doc: add Meson coding style to contributors guide
[dpdk.git] / lib / librte_timer / rte_timer.c
index ca88454..6d19ce4 100644 (file)
@@ -10,7 +10,6 @@
 #include <assert.h>
 #include <sys/queue.h>
 
-#include <rte_atomic.h>
 #include <rte_common.h>
 #include <rte_cycles.h>
 #include <rte_eal_memconfig.h>
@@ -26,7 +25,6 @@
 #include <rte_memzone.h>
 #include <rte_malloc.h>
 #include <rte_errno.h>
-#include <rte_function_versioning.h>
 
 #include "rte_timer.h"
 
@@ -146,11 +144,13 @@ rte_timer_subsystem_init(void)
        const size_t mem_size = data_arr_size + sizeof(*rte_timer_mz_refcnt);
        bool do_full_init = true;
 
-       if (rte_timer_subsystem_initialized)
-               return -EALREADY;
-
        rte_mcfg_timer_lock();
 
+       if (rte_timer_subsystem_initialized) {
+               rte_mcfg_timer_unlock();
+               return -EALREADY;
+       }
+
        mz = rte_memzone_lookup(mz_name);
        if (mz == NULL) {
                mz = rte_memzone_reserve_aligned(mz_name, mem_size,
@@ -184,27 +184,29 @@ rte_timer_subsystem_init(void)
        rte_timer_data_arr[default_data_id].internal_flags |= FL_ALLOCATED;
        (*rte_timer_mz_refcnt)++;
 
-       rte_mcfg_timer_unlock();
-
        rte_timer_subsystem_initialized = 1;
 
+       rte_mcfg_timer_unlock();
+
        return 0;
 }
 
 void
 rte_timer_subsystem_finalize(void)
 {
-       if (!rte_timer_subsystem_initialized)
-               return;
-
        rte_mcfg_timer_lock();
 
+       if (!rte_timer_subsystem_initialized) {
+               rte_mcfg_timer_unlock();
+               return;
+       }
+
        if (--(*rte_timer_mz_refcnt) == 0)
                rte_memzone_free(rte_timer_data_mz);
 
-       rte_mcfg_timer_unlock();
-
        rte_timer_subsystem_initialized = 0;
+
+       rte_mcfg_timer_unlock();
 }
 
 /* Initialize the timer handle tim for use */
@@ -215,7 +217,7 @@ rte_timer_init(struct rte_timer *tim)
 
        status.state = RTE_TIMER_STOP;
        status.owner = RTE_TIMER_NO_OWNER;
-       tim->status.u32 = status.u32;
+       __atomic_store_n(&tim->status.u32, status.u32, __ATOMIC_RELAXED);
 }
 
 /*
@@ -236,9 +238,9 @@ timer_set_config_state(struct rte_timer *tim,
 
        /* wait that the timer is in correct status before update,
         * and mark it as being configured */
-       while (success == 0) {
-               prev_status.u32 = tim->status.u32;
+       prev_status.u32 = __atomic_load_n(&tim->status.u32, __ATOMIC_RELAXED);
 
+       while (success == 0) {
                /* timer is running on another core
                 * or ready to run on local core, exit
                 */
@@ -255,9 +257,15 @@ timer_set_config_state(struct rte_timer *tim,
                 * mark it atomically as being configured */
                status.state = RTE_TIMER_CONFIG;
                status.owner = (int16_t)lcore_id;
-               success = rte_atomic32_cmpset(&tim->status.u32,
-                                             prev_status.u32,
-                                             status.u32);
+               /* CONFIG states are acting as locked states. If the
+                * timer is in CONFIG state, the state cannot be changed
+                * by other threads. So, we should use ACQUIRE here.
+                */
+               success = __atomic_compare_exchange_n(&tim->status.u32,
+                                             &prev_status.u32,
+                                             status.u32, 0,
+                                             __ATOMIC_ACQUIRE,
+                                             __ATOMIC_RELAXED);
        }
 
        ret_prev_status->u32 = prev_status.u32;
@@ -276,20 +284,27 @@ timer_set_running_state(struct rte_timer *tim)
 
        /* wait that the timer is in correct status before update,
         * and mark it as running */
-       while (success == 0) {
-               prev_status.u32 = tim->status.u32;
+       prev_status.u32 = __atomic_load_n(&tim->status.u32, __ATOMIC_RELAXED);
 
+       while (success == 0) {
                /* timer is not pending anymore */
                if (prev_status.state != RTE_TIMER_PENDING)
                        return -1;
 
-               /* here, we know that timer is stopped or pending,
-                * mark it atomically as being configured */
+               /* we know that the timer will be pending at this point
+                * mark it atomically as being running
+                */
                status.state = RTE_TIMER_RUNNING;
                status.owner = (int16_t)lcore_id;
-               success = rte_atomic32_cmpset(&tim->status.u32,
-                                             prev_status.u32,
-                                             status.u32);
+               /* RUNNING states are acting as locked states. If the
+                * timer is in RUNNING state, the state cannot be changed
+                * by other threads. So, we should use ACQUIRE here.
+                */
+               success = __atomic_compare_exchange_n(&tim->status.u32,
+                                             &prev_status.u32,
+                                             status.u32, 0,
+                                             __ATOMIC_ACQUIRE,
+                                             __ATOMIC_RELAXED);
        }
 
        return 0;
@@ -517,10 +532,12 @@ __rte_timer_reset(struct rte_timer *tim, uint64_t expire,
 
