizes and sync logic
>>>> real_buf=net.ipv4.tcp_{w|r}mem[1]   send: 16k (16*1024) recv: 128k (131072)
>>>>
>>>> after net/smc: Unbind smc control from tcp control
>>>> real_buf=SMC_*BUF_INIT_SIZE   send: 16k (16384) recv: 64k (65536)
>>>>
>>>> If my understanding is correct, then I nack this.
>>>> Defaults should be restored to the values before 0227f058aa29.
>>>> Otherwise users will notice a change in memory usage that needs to
>>>> be avoided or announced more explicitely. (and don't change them twice)
>>> See above, I stand corrected. However this patch fixes/restores the buffersize
>>> for setsockopt, but not for the default recieve path.
>>> Could you please clarify that in the title and description?
>>>
>>
>> I am trying to keep the commit title as crisp as possible while providing
>> enough information and set the context in the commit message:
>>
>> "The fixed commit changed the expected behavior of buffersizes set by the
>> user using the setsockopt mechanism."
>>
>>  + There is now a whole e-mail thread to consult in case of any further
>> questions.
>>
>> Thank you for your comments
>> - Jan
>>
>>> Reviewed-by: Alexandra Winter <wintera@...ux.ibm.com>
>>>>>> - Jan
>>>>>>
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Tony Lu
>>>>>>>
>>>>>>>> Since we now have our own sysctl, which is also exposed to the user,
>>>>>>>> we should sync the logic in a way that both values are the real value
>>>>>>>> used by our code and shown by smc_stats. To achieve this this patch
>>>>>>>> changes the behavior to:
>>>>>>>>
>>>>>>>> default  = net.smc.{w|r}mem (which defaults to net.ipv4.tcp_{w|r}mem[1])
>>>>>>>> sockopt  = the setsockopt mechanism
>>>>>>>> val      = the value assigned in default or via setsockopt
>>>>>>>> sk_buf   = short for sk_{snd|rcv}buf
>>>>>>>> real_buf = the real size of the buffer (sk_buf_size in __smc_buf_create)
>>>>>>>>
>>>>>>>>     exposed   | net/core/sock.c  |    af_smc.c     |  smc_core.c
>>>>>>>>               |                  |                 |
>>>>>>>> +---------+ |                  | +-------------+ | +-----------------+
>>>>>>>> | default |----------------------| sk_buf=val*2|---|real_buf=sk_buf/2|
>>>>>>>> +---------+ |                  | +-------------+ | +-----------------+
>>>>>>>>               |                  |                 |    ^
>>>>>>>>               |                  |                 |    |
>>>>>>>> +---------+ | +--------------+ |                 |    |
>>>>>>>> | sockopt |---| sk_buf=val*2 |------------------------|
>>>>>>>> +---------+ | +--------------+ |                 |
>>>>>>>>               |                  |                 |
>>>>>>>>
>>>>>>>> This way both paths follow the same pattern and the expected behavior
>>>>>>>> is re-established.
>>>>>>>>
>>>>>>>> Fixes: 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock and make them tunable")
>>>>>>>> Signed-off-by: Jan Karcher <jaka@...ux.ibm.com>
>>>>>>>> Reviewed-by: Wenjia Zhang <wenjia@...ux.ibm.com>
>>>>>>>> ---
>>>>>>>>    net/smc/af_smc.c   | 9 +++++++--
>>>>>>>>    net/smc/smc_core.c | 8 ++++----
>>>>>>>>    2 files changed, 11 insertions(+), 6 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
>>>>>>>> index 036532cf39aa..a8c84e7bac99 100644
>>>>>>>> --- a/net/smc/af_smc.c
>>>>>>>> +++ b/net/smc/af_smc.c
>>>>>>>> @@ -366,6 +366,7 @@ static void smc_destruct(struct sock *sk)
>>>>>>>>    static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
>>>>>>>>    				   int protocol)
>>>>>>>>    {
>>>>>>>> +	int buffersize_without_overhead;
>>>>>>>>    	struct smc_sock *smc;
>>>>>>>>    	struct proto *prot;
>>>>>>>>    	struct sock *sk;
>>>>>>>> @@ -379,8 +380,12 @@ static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
>>>>>>>>    	sk->sk_state = SMC_INIT;
>>>>>>>>    	sk->sk_destruct = smc_destruct;
>>>>>>>>    	sk->sk_protocol = protocol;
>>>>>>>> -	WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(net->smc.sysctl_wmem));
>>>>>>>> -	WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(net->smc.sysctl_rmem));
>>>>>>>> +	buffersize_without_overhead =
>>>>>>>> +		min_t(int, READ_ONCE(net->smc.sysctl_wmem), INT_MAX / 2);
>>>>>>>> +	WRITE_ONCE(sk->sk_sndbuf, buffersize_without_overhead * 2);
>>>>>>>> +	buffersize_without_overhead =
>>>>>>>> +		min_t(int, READ_ONCE(net->smc.sysctl_rmem), INT_MAX / 2);
>>>>>>>> +	WRITE_ONCE(sk->sk_rcvbuf, buffersize_without_overhead * 2);
>>>>>>>>    	smc = smc_sk(sk);
>>>>>>>>    	INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
>>>>>>>>    	INIT_WORK(&smc->connect_work, smc_connect_work);
>>>>>>>> diff --git a/net/smc/smc_core.c b/net/smc/smc_core.c
>>>>>>>> index 00fb352c2765..36850a2ae167 100644
>>>>>>>> --- a/net/smc/smc_core.c
>>>>>>>> +++ b/net/smc/smc_core.c
>>>>>>>> @@ -2314,10 +2314,10 @@ static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
>>>>>>>>    	if (is_rmb)
>>>>>>>>    		/* use socket recv buffer size (w/o overhead) as start value */
>>>>>>>> -		sk_buf_size = smc->sk.sk_rcvbuf;
>>>>>>>> +		sk_buf_size = smc->sk.sk_rcvbuf / 2;
>>>>>>>>    	else
>>>>>>>>    		/* use socket send buffer size (w/o overhead) as start value */
>>>>>>>> -		sk_buf_size = smc->sk.sk_sndbuf;
>>>>>>>> +		sk_buf_size = smc->sk.sk_sndbuf / 2;
>>>>>>>>    	for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
>>>>>>>>    	     bufsize_short >= 0; bufsize_short--) {
>>>>>>>> @@ -2376,7 +2376,7 @@ static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
>>>>>>>>    	if (is_rmb) {
>>>>>>>>    		conn->rmb_desc = buf_desc;
>>>>>>>>    		conn->rmbe_size_short = bufsize_short;
>>>>>>>> -		smc->sk.sk_rcvbuf = bufsize;
>>>>>>>> +		smc->sk.sk_rcvbuf = bufsize * 2;
>>>>>>>>    		atomic_set(&conn->bytes_to_rcv, 0);
>>>>>>>>    		conn->rmbe_update_limit =
>>>>>>>>    			smc_rmb_wnd_update_limit(buf_desc->len);
>>>>>>>> @@ -2384,7 +2384,7 @@ static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
>>>>>>>>    			smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
>>>>>>>>    	} else {
>>>>>>>>    		conn->sndbuf_desc = buf_desc;
>>>>>>>> -		smc->sk.sk_sndbuf = bufsize;
>>>>>>>> +		smc->sk.sk_sndbuf = bufsize * 2;
>>>>>>>>    		atomic_set(&conn->sndbuf_space, bufsize);
>>>>>>>>    	}
>>>>>>>>    	return 0;
>>>>>>>> -- 
>>>>>>>> 2.34.1

Powered by blists - more mailing lists