rds cq event handler issue

rds cq event handler issue

[Steve Wise]

Hey Venkat,

I think I see a bug in the RDS RDMA module where RDS is not adhering to the RDMA locking context.  From the kernel tree 
Documentation/infiniband/core_locking.txt:

---
   The context in which completion event and asynchronous event
   callbacks run is not defined.  Depending on the low-level driver, it
   may be process context, softirq context, or interrupt context.
   Upper level protocol consumers may not sleep in a callback.
---

So RDMA ULPs cannot assume any certain context for their callback functions.  Yet I get a BUG_ON() when running RDS with 
iw_cxgb3 where RDS is bugging in rds_rdma_free_op():

---
                 /* Mark page dirty if it was possibly modified, which
                  * is the case for a RDMA_READ which copies from remote
                  * to local memory */
                 if (!ro->op_write) {
                         BUG_ON(irqs_disabled());
                         set_page_dirty(page);
                 }
---

And rds_rdma_free_op() can be called in the cq callback path.  Here's a stack trace when it bugged:

---
Call Trace:
<IRQ> [<ffffffff886ca0fc>] :rds:rds_message_purge+0x54/0x79
[<ffffffff886ca162>] :rds:rds_message_put+0x41/0x4c
[<ffffffff886f616b>] :rds_rdma:rds_iw_send_unmap_rm+0xe2/0xf2
[<ffffffff886f63c4>] :rds_rdma:rds_iw_send_cq_comp_handler+0x193/0x2e5
[<ffffffff88698a56>] :iw_cxgb3:iwch_ev_dispatch+0x1df/0x2b1
[<ffffffff8869f0b2>] :iw_cxgb3:cxio_hal_ev_handler+0x6b/0xb4
[<ffffffff882746cd>] :cxgb3:process_rx+0x3d/0xa0
[<ffffffff8827b28c>] :cxgb3:process_responses+0x120c/0x1350
---

iwch_ev_dispatch() explicitly disables irqs to ensure proper serialization:

---
         spin_lock_irqsave(&chp->comp_handler_lock, flag);
         (*chp->ibcq.comp_handler)(&chp->ibcq, chp->ibcq.cq_context);
         spin_unlock_irqrestore(&chp->comp_handler_lock, flag);
---

I'm not sure if that BUG_ON() in rds_rdma_free_op() is valid or not.  If it is valid, then RDS needs to run this logic 
in a safe context, not in the context of the CQ callback.  It BUG_ON() is not valid, we can remove it :).

Can you comment?

Thanks,

Steve.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: rds cq event handler issue

[Venkat Venkatsubra]

On 3/6/2012 12:05 PM, Steve Wise wrote:
> Hey Venkat,
>
> I think I see a bug in the RDS RDMA module where RDS is not adhering 
> to the RDMA locking context.  From the kernel tree 
> Documentation/infiniband/core_locking.txt:
>
> ---
>   The context in which completion event and asynchronous event
>   callbacks run is not defined.  Depending on the low-level driver, it
>   may be process context, softirq context, or interrupt context.
>   Upper level protocol consumers may not sleep in a callback.
> ---
>
> So RDMA ULPs cannot assume any certain context for their callback 
> functions.  Yet I get a BUG_ON() when running RDS with iw_cxgb3 where 
> RDS is bugging in rds_rdma_free_op():
>
> ---
>                 /* Mark page dirty if it was possibly modified, which
>                  * is the case for a RDMA_READ which copies from remote
>                  * to local memory */
>                 if (!ro->op_write) {
>                         BUG_ON(irqs_disabled());
>                         set_page_dirty(page);
>                 }
> ---
>
> And rds_rdma_free_op() can be called in the cq callback path.  Here's 
> a stack trace when it bugged:
>
> ---
> Call Trace:
> <IRQ> [<ffffffff886ca0fc>] :rds:rds_message_purge+0x54/0x79
> [<ffffffff886ca162>] :rds:rds_message_put+0x41/0x4c
> [<ffffffff886f616b>] :rds_rdma:rds_iw_send_unmap_rm+0xe2/0xf2
> [<ffffffff886f63c4>] :rds_rdma:rds_iw_send_cq_comp_handler+0x193/0x2e5
> [<ffffffff88698a56>] :iw_cxgb3:iwch_ev_dispatch+0x1df/0x2b1
> [<ffffffff8869f0b2>] :iw_cxgb3:cxio_hal_ev_handler+0x6b/0xb4
> [<ffffffff882746cd>] :cxgb3:process_rx+0x3d/0xa0
> [<ffffffff8827b28c>] :cxgb3:process_responses+0x120c/0x1350
> ---
>
> iwch_ev_dispatch() explicitly disables irqs to ensure proper 
> serialization:
>
> ---
>         spin_lock_irqsave(&chp->comp_handler_lock, flag);
>         (*chp->ibcq.comp_handler)(&chp->ibcq, chp->ibcq.cq_context);
>         spin_unlock_irqrestore(&chp->comp_handler_lock, flag);
> ---
>
> I'm not sure if that BUG_ON() in rds_rdma_free_op() is valid or not.  
> If it is valid, then RDS needs to run this logic in a safe context, 
> not in the context of the CQ callback.  It BUG_ON() is not valid, we 
> can remove it :).
>
> Can you comment?
>
> Thanks,
>
> Steve.
>
Hi Steve,

