vppinfra: refactor test_and_set spinlocks to use clib_spinlock_t

Spinlock performance improved when implemented with compare_and_exchange
instead of test_and_set. All instances of test_and_set locks were refactored
to use clib_spinlock_t when possible. Some locks e.g. ssvm synchronize
between processes rather than threads, so they cannot directly use
clib_spinlock_t.

Type: refactor

Change-Id: Ia16b5d4cd49209b2b57b8df6c94615c28b11bb60
Signed-off-by: Jason Zhang <jason.zhang2@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Lijian Zhang <Lijian.Zhang@arm.com>

8 files changed, 54 insertions, 69 deletions

diff --git a/src/plugins/ioam/analyse/ioam_analyse.h b/src/plugins/ioam/analyse/ioam_analyse.h
index 3aec7a792d8..2a2cc15823f 100644
--- a/src/plugins/ioam/analyse/ioam_analyse.h
+++ b/src/plugins/ioam/analyse/ioam_analyse.h
@@ -22,6 +22,7 @@
 #include <ioam/lib-e2e/e2e_util.h>
 #include <ioam/lib-trace/trace_util.h>
 #include <ioam/lib-trace/trace_config.h>
+#include <vppinfra/lock.h>
 
 #define IOAM_FLOW_TEMPLATE_ID    260
 #define IOAM_TRACE_MAX_NODES      10
@@ -123,7 +124,7 @@ typedef struct ioam_analyser_data_t_
   struct ioam_analyser_data_t_ *chached_data_list;
 
   /** Lock to since we use this to export the data in other thread. */
-  volatile u32 *writer_lock;
+  clib_spinlock_t writer_lock;
 } ioam_analyser_data_t;
 
 always_inline f64
@@ -191,8 +192,7 @@ ip6_ioam_analyse_set_paths_down (ioam_analyser_data_t * data)
   ioam_path_map_t *path;
   u8 k, i;
 
-  while (clib_atomic_test_and_set (data->writer_lock))
-    ;
+  clib_spinlock_lock (&data->writer_lock);
 
   trace_data = &data->trace_data;
 
@@ -208,7 +208,7 @@ ip6_ioam_analyse_set_paths_down (ioam_analyser_data_t * data)
       for (k = 0; k < trace_record->num_nodes; k++)
 	path[k].state_up = 0;
     }
-  clib_atomic_release (data->writer_lock);
+  clib_spinlock_unlock (&data->writer_lock);
 }
 
 always_inline void
@@ -225,8 +225,7 @@ ip6_ioam_analyse_hbh_trace_loopback (ioam_analyser_data_t * data,
   u16 size_of_traceopt_per_node;
   u16 size_of_all_traceopts;
 
-  while (clib_atomic_test_and_set (data->writer_lock))
-    ;
+  clib_spinlock_lock (&data->writer_lock);
 
   trace_data = &data->trace_data;
 
@@ -277,7 +276,7 @@ ip6_ioam_analyse_hbh_trace_loopback (ioam_analyser_data_t * data,
 	}
     }
 end:
-  clib_atomic_release (data->writer_lock);
+  clib_spinlock_unlock (&data->writer_lock);
 }
 
 always_inline int
@@ -295,8 +294,7 @@ ip6_ioam_analyse_hbh_trace (ioam_analyser_data_t * data,
   ioam_path_map_t *path = NULL;
   ioam_analyse_trace_record *trace_record;
 
-  while (clib_atomic_test_and_set (data->writer_lock))
-    ;
+  clib_spinlock_lock (&data->writer_lock);
 
   trace_data = &data->trace_data;
 
@@ -409,7 +407,7 @@ found_match:
 	(u32) ((sum + delay) / (data->seqno_data.rx_packets + 1));
     }
 DONE:
-  clib_atomic_release (data->writer_lock);
+  clib_spinlock_unlock (&data->writer_lock);
   return 0;
 }
 
