[PATCH v4 net-next 06/21] nvme-tcp: Add DDP offload control path

[Boris Pismenny]

On 17/02/2021 15:55, Or Gerlitz wrote:
> On Sun, Feb 14, 2021 at 8:20 PM David Ahern <dsahern at gmail.com> wrote:
>> On 2/11/21 2:10 PM, Boris Pismenny wrote:
>>> @@ -223,6 +229,164 @@ static inline size_t nvme_tcp_pdu_last_send(struct nvme_tcp_request *req,
>>>       return nvme_tcp_pdu_data_left(req) <= len;
>>>  }
>>>
>>> +#ifdef CONFIG_TCP_DDP
>>> +
>>> +static bool nvme_tcp_resync_request(struct sock *sk, u32 seq, u32 flags);
>>> +static const struct tcp_ddp_ulp_ops nvme_tcp_ddp_ulp_ops = {
>>> +     .resync_request         = nvme_tcp_resync_request,
>>> +};
>>> +
>>> +static int nvme_tcp_offload_socket(struct nvme_tcp_queue *queue)
>>> +{
>>> +     struct net_device *netdev = queue->ctrl->offloading_netdev;
>>> +     struct nvme_tcp_ddp_config config = {};
>>> +     int ret;
>>> +
>>> +     if (!(netdev->features & NETIF_F_HW_TCP_DDP))
>>
>> If nvme_tcp_offload_limits does not find a dst_entry on the socket then
>> offloading_netdev may not NULL at this point.
> 
> correct :( will look on that
> 

That's only partially true.
If nvme_tcp_offload_limits finds a dst_entry, but then the netdevice
goes down, then the check here will catch it. This is needed because
nvme_tcp_offload_limits doesn't hold a reference! We opted not to grab a
reference on nvme_tcp_offload_limits because it doesn't create a context.


>>> +             queue->ctrl->offloading_netdev = NULL;
>>> +             return -ENODEV;
>>> +     }
>>> +
>>> +     if (netdev->features & NETIF_F_HW_TCP_DDP &&
>>> +         netdev->tcp_ddp_ops &&
>>> +         netdev->tcp_ddp_ops->tcp_ddp_limits)
>>> +             ret = netdev->tcp_ddp_ops->tcp_ddp_limits(netdev, &limits);
>>> +     else
>>> +             ret = -EOPNOTSUPP;
>>> +
>>> +     if (!ret) {
>>> +             queue->ctrl->offloading_netdev = netdev;
>>
>> you save a reference to the netdev here, but then release the refcnt
>> below. That device could be deleted between this point in time and the
>> initialization of all queues.
> 

That's true, and this is why we repeat the checks there.

We avoid holding the reference here because there is no obvious
complementary release point for nvme_tcp_offload_limits and there is no
hardware context created here, so there is no real need to hold it at
this stage.