In current implementation RDMA qpair is destroyed right after
disconnect. That is not graceful qpair shutdown process since
there can be requests submitted to HW and we may receive
completions for already destroyed/freed qpair.
To avoid this, only disconnect qpair in ctrlr_disconnect_qpair
transport callback, all other resources will be released in
ctrlr_delete_io_qpair cb.
This patch is useful when nvme poll groups are used since in
that case we use shared CQ, if the disconnected qpair has WRs
submitted to HW then qpair's destruction will be deferred to
poll group.
When nvme poll groups are not used, this patch doesn't change
anything, in that case destruction flow is still ungraceful.
However since CQ is destroyed immediately after qpair,
we shouldn't receive any requests which point to released
resources. A correct solution for non-poll group case
requires async diconnect API which may lead to significant
rework.
There is a bug when Soft Roce is used - we may receive
a completion with "normal" status when qpair is already
disconnected and all nvme requests are aborted. Added
a workaround for it.
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Change-Id: I0680d9ef9aaa8737d7a6d1454cd70a384bb8efac
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10327
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Shuhei Matsumoto <shuheimatsumoto@gmail.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
eb09178a59