@@ -417,13 +415,12 @@ always_inline int
 ip6_ioam_analyse_hbh_e2e (ioam_analyser_data_t * data,
 			  ioam_e2e_packet_t * e2e, u16 len)
 {
-  while (clib_atomic_test_and_set (data->writer_lock))
-    ;
+  clib_spinlock_lock (&data->writer_lock);
 
   ioam_analyze_seqno (&data->seqno_data,
 		      (u64) clib_net_to_host_u32 (e2e->e2e_data));
 
-  clib_atomic_release (data->writer_lock);
+  clib_spinlock_unlock (&data->writer_lock);
 
   return 0;
 }
@@ -509,10 +506,7 @@ ioam_analyse_init_data (ioam_analyser_data_t * data)
    * get extended in future to maintain history of data */
   vec_validate_aligned (data->chached_data_list, 0, CLIB_CACHE_LINE_BYTES);
 
-  data->writer_lock = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES,
-					      CLIB_CACHE_LINE_BYTES);
-
-  clib_atomic_release (data->writer_lock);
+  clib_spinlock_init (&data->writer_lock);
 
   trace_data = &(data->trace_data);
   for (j = 0; j < IOAM_MAX_PATHS_PER_FLOW; j++)
diff --git a/src/plugins/ioam/analyse/ioam_summary_export.c b/src/plugins/ioam/analyse/ioam_summary_export.c
index 4851491012b..12907d8c54b 100644
--- a/src/plugins/ioam/analyse/ioam_summary_export.c
+++ b/src/plugins/ioam/analyse/ioam_summary_export.c
@@ -150,8 +150,7 @@ ioam_analyse_add_ipfix_record (flow_report_t * fr,
 			       ip6_address_t * src, ip6_address_t * dst,
 			       u16 src_port, u16 dst_port)
 {
-  while (clib_atomic_test_and_set (record->writer_lock))
-    ;
+  clib_spinlock_lock (&record->writer_lock);
 
   int field_index = 0;
   u16 tmp;
@@ -259,7 +258,7 @@ ioam_analyse_add_ipfix_record (flow_report_t * fr,
   *(record->chached_data_list) = *record;
   record->chached_data_list->chached_data_list = NULL;
 
-  clib_atomic_release (record->writer_lock);
+  clib_spinlock_unlock (&record->writer_lock);
   return offset;
 }
 
diff --git a/src/plugins/ioam/analyse/ip6/node.c b/src/plugins/ioam/analyse/ip6/node.c
index 7a8d71de93b..2568e305b39 100644
--- a/src/plugins/ioam/analyse/ip6/node.c
+++ b/src/plugins/ioam/analyse/ip6/node.c
@@ -256,17 +256,15 @@ ip6_ioam_analyse_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  data0 = ioam_analyse_get_data_from_flow_id (flow_id0);
 		  data1 = ioam_analyse_get_data_from_flow_id (flow_id1);
 
-		  while (clib_atomic_test_and_set (data0->writer_lock))
-		    ;
+		  clib_spinlock_lock (&data0->writer_lock);
 		  data0->pkt_counter++;
 		  data0->bytes_counter += p_len0;
-		  clib_atomic_release (data0->writer_lock);
+		  clib_spinlock_unlock (&data0->writer_lock);
 
-		  while (clib_atomic_test_and_set (data1->writer_lock))
-		    ;
+		  clib_spinlock_lock (&data1->writer_lock);
 		  data1->pkt_counter++;
 		  data1->bytes_counter += p_len1;
-		  clib_atomic_release (data1->writer_lock);
+		  clib_spinlock_unlock (&data1->writer_lock);
 		}
 	      else if (error0 == 0)
 		{
@@ -274,11 +272,10 @@ ip6_ioam_analyse_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  pkts_failed++;
 