Our internal latest code has a WARN_ON instead:
---------------------
                 /* Mark page dirty if it was possibly modified, which
                  * is the case for a RDMA_READ which copies from remote
                  * to local memory */
                 if (!ro->op_write) {
                         WARN_ON_ONCE(page_mapping(page) && 
irqs_disabled());
                         set_page_dirty(page);
                 }
---------------------

Venkat
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: rds cq event handler issue

[Steve Wise]

On 03/06/2012 12:58 PM, Venkat Venkatsubra wrote:
> On 3/6/2012 12:05 PM, Steve Wise wrote:
>> Hey Venkat,
>>
>> I think I see a bug in the RDS RDMA module where RDS is not adhering to the RDMA locking context.  From the kernel 
>> tree Documentation/infiniband/core_locking.txt:
>>
>> ---
>>   The context in which completion event and asynchronous event
>>   callbacks run is not defined.  Depending on the low-level driver, it
>>   may be process context, softirq context, or interrupt context.
>>   Upper level protocol consumers may not sleep in a callback.
>> ---
>>
>> So RDMA ULPs cannot assume any certain context for their callback functions.  Yet I get a BUG_ON() when running RDS 
>> with iw_cxgb3 where RDS is bugging in rds_rdma_free_op():
>>
>> ---
>>                 /* Mark page dirty if it was possibly modified, which
>>                  * is the case for a RDMA_READ which copies from remote
>>                  * to local memory */
>>                 if (!ro->op_write) {
>>                         BUG_ON(irqs_disabled());
>>                         set_page_dirty(page);
>>                 }
>> ---
>>
>> And rds_rdma_free_op() can be called in the cq callback path.  Here's a stack trace when it bugged:
>>
>> ---
>> Call Trace:
>> <IRQ> [<ffffffff886ca0fc>] :rds:rds_message_purge+0x54/0x79
>> [<ffffffff886ca162>] :rds:rds_message_put+0x41/0x4c
>> [<ffffffff886f616b>] :rds_rdma:rds_iw_send_unmap_rm+0xe2/0xf2
>> [<ffffffff886f63c4>] :rds_rdma:rds_iw_send_cq_comp_handler+0x193/0x2e5
>> [<ffffffff88698a56>] :iw_cxgb3:iwch_ev_dispatch+0x1df/0x2b1
>> [<ffffffff8869f0b2>] :iw_cxgb3:cxio_hal_ev_handler+0x6b/0xb4
>> [<ffffffff882746cd>] :cxgb3:process_rx+0x3d/0xa0
>> [<ffffffff8827b28c>] :cxgb3:process_responses+0x120c/0x1350
>> ---
>>
>> iwch_ev_dispatch() explicitly disables irqs to ensure proper serialization:
>>
>> ---
>>         spin_lock_irqsave(&chp->comp_handler_lock, flag);
>>         (*chp->ibcq.comp_handler)(&chp->ibcq, chp->ibcq.cq_context);
>>         spin_unlock_irqrestore(&chp->comp_handler_lock, flag);
>> ---
>>
>> I'm not sure if that BUG_ON() in rds_rdma_free_op() is valid or not.  If it is valid, then RDS needs to run this 
>> logic in a safe context, not in the context of the CQ callback.  It BUG_ON() is not valid, we can remove it :).
>>
>> Can you comment?
>>
>> Thanks,
>>
>> Steve.
>>
> Hi Steve,
>
> Our internal latest code has a WARN_ON instead:
> ---------------------
>                 /* Mark page dirty if it was possibly modified, which
>                  * is the case for a RDMA_READ which copies from remote
>                  * to local memory */
>                 if (!ro->op_write) {
>                         WARN_ON_ONCE(page_mapping(page) && irqs_disabled());
>                         set_page_dirty(page);
>                 }
> ---------------------
>
> Venkat

That helps with the crashing. :)   Does set_page_dirty() require irqs enabled?  If so, RDS needs to change such that it 
doesn't do this work in the CQ event handler callback context.

Steve.


--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html