Re: [PATCH v4 0/5] ceph: reduce mdsc->mutex contention in the cephfs kclient

Hi Xiubo,

Patches 1-4 look good — the xarray replacement and the wait_list_lock
are clean, and the mutex narrowing is the right direction. One blocker
in patch 5:

replay_unsafe_requests() calls list_del_init(&req->r_unsafe_item) before
parking the request on the local walk list via r_aux_item. That permanently
removes the request from session->s_unsafe. After replay, the request is
no longer tracked as unsafe — subsequent reconnects won't replay it, and
cleanup_session_requests() can't drop it with -EIO on session teardown.
The request can get stuck in the xarray with no error surfaced to userspace.

The fix is to not unlink from s_unsafe at all — just add to the local list
via r_aux_item and leave r_unsafe_item in place. That's the whole point of
having r_aux_item.

Also r_aux_item is added to the struct in patch 4 but never initialized
with INIT_LIST_HEAD() in ceph_mdsc_create_request().

Could you respin as v5?

--
Alex Markuze