 		  data0 = ioam_analyse_get_data_from_flow_id (flow_id0);
-		  while (clib_atomic_test_and_set (data0->writer_lock))
-		    ;
+		  clib_spinlock_lock (&data0->writer_lock);
 		  data0->pkt_counter++;
 		  data0->bytes_counter += p_len0;
-		  clib_atomic_release (data0->writer_lock);
+		  clib_spinlock_unlock (&data0->writer_lock);
 		}
 	      else if (error1 == 0)
 		{
@@ -286,11 +283,10 @@ ip6_ioam_analyse_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  pkts_failed++;
 
 		  data1 = ioam_analyse_get_data_from_flow_id (flow_id1);
-		  while (clib_atomic_test_and_set (data1->writer_lock))
-		    ;
+		  clib_spinlock_lock (&data1->writer_lock);
 		  data1->pkt_counter++;
 		  data1->bytes_counter += p_len1;
-		  clib_atomic_release (data1->writer_lock);
+		  clib_spinlock_unlock (&data1->writer_lock);
 		}
 	      else
 		pkts_failed += 2;
@@ -327,12 +323,11 @@ ip6_ioam_analyse_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		{
 		  pkts_analysed++;
 		  data0 = ioam_analyse_get_data_from_flow_id (flow_id0);
-		  while (clib_atomic_test_and_set (data0->writer_lock))
-		    ;
+		  clib_spinlock_lock (&data0->writer_lock);
 		  data0->pkt_counter++;
 		  data0->bytes_counter +=
 		    clib_net_to_host_u16 (ip60->payload_length);
-		  clib_atomic_release (data0->writer_lock);
+		  clib_spinlock_unlock (&data0->writer_lock);
 		}
 	      else
 		pkts_failed++;
@@ -393,13 +388,12 @@ ip6_ioam_analyse_hbh_pot (u32 flow_id, ip6_hop_by_hop_option_t * opt0,
   pot_profile = pot_profile_get_active ();
   ret = pot_validate (pot_profile, cumulative, random);
 
-  while (clib_atomic_test_and_set (data->writer_lock))
-    ;
+  clib_spinlock_lock (&data->writer_lock);
 
   (0 == ret) ? (data->pot_data.sfc_validated_count++) :
     (data->pot_data.sfc_invalidated_count++);
 
-  clib_atomic_release (data->writer_lock);
+  clib_spinlock_unlock (&data->writer_lock);
   return 0;
 }
 
diff --git a/src/plugins/ioam/export-common/ioam_export.h b/src/plugins/ioam/export-common/ioam_export.h
index 672576ebded..b1bca9bf91f 100644
--- a/src/plugins/ioam/export-common/ioam_export.h
+++ b/src/plugins/ioam/export-common/ioam_export.h
@@ -28,6 +28,7 @@
 #include <vppinfra/hash.h>
 #include <vppinfra/error.h>
 #include <vppinfra/elog.h>
+#include <vppinfra/lock.h>
 
 #include <vlib/threads.h>
 
@@ -62,7 +63,7 @@ typedef struct
   /* Vector of per thread ioam_export_buffer_t to buffer pool index */
   u32 *buffer_per_thread;
   /* Lock per thread to swap buffers between worker and timer process */
-  volatile u32 **lockp;
+  clib_spinlock_t *lockp;
 
