[ewg] Re: [PATCH] NFS-RDMA backport for RHEL 5.2

[Vladimir Sokolovsky]

Jon Mason wrote:
> On Mon, Mar 02, 2009 at 05:37:45PM +0200, Vladimir Sokolovsky wrote:
>   
>> Jon Mason wrote:
>>     
>>> Hey Vlad,
>>>
>>> I wanted to get you the NFS-RDMA backport patches I have queued up prior
>>> to rc1 being built.  I have this patch (RHEL5.2), as well as 2.6.22, and
>>> 2.6.25 (which I will be sending in separate e-mails).
>>>
>>> This patch contains the changes necessary to get NFS-RDMA 99% working on
>>> RHEL5.2.  It 100% passes of the connectathon testcases when using RDMA.
>>> When using TCP, 1 testcase hangs (for which I've found a code hack to
>>> work around, but not a fix).  The testcase is the 30MB transfer in the
>>> "special" tests, and is 100% reproducible.
>>>
>>> With the attached patch, you can now enable the nfs-rdma-mod by default
>>> and run it through your nightly build tests.
>>>
>>> There is one problem with nfs-utils, which will necessitate including it
>>> in the OFED distribution.  Currently, there is a check for the kernel
>>> version you are running when running the mount.nfs command.  This check
>>> is to verify if one is running a kernel greater than 2.6.22.  If not,
>>> then it will not allow a mount with RDMA as an option.  In earlier
>>> versions, there was an override option of "-i".  This has since been
>>> removed for an unknown reason (but if you run `mount.nfs`, you can see
>>> that they forgot to remove 'i' from "usage" options).  I will send them
>>> a patch to fix this in the latest version, but this will be a problem
>>> that will require it to be built locally until such a time as that is
>>> pulled in.
>>>
>>> In the OFED tree, there is another problem that needs to be addressed.
>>> There are kernel header files that are being automatically imported via
>>> the OFED tree.  The redefinition of these kernel data structs causes
>>> major problems.  Specifically, these header files are
>>> include/linux/pipe_fs_i.h and include/linux/splice.h.  These files
>>> redefine pipe_inode_info and pipe_buf_operations and many other structs
>>> used in OFED and the underlying kernel.  They should NOT be included,
>>> and can cause problems when backporting NFS-RDMA to other kernel
>>> versions.  I have removed their usage in my patches, but they should be
>>> removed from the OFED tree as they should not be needed (as we want to
>>> use the native header files for the kernel we are compiling against).
>>>
>>> Thanks,
>>> Jon
>>>   
>>>       
>> Hi Jon,
>> Is there any reason that you removed sg_next function from  
>> kernel_addons/backport/2.6.18-EL5.2/include/linux/scatterlist.h?
>>
>> -static inline struct scatterlist *sg_next(struct scatterlist *sg)
>> -{
>> -       if (!sg) {
>> -               BUG();
>> -               return NULL;
>> -       }
>> -       return sg + 1;
>> -}
>> -
>>     
>
> Oops, I must've removed it on accident.  NFS-RDMA has no need for it not
> to be there.
>
> If you want, I can provide a patch (or a new patchset for NFS-RDMA) to
> fix this issue.
>
> Thanks,
> Jon
I fixed it already.

Regards,
Vladimir