> On Aug 23, 2018, at 12:33 AM, Parav Pandit <parav@xxxxxxxxxxxx> wrote:
>
>
>
>> -----Original Message-----
>> From: linux-rdma-owner@xxxxxxxxxxxxxxx <linux-rdma-owner@xxxxxxxxxxxxxxx>
>> On Behalf Of Jason Gunthorpe
>> Sent: Wednesday, August 22, 2018 11:13 PM
>> To: Chuck Lever <chuck.lever@xxxxxxxxxx>
>> Cc: linux-rdma <linux-rdma@xxxxxxxxxxxxxxx>
>> Subject: Re: CM event handler: process or IRQ context?
>>
>> On Tue, Aug 14, 2018 at 10:41:19AM -0400, Chuck Lever wrote:
>>> Hi-
>>>
>>> /**
>>> * rdma_create_id - Create an RDMA identifier.
>>> *
>>> * @net: The network namespace in which to create the new id.
>>> * @event_handler: User callback invoked to report events associated with the
>>> * returned rdma_id.
>>> * @context: User specified context associated with the id.
>>> * @ps: RDMA port space.
>>> * @qp_type: type of queue pair associated with the id.
>>> *
>>> * The id holds a reference on the network namespace until it is destroyed.
>>> */
>>> #define rdma_create_id(net, event_handler, context, ps, qp_type)
>>>
>>> Is it guaranteed that the @event_handler function is called in a
>>> process context? Code audit suggests it's always a delayed worker
>>> thread, thus it would always be safe to use spin_lock_bh in the event
>>> handler function.
>>>
> This is my conclusion too. Rdmacm holds mutex while calling event handler.
> I have code refactor for locking in progress in cma.c (at really slow pace). However I do not intent to change calling context.
>
>>> Same question for the consumer's QP event handler:
>>>
>>> struct ib_qp_init_attr {
>>> void (*event_handler)(struct ib_event *, void *);
>>>
>>> Thanks in advance.
>>
> Qp event handler can be called from non blocking context.
>
>> If you find out please add a comment :)
> Yes. Its worth the addition, apart from context what should/can be done specially with rdma_cm_id.
Thanks!
The obvious place to add a documenting comments for these handlers
is where the consumers pass in their event handler functions. That's
rdma_create_id for the CM event handler, but rdma_create_qp is handed
an attr structure. Not sure where there is a noticeable spot for that
one.
--
Chuck Lever