   /* time scale transform */
   u32 unix_time_0;
@@ -194,9 +195,7 @@ ioam_export_thread_buffer_init (ioam_export_main_t * em, vlib_main_t * vm)
 	  ioam_export_thread_buffer_free (em);
 	  return (-2);
 	}
-      em->lockp[i] = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES,
-					     CLIB_CACHE_LINE_BYTES);
-      clib_memset ((void *) em->lockp[i], 0, CLIB_CACHE_LINE_BYTES);
+      clib_spinlock_init (&em->lockp[i]);
     }
   return (1);
 }
@@ -404,7 +403,7 @@ ioam_export_process_common (ioam_export_main_t * em, vlib_main_t * vm,
       for (i = 0; i < vec_len (em->buffer_per_thread); i++)
 	{
 	  /* If the worker thread is processing export records ignore further checks */
-	  if (*em->lockp[i] == 1)
+	  if (CLIB_SPINLOCK_IS_LOCKED (&em->lockp[i]))
 	    continue;
 	  eb = pool_elt_at_index (em->buffer_pool, em->buffer_per_thread[i]);
 	  if (eb->records_in_this_buffer > 0
@@ -436,11 +435,10 @@ ioam_export_process_common (ioam_export_main_t * em, vlib_main_t * vm,
 	   */
 	  for (i = 0; i < vec_len (thread_index); i++)
 	    {
-	      while (clib_atomic_test_and_set (em->lockp[thread_index[i]]))
-		;
+	      clib_spinlock_lock (&em->lockp[thread_index[i]]);
 	      em->buffer_per_thread[thread_index[i]] =
 		vec_pop (vec_buffer_indices);
-	      clib_atomic_release (em->lockp[thread_index[i]]);
+	      clib_spinlock_unlock (&em->lockp[thread_index[i]]);
 	    }
 
 	  /* Send the buffers */
@@ -479,7 +477,7 @@ do {                                                                           \
   from = vlib_frame_vector_args (F);                                           \
   n_left_from = (F)->n_vectors;                                                \
   next_index = (N)->cached_next_index;                                         \
-  while (clib_atomic_test_and_set ((EM)->lockp[(VM)->thread_index]));	       \
+  clib_spinlock_lock (&(EM)->lockp[(VM)->thread_index]);	               \
   my_buf = ioam_export_get_my_buffer (EM, (VM)->thread_index);                 \
   my_buf->touched_at = vlib_time_now (VM);                                     \
   while (n_left_from > 0)                                                      \
@@ -622,7 +620,7 @@ do {                                                                           \
     }                                                                          \
   vlib_node_increment_counter (VM, export_node.index,                          \
 			       EXPORT_ERROR_RECORDED, pkts_recorded);          \
-  *(EM)->lockp[(VM)->thread_index] = 0;                                        \
+  clib_spinlock_unlock (&(EM)->lockp[(VM)->thread_index]);                     \
 } while(0)
 
 #endif /* __included_ioam_export_h__ */
diff --git a/src/plugins/lb/lb.c b/src/plugins/lb/lb.c
index 75ca40f5f00..3ac2b284ea6 100644
--- a/src/plugins/lb/lb.c
+++ b/src/plugins/lb/lb.c
@@ -18,6 +18,7 @@
 #include <vpp/app/version.h>
 #include <vnet/api_errno.h>
 #include <vnet/udp/udp.h>
+#include <vppinfra/lock.h>
 
 //GC runs at most once every so many seconds
 #define LB_GARBAGE_RUN 60
@@ -27,8 +28,8 @@
 
 lb_main_t lb_main;
 
-#define lb_get_writer_lock() do {} while(clib_atomic_test_and_set (lb_main.writer_lock))
-#define lb_put_writer_lock() clib_atomic_release (lb_main.writer_lock)
+#define lb_get_writer_lock() clib_spinlock_lock (&lb_main.writer_lock)
+#define lb_put_writer_lock() clib_spinlock_unlock (&lb_main.writer_lock)
 
 static void lb_as_stack (lb_as_t *as);
 
@@ -289,7 +290,7 @@ static void lb_vip_garbage_collection(lb_vip_t *vip)
   lb_snat6_key_t m_key6;
   clib_bihash_kv_24_8_t kv6, value6;
   lb_snat_mapping_t *m = 0;
-  ASSERT (lbm->writer_lock[0]);
+  CLIB_SPINLOCK_ASSERT_LOCKED (&lbm->writer_lock);
 
   u32 now = (u32) vlib_time_now(vlib_get_main());
   if (!clib_u32_loop_gt(now, vip->last_garbage_collection + LB_GARBAGE_RUN))
@@ -384,7 +385,7 @@ static void lb_vip_update_new_flow_table(lb_vip_t *vip)
   lb_as_t *as;
   lb_pseudorand_t *pr, *sort_arr = 0;
 
