UCT/CM/RDMACM: share dummy CQ per device #6016
Conversation
|
@evgeny-leksikov was it clean CP? |
no, there were 2 conflicts around dev_name parameter |
|
bot:pipe:retest |
src/uct/ib/rdmacm/rdmacm_cm.c
Outdated
| @@ -624,6 +674,8 @@ UCS_CLASS_CLEANUP_FUNC(uct_rdmacm_cm_t) | |||
|
|
|||
| ucs_trace("destroying event_channel %p on cm %p", self->ev_ch, self); | |||
| rdma_destroy_event_channel(self->ev_ch); | |||
|
|
|||
src/uct/ib/rdmacm/rdmacm_cm.h
Outdated
| #include <uct/base/uct_cm.h> | ||
| #include <ucs/datastruct/khash.h> | ||
|
|
||
| #include <infiniband/verbs.h> |
src/uct/ib/rdmacm/rdmacm_cm.c
Outdated
| khiter_t iter; | ||
| int ret; | ||
|
|
||
| iter = kh_put(uct_rdmacm_cm_cqs, &cm->cqs, device_name, &ret); |
There was a problem hiding this comment.
How does it work?? looks like the hash table key can point to memory on the stack (device_name->dev_name)
IMO better use rdma_cm_id::pd pointer as the key
Any chance it's using stack garbage as the key and actually creating new CQ every time?
evgeny-leksikov
force-pushed
the
uct_rdmacm_cm_shared_cq_m
branch
from
2309b48 to
7d38a90
Compare
|
re-cherry-picked squashed yosefe#105 + yosefe#113 to avoid conflicts which were appeared due to wrong approach with |
|
|
||
| /* create a dummy qp in order to get a unique qp_num to provide to librdmacm */ | ||
| status = uct_rdmacm_cm_create_dummy_cq_qp(cep->id, &cq, &qp); | ||
| status = uct_rdmacm_cm_create_dummy_qp(cep->id, cq, &cep->qp); |
There was a problem hiding this comment.
maybe call uct_rdmacm_cm_get_cq from uct_rdmacm_cm_create_dummy_qp()?
There was a problem hiding this comment.
I would like to avoid hidden side effect of such function, let's keep things simpler
What
share dummy CQ per device
Why ?
to reduce resources consumption
How ?
ported from yosefe#105