prov/efa: Map EFA errnos to Libfabric codes #9974
Conversation
prov/efa/src/efa_errno.h
Outdated
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_LKEY: | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_UNRESP_REMOTE: | ||
| case FI_EFA_ERR_ESTABLISHED_RECV_UNRESP: | ||
| return FI_EOPBADSTATE; |
There was a problem hiding this comment.
*-FI_EOPBADSTATE*
: The endpoint's state does not permit the requested operation.
I think it doesn't apply to the error here
prov/efa/src/efa_errno.h
Outdated
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_AH: | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_LKEY: | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_UNRESP_REMOTE: | ||
| case FI_EFA_ERR_ESTABLISHED_RECV_UNRESP: |
There was a problem hiding this comment.
FI_EFA_ERR_ESTABLISHED_RECV_UNRESP should be ECONNABOUTED
| case EFA_IO_COMP_STATUS_OK: | ||
| return FI_SUCCESS; | ||
| case EFA_IO_COMP_STATUS_FLUSHED: | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_QP_INTERNAL_ERROR: |
There was a problem hiding this comment.
The local invalid error seems more close to FI_EINVAL
prov/efa/src/efa_errno.h
Outdated
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_QP_INTERNAL_ERROR: | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_AH: | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_LKEY: | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_UNRESP_REMOTE: |
There was a problem hiding this comment.
This unresponsive remote (no handshake made) should be EHOSTUNREACH
prov/efa/src/efa_errno.h
Outdated
| case FI_EFA_ERR_ESTABLISHED_RECV_UNRESP: | ||
| return FI_EOPBADSTATE; | ||
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_OP_TYPE: | ||
| return FI_EOPNOTSUPP; |
There was a problem hiding this comment.
This should be EINVAL as well
| switch (err) { | ||
| case EFA_IO_COMP_STATUS_OK: | ||
| return FI_SUCCESS; | ||
| case EFA_IO_COMP_STATUS_FLUSHED: |
There was a problem hiding this comment.
The flushed error should be EHOSTDOWN IMO because it usually means the remote destroyed the QP because of crash
prov/efa/src/efa_errno.h
Outdated
| case EFA_IO_COMP_STATUS_LOCAL_ERROR_BAD_LENGTH: | ||
| case EFA_IO_COMP_STATUS_REMOTE_ERROR_BAD_LENGTH: | ||
| return FI_EMSGSIZE; | ||
| case EFA_IO_COMP_STATUS_REMOTE_ERROR_BAD_ADDRESS: |
There was a problem hiding this comment.
According to the newest interpretation of this error code, this means invalid MR address, I think we should map it to FI_EINVAL
darrylabbate
force-pushed
the
efa/cq-err-entry
branch
2 times, most recently
from
b33414b to
25f61a4
Compare
|
It seems windows build failed because the errno is not available |
|
Need to re-run against new CI since old CI logs are gone |
|
bot:aws:new:retest |
|
bot:aws:retest |
1 similar comment
|
bot:aws:retest |
darrylabbate
force-pushed
the
efa/cq-err-entry
branch
from
25f61a4 to
c86b5af
Compare
|
Windows error from AWS CI: |
|
The issue here is that |
Signed-off-by: Darryl Abbate <[email protected]>
This adds a rudimentary function to map proprietary EFA status codes to common Libfabric status codes. This is useful when reporting errors to the application for operations that rely solely on ibverbs or RDMA Core, such as CQ polling. Signed-off-by: Darryl Abbate <[email protected]>
darrylabbate
force-pushed
the
efa/cq-err-entry
branch
from
c86b5af to
c1cfb26
Compare
This adds a rudimentary function to map proprietary EFA status codes to common Libfabric status codes. This is useful when reporting errors to the application for operations that rely solely on ibverbs or RDMA Core, such as CQ polling.