-  ASSERT (lbm->writer_lock[0]); //We must have the lock
+  CLIB_SPINLOCK_ASSERT_LOCKED (&lbm->writer_lock); // We must have the lock
 
   //Check if some AS is configured or not
   i = 0;
@@ -496,7 +497,8 @@ int lb_vip_port_find_index(ip46_address_t *prefix, u8 plen,
 {
   lb_main_t *lbm = &lb_main;
   lb_vip_t *vip;
-  ASSERT (lbm->writer_lock[0]); //This must be called with the lock owned
+  /* This must be called with the lock owned */
+  CLIB_SPINLOCK_ASSERT_LOCKED (&lbm->writer_lock);
   ip46_prefix_normalize(prefix, plen);
   pool_foreach(vip, lbm->vips, {
       if ((vip->flags & LB_AS_FLAGS_USED) &&
@@ -560,7 +562,8 @@ int lb_vip_find_index(ip46_address_t *prefix, u8 plen, u8 protocol,
 static int lb_as_find_index_vip(lb_vip_t *vip, ip46_address_t *address, u32 *as_index)
 {
   lb_main_t *lbm = &lb_main;
-  ASSERT (lbm->writer_lock[0]); //This must be called with the lock owned
+  /* This must be called with the lock owned */
+  CLIB_SPINLOCK_ASSERT_LOCKED (&lbm->writer_lock);
   lb_as_t *as;
   u32 *asi;
   pool_foreach(asi, vip->as_indexes, {
@@ -1384,8 +1387,7 @@ lb_init (vlib_main_t * vm)
 
   lbm->per_cpu = 0;
   vec_validate(lbm->per_cpu, tm->n_vlib_mains - 1);
-  lbm->writer_lock = clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES,  CLIB_CACHE_LINE_BYTES);
-  lbm->writer_lock[0] = 0;
+  clib_spinlock_init (&lbm->writer_lock);
   lbm->per_cpu_sticky_buckets = LB_DEFAULT_PER_CPU_STICKY_BUCKETS;
   lbm->flow_timeout = LB_DEFAULT_FLOW_TIMEOUT;
   lbm->ip4_src_address.as_u32 = 0xffffffff;
diff --git a/src/plugins/lb/lb.h b/src/plugins/lb/lb.h
index d09ac632be6..9899eab1fc8 100644
--- a/src/plugins/lb/lb.h
+++ b/src/plugins/lb/lb.h
@@ -41,6 +41,7 @@
 #include <vppinfra/bihash_8_8.h>
 #include <vppinfra/bihash_24_8.h>
 #include <lb/lbhash.h>
+#include <vppinfra/lock.h>
 
 #define LB_DEFAULT_PER_CPU_STICKY_BUCKETS 1 << 10
 #define LB_DEFAULT_FLOW_TIMEOUT 40
@@ -561,7 +562,7 @@ typedef struct {
    */
   u16 msg_id_base;
 
-  volatile u32 *writer_lock;
+  clib_spinlock_t writer_lock;
 
