Date: Tue, 14 Nov 2023 17:52:44 +0800
From: "D. Wythe" <alibuda@...ux.alibaba.com>
To: Wenjia Zhang <wenjia@...ux.ibm.com>, kgraul@...ux.ibm.com,
 jaka@...ux.ibm.com, wintera@...ux.ibm.com
Cc: kuba@...nel.org, davem@...emloft.net, netdev@...r.kernel.org,
 linux-s390@...r.kernel.org, linux-rdma@...r.kernel.org
Subject: Re: [PATCH net v1] net/smc: avoid data corruption caused by decline



On 11/13/23 6:57 PM, Wenjia Zhang wrote:
>
>
> On 13.11.23 03:50, D. Wythe wrote:
>>
>>
>> On 11/10/23 10:51 AM, D. Wythe wrote:
>>>
>>>
>>> On 11/8/23 9:00 PM, Wenjia Zhang wrote:
>>>>
>>>>
>>>> On 08.11.23 10:48, D. Wythe wrote:
>>>>> From: "D. Wythe" <alibuda@...ux.alibaba.com>
>>>>>
>>>>> We found a data corruption issue during testing of SMC-R on Redis
>>>>> applications.
>>>>>
>>>>> The benchmark has a low probability of reporting a strange error as
>>>>> shown below.
>>>>>
>>>>> "Error: Protocol error, got "\xe2" as reply type byte"
>>>>>
>>>>> Finally, we found that the retrieved error data was as follows:
>>>>>
>>>>> 0xE2 0xD4 0xC3 0xD9 0x04 0x00 0x2C 0x20 0xA6 0x56 0x00 0x16 0x3E 0x0C
>>>>> 0xCB 0x04 0x02 0x01 0x00 0x00 0x20 0x00 0x00 0x00 0x00 0x00 0x00 0x00
>>>>> 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0x00 0xE2
>>>>>
>>>>> It is quite obvious that this is a SMC DECLINE message, which 
>>>>> means that
>>>>> the applications received SMC protocol message.
>>>>> We found that this was caused by the following situations:
>>>>>
>>>>> client            server
>>>>>        proposal
>>>>>     ------------->
>>>>>        accept
>>>>>     <-------------
>>>>>        confirm
>>>>>     ------------->
>>>>> wait confirm
>>>>>
>>>>>      failed llc confirm
>>>>>         x------
>>>>> (after 2s)timeout
>>>>>             wait rsp
>>>>>
>>>>> wait decline
>>>>>
>>>>> (after 1s) timeout
>>>>>             (after 2s) timeout
>>>>>         decline
>>>>>     -------------->
>>>>>         decline
>>>>>     <--------------
>>>>>
>>>>> As a result, a decline message was sent in the implementation, and 
>>>>> this
>>>>> message was read from TCP by the already-fallback connection.
>>>>>
>>>>> This patch double the client timeout as 2x of the server value,
>>>>> With this simple change, the Decline messages should never cross or
>>>>> collide (during Confirm link timeout).
>>>>>
>>>>> This issue requires an immediate solution, since the protocol updates
>>>>> involve a more long-term solution.
>>>>>
>>>>> Fixes: 0fb0b02bd6fd ("net/smc: adapt SMC client code to use the 
>>>>> LLC flow")
>>>>> Signed-off-by: D. Wythe <alibuda@...ux.alibaba.com>
>>>>> ---
>>>>>   net/smc/af_smc.c | 2 +-
>>>>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
>>>>> index abd2667..5b91f55 100644
>>>>> --- a/net/smc/af_smc.c
>>>>> +++ b/net/smc/af_smc.c
>>>>> @@ -599,7 +599,7 @@ static int smcr_clnt_conf_first_link(struct 
>>>>> smc_sock *smc)
>>>>>       int rc;
>>>>>         /* receive CONFIRM LINK request from server over RoCE 
>>>>> fabric */
>>>>> -    qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
>>>>> +    qentry = smc_llc_wait(link->lgr, NULL, 2 * SMC_LLC_WAIT_TIME,
>>>>>                     SMC_LLC_CONFIRM_LINK);
>>>>>       if (!qentry) {
>>>>>           struct smc_clc_msg_decline dclc;
>>>> I'm wondering if the double time (if sufficient) of timeout could 
>>>> be for waiting for CLC_DECLINE on the client's side. i.e.
>>>>
>>>
>>> It depends. We can indeed introduce a sysctl to allow server to 
>>> manager their Confirm Link timeout,
>>> but if there will be protocol updates, this introduction will no 
>>> longer be necessary, and we will
>>> have to maintain it continuously.
>>>
> no, I don't think, either, that we need a sysctl for that.

I am okay about that.

>>> I believe the core of the solution is to ensure that decline 
>>> messages never cross or collide. Increasing
>>> the client's timeout by twice as much as the server's timeout can 
>>> temporarily solve this problem.
>
> I have no objection with that, but my question is why you don't 
> increase the timeout waiting for CLC_DECLINE instead of waiting 
> LLC_Confirm_Link? Shouldn't they have the same effect?
>

Logically speaking, of course, they have the same effect, but there are 
two reasons that i choose to increase LLC timeout here:

1. to avoid DECLINE  cross or collide, we need a bigger time gap, a 
simple math is

     2 ( LLC_Confirm_Link) + 1 (CLC_DECLINE) = 3
     2 (LLC_Confirm_Link)  + 1 * 2 (CLC_DECLINE) = 4
     2 * 2(LLC_Confirm_Link) + 1 (CLC_DECLINE) = 5

Obviously, double the LLC_Confirm_Link will result in more time gaps.

2. increase LLC timeout to allow as many RDMA link as possible to 
succeed, rather than fallback.

D. Wythe

>>> If Jerry's proposed protocol updates are too complex or if there 
>>> won't be any future protocol updates,
>>> it's still not late to let server manager their Confirm Link timeout 
>>> then.
>>>
>>> Best wishes,
>>> D. Wythe
>>>
>>
>> FYI:
>>
>> It seems that my email was not successfully delivered due to some 
>> reasons. Sorry
>> for that.
>>
>> D. Wythe
>>
>>
>
>>>> diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
>>>> index 35ddebae8894..9b1feef1013d 100644
>>>> --- a/net/smc/af_smc.c
>>>> +++ b/net/smc/af_smc.c
>>>> @@ -605,7 +605,7 @@ static int smcr_clnt_conf_first_link(struct 
>>>> smc_sock *smc)
>>>>                 struct smc_clc_msg_decline dclc;
>>>>
>>>>                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
>>>> -                                     SMC_CLC_DECLINE, 
>>>> CLC_WAIT_TIME_SHORT);
>>>> +                                     SMC_CLC_DECLINE, 2 * 
>>>> CLC_WAIT_TIME_SHORT);
>>>>                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
>>>>         }
>>>>         smc_llc_save_peer_uid(qentry);
>>>>
>>>> Because the purpose is to let the server have the control to deline.
>>>
>>

Powered by blists - more mailing lists