Re: [PATCH rdma-next v1 2/4] IB/core: Let IB core distribute cache update events

From: Parav Pandit <parav@mellanox.com>
To: Jason Gunthorpe <jgg@ziepe.ca>, Leon Romanovsky <leon@kernel.org>
Cc: Doug Ledford <dledford@redhat.com>,
	Leon Romanovsky <leonro@mellanox.com>,
	RDMA mailing list <linux-rdma@vger.kernel.org>
Subject: Re: [PATCH rdma-next v1 2/4] IB/core: Let IB core distribute cache update events
Date: Wed, 8 Jan 2020 11:35:30 +0000	[thread overview]
Message-ID: <ff0e5aa8-d931-0270-9aa1-0a8aacd2a253@mellanox.com> (raw)
In-Reply-To: <20200107210230.GA7774@ziepe.ca>

On 1/8/2020 2:32 AM, Jason Gunthorpe wrote:
> On Thu, Dec 12, 2019 at 01:30:22PM +0200, Leon Romanovsky wrote:
> 
>> @@ -2627,7 +2626,11 @@ struct ib_device {
>>  	struct rcu_head rcu_head;
>>
>>  	struct list_head              event_handler_list;
>> -	spinlock_t                    event_handler_lock;
>> +	/* Protects event_handler_list */
>> +	struct rw_semaphore event_handler_rwsem;
>> +
>> +	/* Protects QP's event_handler calls and open_qp list */
>> +	spinlock_t event_handler_lock;
> 
> This only protects the open_qp list really, the event handler call
> doesn't need a spinlock. So lets name it properly. open_list_lock ?
> 
Yes. it protects open_qp list and event handler is called for each list
item. So it doesn't really need to protect event handler calls.

> It is sort of weird that we globally serialize all the qp event
> handlers? ie that this lock isn't in the ib_qp.
> 
It probably isn't in each ib_qp because ib_qp is in hundred thousands
and xrc qp events are not so frequent event that can get contented.
So I think per device qp list lock seems fine.