Re: [PATCHv5 for-rc1 v5 4/8] RDMA/rxe: Implement dellink in rxe

[Zhu Yanjun <yanjun.zhu@linux.dev>]

在 2023/6/21 11:23, Bob Pearson 写道:
> On 6/20/23 21:13, Zhu Yanjun wrote:
>> 在 2023/6/21 4:21, Bob Pearson 写道:
>>> On 4/28/23 04:39, Zhu Yanjun wrote:
>>>> From: Zhu Yanjun <yanjun.zhu@linux.dev>
>>>>
>>>> When running "rdma link del" command, dellink function will be called.
>>>> If the sock refcnt is greater than the refcnt needed for udp tunnel,
>>>> the sock refcnt will be decreased by 1.
>>>>
>>>> If equal, the last rdma link is removed. The udp tunnel will be
>>>> destroyed.
>>>>
>>>> Tested-by: Rain River <rain.1986.08.12@gmail.com>
>>>> Signed-off-by: Zhu Yanjun <yanjun.zhu@linux.dev>
>>>> ---
>>>>    drivers/infiniband/sw/rxe/rxe.c     | 12 +++++++++++-
>>>>    drivers/infiniband/sw/rxe/rxe_net.c | 17 +++++++++++++++--
>>>>    drivers/infiniband/sw/rxe/rxe_net.h |  1 +
>>>>    3 files changed, 27 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/infiniband/sw/rxe/rxe.c b/drivers/infiniband/sw/rxe/rxe.c
>>>> index 0ce6adb43cfc..ebfabc6d6b76 100644
>>>> --- a/drivers/infiniband/sw/rxe/rxe.c
>>>> +++ b/drivers/infiniband/sw/rxe/rxe.c
>>>> @@ -166,10 +166,12 @@ void rxe_set_mtu(struct rxe_dev *rxe, unsigned int ndev_mtu)
>>>>    /* called by ifc layer to create new rxe device.
>>>>     * The caller should allocate memory for rxe by calling ib_alloc_device.
>>>>     */
>>>> +static struct rdma_link_ops rxe_link_ops;
>>>>    int rxe_add(struct rxe_dev *rxe, unsigned int mtu, const char *ibdev_name)
>>>>    {
>>>>        rxe_init(rxe);
>>>>        rxe_set_mtu(rxe, mtu);
>>>> +    rxe->ib_dev.link_ops = &rxe_link_ops;
>>>>          return rxe_register_device(rxe, ibdev_name);
>>>>    }
>>>> @@ -206,9 +208,17 @@ static int rxe_newlink(const char *ibdev_name, struct net_device *ndev)
>>>>        return err;
>>>>    }
>>>>    -struct rdma_link_ops rxe_link_ops = {
>>>> +static int rxe_dellink(struct ib_device *dev)
>>>> +{
>>>> +    rxe_net_del(dev);
>>>> +
>>>> +    return 0;
>>>> +}
>>>> +
>>>> +static struct rdma_link_ops rxe_link_ops = {
>>>>        .type = "rxe",
>>>>        .newlink = rxe_newlink,
>>>> +    .dellink = rxe_dellink,
>>>>    };
>>>>      static int __init rxe_module_init(void)
>>>> diff --git a/drivers/infiniband/sw/rxe/rxe_net.c b/drivers/infiniband/sw/rxe/rxe_net.c
>>>> index 3ca92e062800..4cc7de7b115b 100644
>>>> --- a/drivers/infiniband/sw/rxe/rxe_net.c
>>>> +++ b/drivers/infiniband/sw/rxe/rxe_net.c
>>>> @@ -530,6 +530,21 @@ int rxe_net_add(const char *ibdev_name, struct net_device *ndev)
>>>>        return 0;
>>>>    }
>>>>    +#define SK_REF_FOR_TUNNEL    2
>>>> +void rxe_net_del(struct ib_device *dev)
>>>> +{
>>>> +    if (refcount_read(&recv_sockets.sk6->sk->sk_refcnt) > SK_REF_FOR_TUNNEL)
>>>> +        __sock_put(recv_sockets.sk6->sk);
>>>> +    else
>>>> +        rxe_release_udp_tunnel(recv_sockets.sk6);
>>>> +
>>>> +    if (refcount_read(&recv_sockets.sk4->sk->sk_refcnt) > SK_REF_FOR_TUNNEL)
>>>> +        __sock_put(recv_sockets.sk4->sk);
>>>> +    else
>>>> +        rxe_release_udp_tunnel(recv_sockets.sk4);
>>>> +}
>>>> +#undef SK_REF_FOR_TUNNEL
>>>> +
>>>>    static void rxe_port_event(struct rxe_dev *rxe,
>>>>                   enum ib_event_type event)
>>>>    {
>>>> @@ -689,8 +704,6 @@ int rxe_register_notifier(void)
>>>>      void rxe_net_exit(void)
>>>>    {
>>>> -    rxe_release_udp_tunnel(recv_sockets.sk6);
>>>> -    rxe_release_udp_tunnel(recv_sockets.sk4);
>>>>        unregister_netdevice_notifier(&rxe_net_notifier);
>>>>    }
>>> These calls are moved to rxe_net_del which is called by an explicit unlink command.
>>> But if rxe_net_init fails and returns an error code this will never happen.
>>> This will result in leaking resources.
>> Thanks a lot. Bob.
>>
>> Sure, if ipv6 tunnel fails to be created, the resource related with ipv4 should be released.
>>
>> I will fix it in the latest version.
>>
>> Zhu Yanjun
> I haven't had a chance to test netns yet. I am sure it works but I will test it.
Yes. Please. It is an interesting feature.
> The only other thing I noticed are some stylistic differences with the rest of
> the rxe driver. You use
>
> struct rxe_dev *rdev;
>
> elsewhere it is
>
> struct rxe_dev *rxe;
>
> Yours is more like the mlx drivers where they use dev for ib_device and mdev for mlx_device.
> rxe tries to use ibdev ibqp, ibmr, etc for the ib objects and no prefix for the driver
> specific ones. It's less typing that way.


Got you. I think we should use rxe instead of rdev. I will fix it in the 
latest commits.


>
> With a couple of exceptions all the printk's are now in the form
>
> rxe_[type]_[obj](obj, "message", ...) or rxe_[type] if there isn't an obj to refer to.
>
> where type = info, err, warn, or dbg and obj = rxe, ah, pd, qp, cq, etc. These are basically
> adapted from the siw driver.

If I can get you correctly, you mean that we should use rxe_dbg_qp, .... 
to replace pr_err ....

I have questions:

1). What benefit will this bring?

2). If the log is in module __init or  module __exit functions, we 
should use pr_xxx? Because obj does not exist in these __init and __exit 
functions.

Best Regards,

Zhu Yanjun

>
> Regards,
>
> Bob
>
>>> Bob
>>>>    diff --git a/drivers/infiniband/sw/rxe/rxe_net.h b/drivers/infiniband/sw/rxe/rxe_net.h
>>>> index a222c3eeae12..f48f22f3353b 100644
>>>> --- a/drivers/infiniband/sw/rxe/rxe_net.h
>>>> +++ b/drivers/infiniband/sw/rxe/rxe_net.h
>>>> @@ -17,6 +17,7 @@ struct rxe_recv_sockets {
>>>>    };
>>>>      int rxe_net_add(const char *ibdev_name, struct net_device *ndev);
>>>> +void rxe_net_del(struct ib_device *dev);
>>>>      int rxe_register_notifier(void);
>>>>    int rxe_net_init(void);