[v1] hw/rdma: Last step in eliminating data-path processing

[PATCH v1 0/2] hw/rdma: Last step in eliminating data-path processing

Yuval Shaia posted 2 patches 4 years, 1 month ago

Diff against v1
Download series mbox

Test docker-quick@centos7 passed

Test docker-mingw@fedora passed

Test checkpatch passed

Test FreeBSD passed

Test asan passed

Patches applied successfully (tree, apply log)
git fetch https://github.com/patchew-project/qemu tags/patchew/20200320143429.9490-1-yuval.shaia.ml@gmail.com

Maintainers: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>, Yuval Shaia <yuval.shaia.ml@gmail.com>

hw/rdma/rdma_backend.c | 61 +++++++++++++++++++++---------------------
hw/rdma/rdma_backend.h |  5 ----
hw/rdma/rdma_rm.c      | 13 +++++----
3 files changed, 36 insertions(+), 43 deletions(-)

Expand all Fold all

[PATCH v1 0/2] hw/rdma: Last step in eliminating data-path processing

Posted by Yuval Shaia 4 years, 1 month ago

With emulated device we have the luxury of doing things in data-path.
Considering a future virtio-rdma device we no longer have this luxury. The
driver will directly post WQEs to HW queues, bypassing the device emulator
in qemu. These WQEs will consist of addresses known to the device (even if
it is guest addresses) and ID of an HW memory region (mr).

Commit 68b89aee71 ("Utilize ibv_reg_mr_iova for memory registration") did
the first important step of utilizing a new rdma API so addresses
translation is no longer needed.
This patch-set continues and remove entirely the processing in data-path by
eliminating the need to translate emulated mr_id to backend device mr_id.

v0 -> v1:
	* Accept comment from Marcel

Yuval Shaia (2):
  hw/rdma: Cosmetic change - no need for two sge arrays
  hw/rdma: Skip data-path mr_id translation

 hw/rdma/rdma_backend.c | 61 +++++++++++++++++++++---------------------
 hw/rdma/rdma_backend.h |  5 ----
 hw/rdma/rdma_rm.c      | 13 +++++----
 3 files changed, 36 insertions(+), 43 deletions(-)

-- 
2.20.1