mbox series

[resend,for-next,0/5] RTRS enable write path fast memory regitration

Message ID 20210621055340.11789-1-jinpu.wang@ionos.com (mailing list archive)
Headers show
Series RTRS enable write path fast memory regitration | expand

Message

Jinpu Wang June 21, 2021, 5:53 a.m. UTC
Hi Jason, hi Doug, hi Jens

Please consider to include following changes to the next merge window.

This enables fast memory registration for write IO patch, so rtrs can
support bigger IO than 116k without splitting. With this in place, both
read/write request are more symmetric, and we can also reduce the memory
usage.

The patchset is orgnized as:
- patch1 preparation.
- patch2 implement fast memory registration for write patch.
- patch3 reduce memory usage.
- patch4 raise MAX_SGEMENTs
- patch5 rnbd-clt to query and use the max_sgements setting.

As the main change is in RTRS, so it's easier to go through RDMA tree, hence
send this patchset to linux-rdma.

This is a rebased on the top latest rdma/wip/jgg-for-next commit
7e78dd816e45 ("RDMA/hns: Clear extended doorbell info before using")

v1: https://lore.kernel.org/linux-rdma/20210608113536.42965-1-jinpu.wang@ionos.com/T/#t

Thanks!

Jack Wang (5):
  RDMA/rtrs: Introduce head/tail wr
  RDMA/rtrs-clt: Write path fast memory registration
  RDMA/rtrs_clt: Alloc less memory with write path fast memory
    registration
  RDMA/rtrs-clt: Raise MAX_SEGMENTS
  rnbd/rtrs-clt: Query and use max_segments from rtrs-clt.

 drivers/block/rnbd/rnbd-clt.c          |   5 +-
 drivers/block/rnbd/rnbd-clt.h          |   5 +-
 drivers/infiniband/ulp/rtrs/rtrs-clt.c | 143 ++++++++++++++++---------
 drivers/infiniband/ulp/rtrs/rtrs-clt.h |   1 +
 drivers/infiniband/ulp/rtrs/rtrs-pri.h |   3 +-
 drivers/infiniband/ulp/rtrs/rtrs.c     |  28 ++---
 drivers/infiniband/ulp/rtrs/rtrs.h     |   2 +-
 7 files changed, 119 insertions(+), 68 deletions(-)

Comments

Jason Gunthorpe June 22, 2021, 12:03 a.m. UTC | #1
On Mon, Jun 21, 2021 at 07:53:35AM +0200, Jack Wang wrote:
> Hi Jason, hi Doug, hi Jens
> 
> Please consider to include following changes to the next merge window.
> 
> This enables fast memory registration for write IO patch, so rtrs can
> support bigger IO than 116k without splitting. With this in place, both
> read/write request are more symmetric, and we can also reduce the memory
> usage.
> 
> The patchset is orgnized as:
> - patch1 preparation.
> - patch2 implement fast memory registration for write patch.
> - patch3 reduce memory usage.
> - patch4 raise MAX_SGEMENTs
> - patch5 rnbd-clt to query and use the max_sgements setting.
> 
> As the main change is in RTRS, so it's easier to go through RDMA tree, hence
> send this patchset to linux-rdma.
> 
> This is a rebased on the top latest rdma/wip/jgg-for-next commit
> 7e78dd816e45 ("RDMA/hns: Clear extended doorbell info before using")
> 
> v1: https://lore.kernel.org/linux-rdma/20210608113536.42965-1-jinpu.wang@ionos.com/T/#t
> 
> Thanks!
> 
> Jack Wang (5):
>   RDMA/rtrs: Introduce head/tail wr
>   RDMA/rtrs-clt: Write path fast memory registration
>   RDMA/rtrs_clt: Alloc less memory with write path fast memory
>     registration
>   RDMA/rtrs-clt: Raise MAX_SEGMENTS
>   rnbd/rtrs-clt: Query and use max_segments from rtrs-clt.

Applied to for-next, thanks

Jason