osd/scheduler/mClockScheduler: Use same profile and client ids for all clients

Currently, without the distributed mClock (dmClock) feature, all external
clients classified under op_scheduler_class::client must share the
allocated reservation and limit as set within default_external_client_info.
This is realized only if all the external clients are assigned the same
id. This was not true because the client_id field within the
client_profile_id_t was set to item.get_owner() which uniquely identified
each client. As a result, an increase in the client count resulted in a
proportional increase in the consumption of the allocated limit which is
not the desired outcome in the current implementation.

The fix is to use the same client_id and profile_id within the
client_profile_id_t structure for all clients. These fields are set to '0'.
This ensures that all such clients get the same QoS controls which means
reservation and limit is shared between them. In other words, regardless
of the number of clients active, the reservation and limit consumption
remains constant as per the mClock profile setting and is shared amongst
them.

NOTE: When dmClock feature (client vs client QoS) is implemented, the
above fields would change with client_id set to the global id (client.####)
and the profile_id set to the id generated by the client's QoS profile.

Fixes: https://tracker.ceph.com/issues/62293
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>

diff --git a/src/osd/scheduler/mClockScheduler.h b/src/osd/scheduler/mClockScheduler.h
index 3d958bed9dcd4f3cf012114143be99a6afb80d1d..f708b1d7a1e38cdb487d0c723b97835000db866b 100644 (file)
--- a/src/osd/scheduler/mClockScheduler.h
+++ b/src/osd/scheduler/mClockScheduler.h
@@ -38,12 +38,30 @@ constexpr double default_max = std::numeric_limits<double>::is_iec559 ?
   std::numeric_limits<double>::infinity() :
   std::numeric_limits<double>::max();
 
-using client_id_t = uint64_t;
-using profile_id_t = uint64_t;
-
+/**
+ * client_profile_id_t
+ *
+ * client_id - global id (client.####) for client QoS
+ * profile_id - id generated by client's QoS profile
+ *
+ * Currently (Reef and below), both members are set to
+ * 0 which ensures that all external clients share the
+ * mClock profile allocated reservation and limit
+ * bandwidth.
+ *
+ * Note: Post Reef, both members will be set to non-zero
+ * values when the distributed feature of the mClock
+ * algorithm is utilized.
+ */
 struct client_profile_id_t {
-  client_id_t client_id;
-  profile_id_t profile_id;
+  uint64_t client_id = 0;
+  uint64_t profile_id = 0;
+
+  client_profile_id_t(uint64_t _client_id, uint64_t _profile_id) :
+    client_id(_client_id),
+    profile_id(_profile_id) {}
+
+  client_profile_id_t() = default;
 
   auto operator<=>(const client_profile_id_t&) const = default;
   friend std::ostream& operator<<(std::ostream& out,
@@ -177,10 +195,7 @@ class mClockScheduler : public OpScheduler, md_config_obs_t {
   static scheduler_id_t get_scheduler_id(const OpSchedulerItem &item) {
     return scheduler_id_t{
       item.get_scheduler_class(),
-       client_profile_id_t{
-       item.get_owner(),
-         0
-         }
+      client_profile_id_t()
     };
   }