   /* convenience */
   vlib_main_t *vlib_main;
diff --git a/src/plugins/map/map.c b/src/plugins/map/map.c
index 2f036da7ebc..6ec9e72b413 100644
--- a/src/plugins/map/map.c
+++ b/src/plugins/map/map.c
@@ -2265,9 +2265,7 @@ map_init (vlib_main_t * vm)
   /* IP4 virtual reassembly */
   mm->ip4_reass_hash_table = 0;
   mm->ip4_reass_pool = 0;
-  mm->ip4_reass_lock =
-    clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, CLIB_CACHE_LINE_BYTES);
-  *mm->ip4_reass_lock = 0;
+  clib_spinlock_init (&mm->ip4_reass_lock);
   mm->ip4_reass_conf_ht_ratio = MAP_IP4_REASS_HT_RATIO_DEFAULT;
   mm->ip4_reass_conf_lifetime_ms = MAP_IP4_REASS_LIFETIME_DEFAULT;
   mm->ip4_reass_conf_pool_size = MAP_IP4_REASS_POOL_SIZE_DEFAULT;
@@ -2281,9 +2279,7 @@ map_init (vlib_main_t * vm)
   /* IP6 virtual reassembly */
   mm->ip6_reass_hash_table = 0;
   mm->ip6_reass_pool = 0;
-  mm->ip6_reass_lock =
-    clib_mem_alloc_aligned (CLIB_CACHE_LINE_BYTES, CLIB_CACHE_LINE_BYTES);
-  *mm->ip6_reass_lock = 0;
+  clib_spinlock_init (&mm->ip6_reass_lock);
   mm->ip6_reass_conf_ht_ratio = MAP_IP6_REASS_HT_RATIO_DEFAULT;
   mm->ip6_reass_conf_lifetime_ms = MAP_IP6_REASS_LIFETIME_DEFAULT;
   mm->ip6_reass_conf_pool_size = MAP_IP6_REASS_POOL_SIZE_DEFAULT;
diff --git a/src/plugins/map/map.h b/src/plugins/map/map.h
index 6dc5232bf43..a65a16a06d5 100644
--- a/src/plugins/map/map.h
+++ b/src/plugins/map/map.h
@@ -306,7 +306,7 @@ typedef struct {
   u16 ip4_reass_allocated;
   u16 *ip4_reass_hash_table;
   u16 ip4_reass_fifo_last;
-  volatile u32 *ip4_reass_lock;
+  clib_spinlock_t ip4_reass_lock;
 
   /* Counters */
   u32 ip4_reass_buffered_counter;
@@ -329,7 +329,7 @@ typedef struct {
   u16 ip6_reass_allocated;
   u16 *ip6_reass_hash_table;
   u16 ip6_reass_fifo_last;
-  volatile u32 *ip6_reass_lock;
+  clib_spinlock_t ip6_reass_lock;
 
   /* Counters */
   u32 ip6_reass_buffered_counter;
@@ -502,8 +502,8 @@ map_ip4_reass_get(u32 src, u32 dst, u16 fragment_id,
 void
 map_ip4_reass_free(map_ip4_reass_t *r, u32 **pi_to_drop);
 
-#define map_ip4_reass_lock() while (clib_atomic_test_and_set (map_main.ip4_reass_lock)) { CLIB_PAUSE (); }
-#define map_ip4_reass_unlock() clib_atomic_release (map_main.ip4_reass_lock)
+#define map_ip4_reass_lock() clib_spinlock_lock (&map_main.ip4_reass_lock)
+#define map_ip4_reass_unlock() clib_spinlock_unlock (&map_main.ip4_reass_lock)
 
 static_always_inline void
 map_ip4_reass_get_fragments(map_ip4_reass_t *r, u32 **pi)
@@ -527,8 +527,8 @@ map_ip6_reass_get(ip6_address_t *src, ip6_address_t *dst, u32 fragment_id,
 void
 map_ip6_reass_free(map_ip6_reass_t *r, u32 **pi_to_drop);
 
-#define map_ip6_reass_lock() while (clib_atomic_test_and_set (map_main.ip6_reass_lock)) { CLIB_PAUSE (); }
-#define map_ip6_reass_unlock() clib_atomic_release (map_main.ip6_reass_lock)
+#define map_ip6_reass_lock() clib_spinlock_lock (&map_main.ip6_reass_lock)
+#define map_ip6_reass_unlock() clib_spinlock_unlock (&map_main.ip6_reass_lock)
 
 int
 map_ip6_reass_add_fragment(map_ip6_reass_t *r, u32 pi,
@@ -589,6 +589,7 @@ map_domain_counter_lock (map_main_t *mm)
     while (clib_atomic_test_and_set (mm->counter_lock))
       /* zzzz */ ;
 }
+
 static inline void
 map_domain_counter_unlock (map_main_t *mm)
 {