prov/efa: Cleanup EP/CQ#12167
Conversation
alekswn
force-pushed
the
cleanup-ep-cq
branch
2 times, most recently
from
4a71eae to
c70bce1
Compare
|
bot:aws:retest |
alekswn
force-pushed
the
cleanup-ep-cq
branch
3 times, most recently
from
3d8239c to
314b561
Compare
Remove dead DGRAM-specific ops structs that are no longer referenced after the endpoint unification. Signed-off-by: Alexey Novikov <nalexey@amazon.com>
…ect_ep struct Extract ope_pool and ope_list from efa_base_ep into a new efa_direct_ep struct (defined in efa_direct_ep.h) that wraps efa_base_ep as its first member for container_of castability. This separates efa-direct (DGRAM) endpoint state from the shared base, keeping efa_base_ep clean for both efa-direct and efa-protocol paths. Guard container_of casts with EFA_INFO_TYPE_IS_RDM checks where the base_ep_list contains mixed endpoint types. Signed-off-by: Alexey Novikov <nalexey@amazon.com>
Add test_efa_base_ep_construct_ibv_qp_init_attr_ex_dgram to verify QP initialization attributes for DGRAM endpoints, complementing the existing RDM test. Signed-off-by: Alexey Novikov <nalexey@amazon.com>
Add the new efa_direct_ep.h header to the list of distributed headers so it is included in release tarballs. Signed-off-by: Alexey Novikov <nalexey@amazon.com>
|
Generally, some parts are missing here. The goal of the refactor is to eliminate the fabric condition in the common efa_base_ep code. I am pretty sure there are some still there. like libfabric/prov/efa/src/efa_base_ep.c Line 406 in 7367b95 |
| struct efa_direct_ep { | ||
| struct efa_base_ep base_ep; | ||
|
|
||
| struct ofi_bufpool *ope_pool; /**< pool for efa_direct_ope */ |
There was a problem hiding this comment.
Have a comment here that we make the ope_pool/ope_list additional for efa_direct_ep because of a need to track the outstanding operations that still reference MRs. This should be a temporary addition and finally removed or moved to the efa_base_ep when we have a unified approach to make it cover both efa and efa-direct
There was a problem hiding this comment.
Thinking about it more, I think we should keep ope_pool and ope_list in the base_ep and make efa_rdm_ep use them. efa_rdm_ep also have a ope_pool which is the same type. efa_rdm_ep has separate txe_list and rxe_list but they are only used for final cleanup purpose. There is no reason we cannot combine them into a single ope_list and scan it once and clean up based on the ope type accordingly
| prov/efa/src/efa_cq.h \ | ||
| prov/efa/src/efa_cntr.h \ | ||
| prov/efa/src/efa_base_ep.h \ | ||
| prov/efa/src/efa_direct_ep.h \ |
There was a problem hiding this comment.
this should have been part of your first commit
2 participants
Uh oh!
There was an error while loading. Please reload this page.