OMPI/OSHMEM: bug-fix: store mkeys for each oshmem ctx. #6443
Conversation
Signed-off-by: Xin Zhao <[email protected]>
|
cherrypick from #6290 |
| ucp_rkey_pack(mca_spml_ucx.ucp_context, mkey->key.mem_h, | ||
| &(new_mkey->u.data), &len); | ||
|
|
||
| ucp_ep_rkey_unpack(ucx_ctx->ucp_peers[pe].ucp_conn, |
There was a problem hiding this comment.
this works only for self, where are keys for other processes unpacked?
|
|
||
| new_mkey->spml_context = new_ucx_mkey; | ||
|
|
||
| ucp_rkey_pack(mca_spml_ucx.ucp_context, mkey->key.mem_h, |
There was a problem hiding this comment.
need to release rkey buf with ucp_rkey_buffer_release sometime
| return module->get_mkey_slow(ctx, pe, va, rva); | ||
| } else { | ||
| uint32_t segno = memheap_find_segnum(va); | ||
| sshmem_mkey_t *new_mkey = (sshmem_mkey_t *)calloc(1, sizeof(*new_mkey)); |
| } | ||
| } else { | ||
| assert(module->get_mkey_slow); | ||
| return module->get_mkey_slow(ctx, pe, va, rva); |
There was a problem hiding this comment.
probably it will not request a new key from the remote, but will return the one cached for default context (if it's not a new mem region)
| memheap_oob.mkeys = mkeys; | ||
| memheap_oob.segno = seg; | ||
| memheap_oob.mkeys_rcvd = 0; | ||
| memheap_oob.ctx = ctx; |
There was a problem hiding this comment.
can it be overwritten if several keys are requested in a row?
|
@bwbarrett Why the @xinzhao3 Could you please add the "Cherry Picked from ... " message to the cherry-pick via the |
|
@gpaulsen because I suck? :) |
|
@xinzhao3 we are planning to do another, and hopefully final rc on v4.0.1 on Monday. So please address his concerns by then, otherwise this PR will have to wait to 4.0.2. |
|
@hppritcha: This PR is stale, should be replaced with #6509 |
|
This needs to be closed. |
7 participants
Signed-off-by: Xin Zhao [email protected]