osc/rdma: fix data integrity issue in rdma get_accumulate #10634
Conversation
devreal
left a comment
devreal
left a comment
There was a problem hiding this comment.
Thanks for the fix 👍 One detail though
| /* result buffer is not necessarily contiguous. use the opal datatype engine to | ||
| * copy the data over in this case */ | ||
| struct iovec iov = {.iov_base = result, .iov_len = request->len}; | ||
| struct iovec iov = {.iov_base = result_start, .iov_len = request->len}; |
There was a problem hiding this comment.
If result != NULL then result_start is NULL (per https://github.com/open-mpi/ompi/pull/10634/files#diff-e4bd30eae59fa8aabf9ee87dcb523aa60a292e7c09d1e84e014c1ae7e13e1e3cR281). result_start should always point to the start of the result buffer.
There was a problem hiding this comment.
Thanks for the review, updated the PR with the changes, there was also a memory leak here, https://github.com/open-mpi/ompi/pull/10634/files#diff-e4bd30eae59fa8aabf9ee87dcb523aa60a292e7c09d1e84e014c1ae7e13e1e3cR311 which is fixed.
|
When this PR is complete, will it need to be cherry-picked back to v4.1.x and v4.0.x, too? |
Use start of the result buffer to unpack the data. Signed-off-by: Aboorva Devarajan <abodevar@in.ibm.com>
AboorvaDevarajan
force-pushed
the
fix_osc_rdma
branch
from
874a8cb to
02d2f5c
Compare
Jeff, this test passes in |
|
@AboorvaDevarajan Can you please backport this fix to the release branches? |
3 participants
Signed-off-by: Aboorva Devarajan abodevar@in.ibm.com