Adapt ko2iblnd to latest RDMA changes
(LU-8874)
|
|
| Status: | Open |
| Project: | Lustre |
| Component/s: | None |
| Affects Version/s: | Lustre 2.10.0, Upstream |
| Fix Version/s: | Upstream |
| Type: | Technical task | Priority: | Major |
| Reporter: | Doug Oucharek (Inactive) | Assignee: | Amir Shehata (Inactive) |
| Resolution: | Unresolved | Votes: | 0 |
| Labels: | lnet | ||
| Issue Links: |
|
||||||||||||
| Rank (Obsolete): | 9223372036854775807 | ||||||||||||
| Description |
|
Start using these RDMA APIs and delete the ko2iblnd equivalents:
These three routines represent a new strategy for memory pools so there is a lot of ko2iblnd routines which will no longer be needed when these 3 calls get utilized. Code which maps/unmaps memory and pool creation routines all need to be evaluated and deleted if no longer needed. |
| Comments |
| Comment by Doug Oucharek (Inactive) [ 03/Jun/17 ] |
|
Whenever I ask questions about the RDMA API, I am told it is not documented so just do what the iSer driver does. Sigh. James: would you agree that we should "duplicate" the iSer driver's iser_memory.c approach to differentiating between fmr and fastreg and how it makes use of the new RDMA calls? Doing this would get us away from being "so different" from the other RDMA users who all seem to just copy iSer anyway. We would keep our existing runtime model (i.e. our scheduler), but would get us to use the RDMA data structures in exactly the same way as iSer. |