        /* update state: as we are in CONFIG state, only us can modify
         * the state so we don't need to use cmpset() here */
-       rte_wmb();
        status.state = RTE_TIMER_PENDING;
        status.owner = (int16_t)tim_lcore;
-       tim->status.u32 = status.u32;
+       /* The "RELEASE" ordering guarantees the memory operations above
+        * the status update are observed before the update by all threads
+        */
+       __atomic_store_n(&tim->status.u32, status.u32, __ATOMIC_RELEASE);
 
        if (tim_lcore != lcore_id || !local_is_locked)
                rte_spinlock_unlock(&priv_timer[tim_lcore].list_lock);
@@ -597,10 +614,12 @@ __rte_timer_stop(struct rte_timer *tim, int local_is_locked,
        }
 
        /* mark timer as stopped */
-       rte_wmb();
        status.state = RTE_TIMER_STOP;
        status.owner = RTE_TIMER_NO_OWNER;
-       tim->status.u32 = status.u32;
+       /* The "RELEASE" ordering guarantees the memory operations above
+        * the status update are observed before the update by all threads
+        */
+       __atomic_store_n(&tim->status.u32, status.u32, __ATOMIC_RELEASE);
 
        return 0;
 }
@@ -634,7 +653,8 @@ rte_timer_stop_sync(struct rte_timer *tim)
 int
 rte_timer_pending(struct rte_timer *tim)
 {
-       return tim->status.state == RTE_TIMER_PENDING;
+       return __atomic_load_n(&tim->status.state,
+                               __ATOMIC_RELAXED) == RTE_TIMER_PENDING;
 }
 
 /* must be called periodically, run all timer that expired */
@@ -736,8 +756,12 @@ __rte_timer_manage(struct rte_timer_data *timer_data)
                        /* remove from done list and mark timer as stopped */
                        status.state = RTE_TIMER_STOP;
                        status.owner = RTE_TIMER_NO_OWNER;
-                       rte_wmb();
-                       tim->status.u32 = status.u32;
+                       /* The "RELEASE" ordering guarantees the memory
+                        * operations above the status update are observed
+                        * before the update by all threads
+                        */
+                       __atomic_store_n(&tim->status.u32, status.u32,
+                               __ATOMIC_RELEASE);
                }
                else {
                        /* keep it in list and mark timer as pending */
@@ -745,8 +769,12 @@ __rte_timer_manage(struct rte_timer_data *timer_data)
                        status.state = RTE_TIMER_PENDING;
                        __TIMER_STAT_ADD(priv_timer, pending, 1);
                        status.owner = (int16_t)lcore_id;
-                       rte_wmb();
-                       tim->status.u32 = status.u32;
+                       /* The "RELEASE" ordering guarantees the memory
+                        * operations above the status update are observed
+                        * before the update by all threads
+                        */
+                       __atomic_store_n(&tim->status.u32, status.u32,
+                               __ATOMIC_RELEASE);
                        __rte_timer_reset(tim, tim->expire + tim->period,
                                tim->period, lcore_id, tim->f, tim->arg, 1,
                                timer_data);
@@ -916,8 +944,12 @@ rte_timer_alt_manage(uint32_t timer_data_id,
                        /* remove from done list and mark timer as stopped */
                        status.state = RTE_TIMER_STOP;
                        status.owner = RTE_TIMER_NO_OWNER;
-                       rte_wmb();
-                       tim->status.u32 = status.u32;
+                       /* The "RELEASE" ordering guarantees the memory
+                        * operations above the status update are observed
+                        * before the update by all threads
+                        */
+                       __atomic_store_n(&tim->status.u32, status.u32,
+                               __ATOMIC_RELEASE);
                } else {
                        /* keep it in list and mark timer as pending */
                        rte_spinlock_lock(
@@ -925,8 +957,12 @@ rte_timer_alt_manage(uint32_t timer_data_id,
                        status.state = RTE_TIMER_PENDING;
                        __TIMER_STAT_ADD(data->priv_timer, pending, 1);
                        status.owner = (int16_t)this_lcore;
-                       rte_wmb();
-                       tim->status.u32 = status.u32;
+                       /* The "RELEASE" ordering guarantees the memory
+                        * operations above the status update are observed
+                        * before the update by all threads
+                        */
+                       __atomic_store_n(&tim->status.u32, status.u32,
+                               __ATOMIC_RELEASE);
                        __rte_timer_reset(tim, tim->expire + tim->period,
                                tim->period, this_lcore, tim->f, tim->arg, 1,
                                data);
@@ -978,6 +1014,33 @@ rte_timer_stop_all(uint32_t timer_data_id, unsigned int *walk_lcores,
        return 0;
 }
 
+int64_t
+rte_timer_next_ticks(void)
+{
+       unsigned int lcore_id = rte_lcore_id();
+       struct rte_timer_data *timer_data;
+       struct priv_timer *priv_timer;
+       const struct rte_timer *tm;
+       uint64_t cur_time;
+       int64_t left = -ENOENT;
+
+       TIMER_DATA_VALID_GET_OR_ERR_RET(default_data_id, timer_data, -EINVAL);
+
+       priv_timer = timer_data->priv_timer;
+       cur_time = rte_get_timer_cycles();
+
+       rte_spinlock_lock(&priv_timer[lcore_id].list_lock);
+       tm = priv_timer[lcore_id].pending_head.sl_next[0];
+       if (tm) {
+               left = tm->expire - cur_time;
+               if (left < 0)
+                       left = 0;
+       }
+       rte_spinlock_unlock(&priv_timer[lcore_id].list_lock);
+
+       return left;
+}
+
 /* dump statistics about timers */
 static void
 __rte_timer_dump_stats(struct rte_timer_data *timer_data __rte_unused, FILE *f)