Re: [PATCH v1 1/4] xprtrdma: Remove support for FMR memory registration

[Chuck Lever <chuck.lever@oracle.com>]

> On Nov 19, 2018, at 5:41 PM, Jason Gunthorpe <jgg@ziepe.ca> wrote:
> 
> On Mon, Nov 19, 2018 at 08:16:37AM -0800, Bart Van Assche wrote:
>> On Mon, 2018-11-19 at 10:45 -0500, Chuck Lever wrote:
>>> FMR is not supported on most recent RDMA devices. It is also slower
>>> and less secure than FRWR. As discussed during the RDMA BoF at LPC
>>> 2018, it is time to remove support for FMR in the NFS/RDMA client
>>> stack. NFS/RDMA server-side uses either local memory registration or
>>> FRWR.  There is no change required there to deprecate FMR.
>>> 
>>> There are a few Infiniband/RoCE devices in the kernel tree that do
>>> not support MEM_MGT_EXTENSIONS, and therefore will no longer support
>>> client-side NFS/RDMA. These are:
>>> 
>>> - mthca
>>> - qib
> 
> Ooh, qib was not in the list when we talked on this in plumbers. That
> does change things. 
> 
> Dennis: Could qib be changed? HFI1 supports FRWR right?
> 
>>> - usnic
>>> - hns (RoCE)
>> 
>> Can someone from Intel comment on how many of their customers rely on the qib
>> driver?
>> 
>> Can someone from Cisco comment on how many of their customers rely on the usnic
>> driver?
> 
> Ignore usnic, it is not a RDMA driver
> 
>> Can someone from Huawei comment on how many of their customers rely on the hns
>> driver?
> 
> Nor was this, this is an actively used driver..
> 
> Could hns do FRWR? It is very disappointing that such a new driver
> does not support FRWR.
> 
> I guess it is worth stating again that FMR is deprecated. People are
> going to be disappointed if they try to submit new drivers that don't
> support FRWR :(

My methodology was to search for whether each driver asserts
IB_DEVICE_MEM_MGT_EXTENSIONS in device_cap_flags.

Interestingly, nes and ocrdma assert it, but qib and hns do not.


--
Chuck Lever