timer: remove unnecessary list insertion
[dpdk.git] / lib / librte_timer / rte_timer.c
old mode 100755 (executable)
new mode 100644 (file)
index f0f0c2f..3f21f2d
@@ -1,13 +1,13 @@
 /*-
  *   BSD LICENSE
- * 
+ *
  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
  *   All rights reserved.
- * 
+ *
  *   Redistribution and use in source and binary forms, with or without
  *   modification, are permitted provided that the following conditions
  *   are met:
- * 
+ *
  *     * Redistributions of source code must retain the above copyright
  *       notice, this list of conditions and the following disclaimer.
  *     * Redistributions in binary form must reproduce the above copyright
@@ -17,7 +17,7 @@
  *     * Neither the name of Intel Corporation nor the names of its
  *       contributors may be used to endorse or promote products derived
  *       from this software without specific prior written permission.
- * 
+ *
  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
@@ -35,6 +35,8 @@
 #include <stdio.h>
 #include <stdint.h>
 #include <inttypes.h>
+#include <assert.h>
+#include <sys/queue.h>
 
 #include <rte_atomic.h>
 #include <rte_common.h>
@@ -43,7 +45,6 @@
 #include <rte_memory.h>
 #include <rte_memzone.h>
 #include <rte_launch.h>
-#include <rte_tailq.h>
 #include <rte_eal.h>
 #include <rte_per_lcore.h>
 #include <rte_lcore.h>
@@ -79,9 +80,10 @@ static struct priv_timer priv_timer[RTE_MAX_LCORE];
 
 /* when debug is enabled, store some statistics */
 #ifdef RTE_LIBRTE_TIMER_DEBUG
-#define __TIMER_STAT_ADD(name, n) do {                         \
-               unsigned __lcore_id = rte_lcore_id();           \
-               priv_timer[__lcore_id].stats.name += (n);       \
+#define __TIMER_STAT_ADD(name, n) do {                                 \
+               unsigned __lcore_id = rte_lcore_id();                   \
+               if (__lcore_id < RTE_MAX_LCORE)                         \
+                       priv_timer[__lcore_id].stats.name += (n);       \
        } while(0)
 #else
 #define __TIMER_STAT_ADD(name, n) do {} while(0)
@@ -135,7 +137,7 @@ timer_set_config_state(struct rte_timer *tim,
 
                /* timer is running on another core, exit */
                if (prev_status.state == RTE_TIMER_RUNNING &&
-                   (unsigned)prev_status.owner != lcore_id)
+                   prev_status.owner != (uint16_t)lcore_id)
                        return -1;
 
                /* timer is being configured on another core */
@@ -366,9 +368,16 @@ __rte_timer_reset(struct rte_timer *tim, uint64_t expire,
 
        /* round robin for tim_lcore */
        if (tim_lcore == (unsigned)LCORE_ID_ANY) {
-               tim_lcore = rte_get_next_lcore(priv_timer[lcore_id].prev_lcore,
-                                              0, 1);
-               priv_timer[lcore_id].prev_lcore = tim_lcore;
+               if (lcore_id < RTE_MAX_LCORE) {
+                       /* EAL thread with valid lcore_id */
+                       tim_lcore = rte_get_next_lcore(
+                               priv_timer[lcore_id].prev_lcore,
+                               0, 1);
+                       priv_timer[lcore_id].prev_lcore = tim_lcore;
+               } else
+                       /* non-EAL thread do not run rte_timer_manage(),
+                        * so schedule the timer on the first enabled lcore. */
+                       tim_lcore = rte_get_next_lcore(LCORE_ID_ANY, 0, 1);
        }
 
        /* wait that the timer is in correct status before update,
@@ -378,7 +387,8 @@ __rte_timer_reset(struct rte_timer *tim, uint64_t expire,
                return -1;
 
        __TIMER_STAT_ADD(reset, 1);
-       if (prev_status.state == RTE_TIMER_RUNNING) {
+       if (prev_status.state == RTE_TIMER_RUNNING &&
+           lcore_id < RTE_MAX_LCORE) {
                priv_timer[lcore_id].updated = 1;
        }
 
@@ -424,10 +434,8 @@ rte_timer_reset(struct rte_timer *tim, uint64_t ticks,
        else
                period = 0;
 
-       __rte_timer_reset(tim,  cur_time + ticks, period, tim_lcore,
+       return __rte_timer_reset(tim,  cur_time + ticks, period, tim_lcore,
                          fct, arg, 0);
-
-       return 0;
 }
 
 /* loop until rte_timer_reset() succeed */
@@ -437,7 +445,8 @@ rte_timer_reset_sync(struct rte_timer *tim, uint64_t ticks,
                     rte_timer_cb_t fct, void *arg)
 {
        while (rte_timer_reset(tim, ticks, type, tim_lcore,
-                              fct, arg) != 0);
+                              fct, arg) != 0)
+               rte_pause();
 }
 
 /* Stop the timer associated with the timer handle tim */
@@ -455,7 +464,8 @@ rte_timer_stop(struct rte_timer *tim)
                return -1;
 
        __TIMER_STAT_ADD(stop, 1);
-       if (prev_status.state == RTE_TIMER_RUNNING) {
+       if (prev_status.state == RTE_TIMER_RUNNING &&
+           lcore_id < RTE_MAX_LCORE) {
                priv_timer[lcore_id].updated = 1;
        }
 
