Re: Kernel oops/panic with NFS over RDMA mount after disrupted Infiniband connection

[sagi grimberg <sagig@mellanox.com>]

On 3/29/2014 3:05 AM, Chuck Lever wrote:
> On Mar 28, 2014, at 4:06 PM, sagi grimberg <sagig@mellanox.com> wrote:
>
>> On 3/29/2014 1:30 AM, Chuck Lever wrote:
>>> On Mar 28, 2014, at 2:42 AM, Senn Klemens <klemens.senn@ims.co.at> wrote:
>>>
>>>> Hi Chuck,
>>>>
>>>> On 03/27/2014 04:59 PM, Chuck Lever wrote:
>>>>> Hi-
>>>>>
>>>>>
>>>>> On Mar 27, 2014, at 12:53 AM, Reiter Rafael <rafael.reiter@ims.co.at> wrote:
>>>>>
>>>>>> On 03/26/2014 07:15 PM, Chuck Lever wrote:
>>>>>>> Hi Rafael-
>>>>>>>
>>>>>>> I’ll take a look. Can you report your HCA and how you reproduce this issue?
>>>>>> The HCA is Mellanox Technologies MT26428.
>>>>>>
>>>>>> Reproduction:
>>>>>> 1) Mount a directory via NFS/RDMA
>>>>>> mount -t nfs -o port=20049,rdma,vers=4.0,timeo=900 172.16.100.2:/ /mnt/
>>>> An additional "ls /mnt" is needed here (between step 1 and 2)
>>>>
>>>>>> 2) Pull the Infiniband cable or use ibportstate to disrupt the Infiniband connection
>>>>>> 3) ls /mnt
>>>>>> 4) wait 5-30 seconds
>>>>> Thanks for the information.
>>>>>
>>>>> I have that HCA, but I won’t have access to my test systems for a week (traveling). So can you try this:
>>>>>
>>>>> # rpcdebug -m rpc -s trans
>>>>>
>>>>> then reproduce (starting with step 1 above). Some debugging output will appear at the tail of /var/log/messages. Copy it to this thread.
>>>>>
>>>> The output of /var/log/messages is:
>>>>
>>>> [  143.233701] RPC:  1688 xprt_rdma_allocate: size 1112 too large for
>>>> buffer[1024]: prog 100003 vers 4 proc 1
>>>> [  143.233708] RPC:  1688 xprt_rdma_allocate: size 1112, request
>>>> 0xffff88105894c000
>>>> [  143.233715] RPC:  1688 rpcrdma_inline_pullup: pad 0 destp
>>>> 0xffff88105894d7dc len 124 hdrlen 124
>>>> [  143.233718] RPC:       rpcrdma_register_frmr_external: Using frmr
>>>> ffff88084e589260 to map 1 segments
>>>> [  143.233722] RPC:  1688 rpcrdma_create_chunks: reply chunk elem
>>>> 652@0x105894d92c:0xced01 (last)
>>>> [  143.233725] RPC:  1688 rpcrdma_marshal_req: reply chunk: hdrlen 48
>>>> rpclen 124 padlen 0 headerp 0xffff88105894d100 base 0xffff88105894d760
>>>> lkey 0x8000
>>>> [  143.233785] RPC:       rpcrdma_event_process: event rep
>>>> ffff88084e589260 status 0 opcode 8 length 0
>>>> [  177.272397] RPC:       rpcrdma_event_process: event rep
>>>> (null) status C opcode FFFF8808 length 4294967295
>>>> [  177.272649] RPC:       rpcrdma_event_process: event rep
>>>> ffff880848ed0000 status 5 opcode FFFF8808 length 4294936584
>>> The mlx4 provider is returning a WC completion status of
>>> IB_WC_WR_FLUSH_ERR.
>>>
>>>> [  177.272651] RPC:       rpcrdma_event_process: WC opcode -30712 status
>>>> 5, connection lost
>>> -30712 is a bogus WC opcode. So the mlx4 provider is not filling in the
>>> WC opcode. rpcrdma_event_process() thus can’t depend on the contents of
>>> the ib_wc.opcode field when the WC completion status != IB_WC_SUCCESS.
>> Hey Chuck,
>>
>> That is correct, the opcode field in the wc is not reliable in FLUSH errors.
>>
>>> A copy of the opcode reachable from the incoming rpcrdma_rep could be
>>> added, initialized in the forward paths. rpcrdma_event_process() could
>>> use the copy in the error case.
>> How about suppressing completions alltogether for fast_reg and local_inv work requests?
>> if these shall fail you will get an error completion and the QP will transition to error state
>> generating FLUSH_ERR completions for all pending WRs. In this case, you can just ignore
>> flush fast_reg + local_inv errors.
>>
>> see http://marc.info/?l=linux-rdma&m=139047309831997&w=2
> While considering your suggestion, I see that my proposed fix doesn’t work. In the FAST_REG_MR and LOCAL_INV cases, wr_id points to a struct rpcrdma_mw, not a struct rpcrdma_rep. Putting a copy of the opcode in rpcrdma_rep would have no effect. Worse:
>
>>   158         if (IB_WC_SUCCESS != wc->status) {
>>   159                 dprintk("RPC:       %s: WC opcode %d status %X, connection lost\n",
>>   160                         __func__, wc->opcode, wc->status);
>>   161                 rep->rr_len = ~0U;
> Suppose this is an IB_WC_FAST_REG_MR completion, so “rep” here is actually a struct rpcrdma_mw, not a struct rpcrdma_rep. Line 161 pokes 32 one-bits at the top of that struct rpcrdma_mw. If wc->opcode was always usable, we’d at least have to fix that.

<nod>
So for error completions one needs to be careful dereferencing wr_id as 
the opcode might not reliable. it will be better to first identify that 
wr_id is indeed a pointer to rep.

>>   162                 if (wc->opcode != IB_WC_FAST_REG_MR && wc->opcode != IB_WC_LOCAL_INV)
>>   163                         rpcrdma_schedule_tasklet(rep);
>>   164                 return;
>>   165         }
>>   166
>>   167         switch (wc->opcode) {
>>   168         case IB_WC_FAST_REG_MR:
>>   169                 frmr = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
>>   170                 frmr->r.frmr.state = FRMR_IS_VALID;
>>   171                 break;
>
> To make my initial solution work, you’d have to add a field to both struct rpcrdma_mw and struct rpcrdma_rep, and ensure they are at the same offset in both structures. Ewe.
>
> Eliminating completions for FAST_REG_MR and LOCAL_INV might be a preferable way to address this.

Agree, Same applies for MW.

Sagi.