prov/shm: add iface parameter to smr_select_proto #8849
Conversation
| { | ||
| if (op == ofi_op_read_req) { | ||
| if (use_ipc) | ||
| return smr_src_ipc; | ||
| if (cma_avail && FI_HMEM_SYSTEM) |
There was a problem hiding this comment.
I guess cma_avail && FI_HMEM_SYSTEM does not make much sense. Are we trying to test cma_avail && FI_HMEM_SYSTEM == iface instead?
There was a problem hiding this comment.
Yeah definitely a bug that this fixes. Thank you!
There was a problem hiding this comment.
Thanks for confirming!
| { | ||
| if (op == ofi_op_read_req) { | ||
| if (use_ipc) | ||
| return smr_src_ipc; | ||
| if (cma_avail && FI_HMEM_SYSTEM) |
There was a problem hiding this comment.
Woah, that's a big miss. Can you add something in the commit message about how this fixes a bug where CMA would never be selected for a read?
There was a problem hiding this comment.
Thanks. Updated commit message. I will backport the change to 1.18.x branch.
Do we need it for 1.17.x as well?
There was a problem hiding this comment.
Well the only real use case of shm is efa so I'll let you decide how far back you want to go. The issue technically goes all the way back to 1.15.x
There was a problem hiding this comment.
Cool then we only need to backport to 1.18.x
Pass through hmem iface to proto selection logic. This change re-enables CMA for read ops. Signed-off-by: Wenduo Wang <wenduwan@amazon.com>
wenduwan
force-pushed
the
shm_gdrcopy_aware
branch
from
3f4c274
to
c1de6b8
Compare
|
bot:aws:retest |
|
Please back port |
Pass through hmem iface to proto selection logic.