@@ -494,11 +504,15 @@ void rte_timer_manage(void)
 {
        union rte_timer_status status;
        struct rte_timer *tim, *next_tim;
+       struct rte_timer *run_first_tim, **pprev;
        unsigned lcore_id = rte_lcore_id();
        struct rte_timer *prev[MAX_SKIPLIST_DEPTH + 1];
        uint64_t cur_time;
        int i, ret;
 
+       /* timer manager only runs on EAL thread with valid lcore_id */
+       assert(lcore_id < RTE_MAX_LCORE);
+
        __TIMER_STAT_ADD(manage, 1);
        /* optimize for the case where per-cpu list is empty */
        if (priv_timer[lcore_id].pending_head.sl_next[0] == NULL)
@@ -506,9 +520,9 @@ void rte_timer_manage(void)
        cur_time = rte_get_timer_cycles();
 
 #ifdef RTE_ARCH_X86_64
-       /* on 64-bit the value cached in the pending_head.expired will be updated
-        * atomically, so we can consult that for a quick check here outside the
-        * lock */
+       /* on 64-bit the value cached in the pending_head.expired will be
+        * updated atomically, so we can consult that for a quick check here
+        * outside the lock */
        if (likely(priv_timer[lcore_id].pending_head.expire > cur_time))
                return;
 #endif
@@ -518,8 +532,10 @@ void rte_timer_manage(void)
 
        /* if nothing to do just unlock and return */
        if (priv_timer[lcore_id].pending_head.sl_next[0] == NULL ||
-                       priv_timer[lcore_id].pending_head.sl_next[0]->expire > cur_time)
-               goto done;
+           priv_timer[lcore_id].pending_head.sl_next[0]->expire > cur_time) {
+               rte_spinlock_unlock(&priv_timer[lcore_id].list_lock);
+               return;
+       }
 
        /* save start of list of expired timers */
        tim = priv_timer[lcore_id].pending_head.sl_next[0];
@@ -527,30 +543,48 @@ void rte_timer_manage(void)
        /* break the existing list at current time point */
        timer_get_prev_entries(cur_time, lcore_id, prev);
        for (i = priv_timer[lcore_id].curr_skiplist_depth -1; i >= 0; i--) {
-               priv_timer[lcore_id].pending_head.sl_next[i] = prev[i]->sl_next[i];
+               if (prev[i] == &priv_timer[lcore_id].pending_head)
+                       continue;
+               priv_timer[lcore_id].pending_head.sl_next[i] =
+                   prev[i]->sl_next[i];
                if (prev[i]->sl_next[i] == NULL)
                        priv_timer[lcore_id].curr_skiplist_depth--;
                prev[i] ->sl_next[i] = NULL;
        }
 
-       /* now scan expired list and call callbacks */
+       /* transition run-list from PENDING to RUNNING */
+       run_first_tim = tim;
+       pprev = &run_first_tim;
+
        for ( ; tim != NULL; tim = next_tim) {
                next_tim = tim->sl_next[0];
 
                ret = timer_set_running_state(tim);
+               if (likely(ret == 0)) {
+                       pprev = &tim->sl_next[0];
+               } else {
+                       /* another core is trying to re-config this one,
+                        * remove it from local expired list
+                        */
+                       *pprev = next_tim;
+               }
+       }
 
-               /* this timer was not pending, continue */
-               if (ret < 0)
-                       continue;
+       /* update the next to expire timer value */
+       priv_timer[lcore_id].pending_head.expire =
+           (priv_timer[lcore_id].pending_head.sl_next[0] == NULL) ? 0 :
+               priv_timer[lcore_id].pending_head.sl_next[0]->expire;
 
-               rte_spinlock_unlock(&priv_timer[lcore_id].list_lock);
+       rte_spinlock_unlock(&priv_timer[lcore_id].list_lock);
 
+       /* now scan expired list and call callbacks */
+       for (tim = run_first_tim; tim != NULL; tim = next_tim) {
+               next_tim = tim->sl_next[0];
                priv_timer[lcore_id].updated = 0;
 
                /* execute callback function with list unlocked */
                tim->f(tim, tim->arg);
 
-               rte_spinlock_lock(&priv_timer[lcore_id].list_lock);
                __TIMER_STAT_ADD(pending, -1);
                /* the timer was stopped or reloaded by the callback
                 * function, we have nothing to do here */
@@ -566,23 +600,17 @@ void rte_timer_manage(void)
                }
                else {
                        /* keep it in list and mark timer as pending */
+                       rte_spinlock_lock(&priv_timer[lcore_id].list_lock);
                        status.state = RTE_TIMER_PENDING;
                        __TIMER_STAT_ADD(pending, 1);
                        status.owner = (int16_t)lcore_id;
                        rte_wmb();
                        tim->status.u32 = status.u32;
                        __rte_timer_reset(tim, cur_time + tim->period,
-                                       tim->period, lcore_id, tim->f, tim->arg, 1);
+                               tim->period, lcore_id, tim->f, tim->arg, 1);
+                       rte_spinlock_unlock(&priv_timer[lcore_id].list_lock);
                }
        }
-
-       /* update the next to expire timer value */
-       priv_timer[lcore_id].pending_head.expire =
-                       (priv_timer[lcore_id].pending_head.sl_next[0] == NULL) ? 0 :
-                                       priv_timer[lcore_id].pending_head.sl_next[0]->expire;
-done:
-       /* job finished, unlock the list lock */
-       rte_spinlock_unlock(&priv_timer[lcore_id].list_lock);
 }
 
 /* dump statistics about timers */