Re: [PATCH net] net/smc: Fix lookup of netdev by using ib_device_get_netdev()

[Leon Romanovsky <leon@kernel.org>]

On Wed, Nov 06, 2024 at 10:24:39AM +0100, Halil Pasic wrote:
> On Tue, 5 Nov 2024 13:23:13 +0200
> Leon Romanovsky <leon@kernel.org> wrote:
> 
> > On Tue, Nov 05, 2024 at 10:50:45AM +0100, Wenjia Zhang wrote:
> > > 
> > > 
> > > On 27.10.24 21:18, Leon Romanovsky wrote:  
> > > > On Fri, Oct 25, 2024 at 09:23:55AM +0200, Wenjia Zhang wrote:  
> > > > > Commit c2261dd76b54 ("RDMA/device: Add ib_device_set_netdev() as
> > > > > an alternative to get_netdev") introduced an API
> > > > > ib_device_get_netdev. The SMC-R variant of the SMC protocol
> > > > > continued to use the old API ib_device_ops.get_netdev() to
> > > > > lookup netdev.  
> > > > 
> > > > I would say that calls to ibdev ops from ULPs was never been right
> > > > thing to do. The ib_device_set_netdev() was introduced for the
> > > > drivers.
> > > > 
> > > > So the whole commit message is not accurate and better to be
> > > > rewritten. 
> > > > > As this commit 8d159eb2117b
> > > > > ("RDMA/mlx5: Use IB set_netdev and get_netdev functions")
> > > > > removed the get_netdev callback from
> > > > > mlx5_ib_dev_common_roce_ops, calling ib_device_ops.get_netdev
> > > > > didn't work any more at least by using a mlx5 device driver.  
> > > > 
> > > > It is not a correct statement too. All modern drivers (for last 5
> > > > years) don't have that .get_netdev() ops, so it is not mlx5
> > > > specific, but another justification to say that SMC-R was doing it
> > > > wrong. 
> > > > > Thus, using ib_device_set_netdev() now became mandatory.  
> > > > 
> > > > ib_device_set_netdev() is mandatory for the drivers, it is nothing
> > > > to do with ULPs.
> > > >   
> > > > > 
> > > > > Replace ib_device_ops.get_netdev() with ib_device_get_netdev().  
> > > > 
> > > > It is too late for me to do proper review for today, but I would
> > > > say that it is worth to pay attention to multiple dev_put() calls
> > > > in the functions around the ib_device_get_netdev().
> > > >   
> > > > > 
> > > > > Fixes: 54903572c23c ("net/smc: allow pnetid-less configuration")
> > > > > Fixes: 8d159eb2117b ("RDMA/mlx5: Use IB set_netdev and
> > > > > get_netdev functions")  
> > > > 
> > > > It is not related to this change Fixes line.
> > > >   
> > > 
> > > Hi Leon,
> > > 
> > > Thank you for the review! I agree that SMC could do better. However,
> > > we should fix it and give enough information and reference on the
> > > changes, since the code has already existed and didn't work with the
> > > old way.   
> > 
> > The code which you change worked by chance and was wrong from day one.
> 
> I absolutely agree with that statement. But please notice that the
> commit date of commit c2261dd76b54 ("RDMA/device: Add
> ib_device_set_netdev() as an alternative to get_netdev") predates the
> commit date of commit 54903572c23c ("net/smc: allow pnetid-less
> configuration") only by 9 days. And before commit c2261dd76b54
> ("RDMA/device: Add ib_device_set_netdev() as an alternative to
> get_netdev") there was no 
> ib_device_get_netdev() AFAICT.

It doesn't make it right.

1. While commit c2261dd76b54 was submitted and discussed, RDMA was not CCed.
2. Author didn't try to add his version of ib_device_get_netdev() as it
is done for all APIs exposed by RDMA core.

> 
> Maybe the two patches crossed mid air so to say.
> 
> > 
> > > I can rewrite the commit message.
> > > 
> > > What about:
> > > "
> > > The SMC-R variant of the SMC protocol still called
> > > ib_device_ops.get_netdev() to lookup netdev. As we used mlx5 device
> > > driver to run SMC-R, it failed to find a device, because in mlx5_ib
> > > the internal net device management for retrieving net devices was
> > > replaced by a common interface ib_device_get_netdev() in commit
> > > 8d159eb2117b ("RDMA/mlx5: Use IB set_netdev and get_netdev
> > > functions"). Thus, replace ib_device_ops.get_netdev() with
> > > ib_device_get_netdev() in SMC. "  
> > 
> >  The SMC-R variant of the SMC protocol used direct call to
> > ib_device_ops.get_netdev() function to lookup netdev. Such direct
> > accesses are not correct for any usage outside of RDMA core code. 
> > 
> 
> I agree, it is not correct since c2261dd76b54 ("RDMA/device: Add
> ib_device_set_netdev() as an alternative to get_netdev").
> 
> Does fs/smb/server/transport_rdma.c qualify as inside of RDMA core code?

RDMA core code is drivers/infiniband/core/*.

> I would guess it is not, and I would not actually mind sending a patch
> but I have trouble figuring out the logic behind  commit ecce70cf17d9
> ("ksmbd: fix missing RDMA-capable flag for IPoIB device in
> ksmbd_rdma_capable_netdev()").

It is strange version of RDMA-CM. All other ULPs use RDMA-CM to avoid
GID, netdev and fabric complexity.

> 
> 
> >  RDMA subsystem provides ib_device_get_netdev() function that works on
> >  all RDMA drivers returns valid netdev with proper locking an reference
> >  counting. The commit 8d159eb2117b ("RDMA/mlx5: Use IB set_netdev and
> > get_netdev functions") exposed that SMC-R didn't use that function.
> > 
> 
> I believe the intention was this all along. I think the commit message
> was written with the idea that 54903572c23c happened before c2261dd76b54
> which is not the case.
> 
> >  So update the SMC-R to use proper API,
> > 
> 
> I believe this is exactly what the patch does! And I agree we need to
> improve on the commit message.
> 
> Regards,
> Halil