issue: 3925885 Adjusting XLIO default allocation logic

- Modified the default ring allocation logic from ‘ring per interface’ to ‘ring per thread’.
- Adjusted the MCE_SPEC_NGINX_DPU and MCE_SPEC_NGINX by setting
ring_allocation_logic_tx/rx to ‘ring per interface’.
- Fixed a minor side-effect in the ring_allocation_logic constructor.
- Disable progress engine by default to avoid performance degradation
with per-thread allocation logic.

Signed-off-by: Bashar Abdelgafer  <babdelgafer@nvidia.com>

src/core/dev/net_device_val.cpp

@@ -70,8 +70,9 @@
 #define nd_logfunc    __log_info_func
 #define nd_logfuncall __log_info_funcall
 
+// The value given in this constructor is later always overriden
 ring_alloc_logic_attr::ring_alloc_logic_attr()
-    : m_ring_alloc_logic(RING_LOGIC_PER_INTERFACE)
+    : m_ring_alloc_logic(RING_LOGIC_PER_THREAD)
     , m_use_locks(true)
     , m_user_id_key(0)
 {

src/core/dev/ring_allocation_logic.cpp

@@ -56,16 +56,12 @@ ring_allocation_logic::ring_allocation_logic()
 {
 }
 
-ring_allocation_logic::ring_allocation_logic(ring_logic_t allocation_logic,
-                                             int ring_migration_ratio, source_t source,
+ring_allocation_logic::ring_allocation_logic(int ring_migration_ratio, source_t source,
                                              resource_allocation_key &ring_profile)
     : m_ring_migration_ratio(ring_migration_ratio)
     , m_migration_try_count(ring_migration_ratio)
     , m_source(source)
 {
-    if (ring_profile.get_ring_alloc_logic() == RING_LOGIC_PER_INTERFACE) {
-        ring_profile.set_ring_alloc_logic(allocation_logic);
-    }
     m_res_key = resource_allocation_key(ring_profile);
     m_migration_candidate = 0;
     m_res_key.set_user_id_key(calc_res_key_by_logic());

src/core/dev/ring_allocation_logic.h

@@ -81,7 +81,7 @@ class source_t {
 class ring_allocation_logic {
 protected:
     ring_allocation_logic();
-    ring_allocation_logic(ring_logic_t ring_allocation_logic, int ring_migration_ratio,
+    ring_allocation_logic(int ring_migration_ratio,
                           source_t source, resource_allocation_key &ring_profile);
 
     void debug_print_type(const char *type);
@@ -121,8 +121,7 @@ class ring_allocation_logic_rx : public ring_allocation_logic {
         debug_print_type("Rx");
     }
     ring_allocation_logic_rx(source_t source, resource_allocation_key &ring_profile)
-        : ring_allocation_logic(safe_mce_sys().ring_allocation_logic_rx,
-                                safe_mce_sys().ring_migration_ratio_rx, source, ring_profile)
+        : ring_allocation_logic(safe_mce_sys().ring_migration_ratio_rx, source, ring_profile)
     {
         debug_print_type("Rx");
     }
@@ -136,8 +135,7 @@ class ring_allocation_logic_tx : public ring_allocation_logic {
         debug_print_type("Tx");
     }
     ring_allocation_logic_tx(source_t source, resource_allocation_key &ring_profile)
-        : ring_allocation_logic(safe_mce_sys().ring_allocation_logic_tx,
-                                safe_mce_sys().ring_migration_ratio_tx, source, ring_profile)
+        : ring_allocation_logic(safe_mce_sys().ring_migration_ratio_tx, source, ring_profile)
     {
         debug_print_type("Tx");
     }

src/core/util/sys_vars.cpp

@@ -1042,6 +1042,8 @@ void mce_sys_var::get_env_params()
 
 #ifdef DEFINED_NGINX
     case MCE_SPEC_NGINX:
+        ring_allocation_logic_tx = RING_LOGIC_PER_INTERFACE;
+        ring_allocation_logic_rx = RING_LOGIC_PER_INTERFACE;
         memory_limit = (app.workers_num > 16 ? 3072LU : 4096LU) * 1024 * 1024;
         memory_limit *= std::max(app.workers_num, 1);
         rx_bufs_batch = 8; // MCE_DEFAULT_RX_BUFS_BATCH (64), RX buffers batch size.
@@ -1081,6 +1083,8 @@ void mce_sys_var::get_env_params()
         break;
 
     case MCE_SPEC_NGINX_DPU:
+        ring_allocation_logic_tx = RING_LOGIC_PER_INTERFACE;
+        ring_allocation_logic_rx = RING_LOGIC_PER_INTERFACE;
         // The top part is different from NGINX SPEC
         memory_limit = (app.workers_num == 16 ? 512LU : 1024LU) * 1024 * 1024;
         memory_limit *= std::max(app.workers_num, 1);

src/core/util/sys_vars.h

@@ -727,8 +727,8 @@ extern mce_sys_var &safe_mce_sys();
 #define MCE_DEFAULT_HANDLE_SIGINTR           (true)
 #define MCE_DEFAULT_HANDLE_SIGFAULT          (false)
 #define MCE_DEFAULT_STATS_FD_NUM             0
-#define MCE_DEFAULT_RING_ALLOCATION_LOGIC_TX (RING_LOGIC_PER_INTERFACE)
-#define MCE_DEFAULT_RING_ALLOCATION_LOGIC_RX (RING_LOGIC_PER_INTERFACE)
+#define MCE_DEFAULT_RING_ALLOCATION_LOGIC_TX (RING_LOGIC_PER_THREAD)
+#define MCE_DEFAULT_RING_ALLOCATION_LOGIC_RX (RING_LOGIC_PER_THREAD)
 #define MCE_DEFAULT_RING_MIGRATION_RATIO_TX  (-1)
 #define MCE_DEFAULT_RING_MIGRATION_RATIO_RX  (-1)
 #define MCE_DEFAULT_RING_LIMIT_PER_INTERFACE (0)
@@ -800,7 +800,7 @@ extern mce_sys_var &safe_mce_sys();
 #define MCE_DEFAULT_CQ_AIM_INTERVAL_MSEC           (1000)
 #define MCE_DEFAULT_CQ_AIM_INTERRUPTS_RATE_PER_SEC (1000)
 #define MCE_DEFAULT_CQ_POLL_BATCH                  (16)
-#define MCE_DEFAULT_PROGRESS_ENGINE_INTERVAL_MSEC  (10)
+#define MCE_DEFAULT_PROGRESS_ENGINE_INTERVAL_MSEC  (0)
 #define MCE_DEFAULT_PROGRESS_ENGINE_WCE_MAX        (10000)
 #define MCE_DEFAULT_CQ_KEEP_QP_FULL                (true)
 #define MCE_DEFAULT_QP_COMPENSATION_LEVEL          (256)