[PATCH rdma-next v1 0/7] Use ib_umem_find_best_pgsz() for all umems

[PATCH rdma-next v1 0/7] Use ib_umem_find_best_pgsz() for all umems

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

Changelog:
v1:
 * Added patch for raw QP
 * Fixed commit messages
v0: https://lore.kernel.org/lkml/20201026132635.1337663-1-leon@kernel.org

-------------------------
From Jason:

Move the remaining cases working with umems to use versions of
ib_umem_find_best_pgsz() tailored to the calculations the devices
requires.

Unlike a MR there is no IOVA, instead a page offset from the starting page
is possible, with various restrictions.

Compute the best page size to meet the page_offset restrictions.

Thanks

Jason Gunthorpe (7):
  RDMA/mlx5: Use ib_umem_find_best_pgoff() for SRQ
  RDMA/mlx5: Use mlx5_umem_find_best_quantized_pgoff() for WQ
  RDMA/mlx5: Directly compute the PAS list for raw QP RQ's
  RDMA/mlx5: Use mlx5_umem_find_best_quantized_pgoff() for QP
  RDMA/mlx5: mlx5_umem_find_best_quantized_pgoff() for CQ
  RDMA/mlx5: Use ib_umem_find_best_pgsz() for devx
  RDMA/mlx5: Lower setting the umem's PAS for SRQ

 drivers/infiniband/hw/mlx5/cq.c      |  48 +++++---
 drivers/infiniband/hw/mlx5/devx.c    |  66 ++++++-----
 drivers/infiniband/hw/mlx5/mem.c     | 115 +++++++------------
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  47 +++++++-
 drivers/infiniband/hw/mlx5/qp.c      | 165 ++++++++++++---------------
 drivers/infiniband/hw/mlx5/srq.c     |  27 +----
 drivers/infiniband/hw/mlx5/srq.h     |   1 +
 drivers/infiniband/hw/mlx5/srq_cmd.c |  80 ++++++++++++-
 include/rdma/ib_umem.h               |  42 +++++++
 9 files changed, 343 insertions(+), 248 deletions(-)

--
2.28.0

Re: [PATCH rdma-next v1 0/7] Use ib_umem_find_best_pgsz() for all umems

[Jason Gunthorpe]

On Sun, Nov 15, 2020 at 01:43:04PM +0200, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> Changelog:
> v1:
>  * Added patch for raw QP
>  * Fixed commit messages
> v0: https://lore.kernel.org/lkml/20201026132635.1337663-1-leon@kernel.org
> 
> -------------------------
> >From Jason:
> 
> Move the remaining cases working with umems to use versions of
> ib_umem_find_best_pgsz() tailored to the calculations the devices
> requires.
> 
> Unlike a MR there is no IOVA, instead a page offset from the starting page
> is possible, with various restrictions.
> 
> Compute the best page size to meet the page_offset restrictions.
> 
> Thanks
> 
> Jason Gunthorpe (7):
>   RDMA/mlx5: Use ib_umem_find_best_pgoff() for SRQ
>   RDMA/mlx5: Use mlx5_umem_find_best_quantized_pgoff() for WQ
>   RDMA/mlx5: Directly compute the PAS list for raw QP RQ's
>   RDMA/mlx5: Use mlx5_umem_find_best_quantized_pgoff() for QP
>   RDMA/mlx5: mlx5_umem_find_best_quantized_pgoff() for CQ
>   RDMA/mlx5: Use ib_umem_find_best_pgsz() for devx
>   RDMA/mlx5: Lower setting the umem's PAS for SRQ

Applied to for-next

Jason