Whamcloud - gitweb
LU-2924 Only wake up ldlm_poold as frequently as the check interval
[fs/lustre-release.git] / lustre / ldlm / ldlm_pool.c
index 820a50d..5289a4e 100644 (file)
@@ -518,8 +518,8 @@ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
          * It may be called when SLV has changed much, this is why we do not
          * take into account pl->pl_recalc_time here.
          */
-        RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LDLM_SYNC, 
-                               LDLM_CANCEL_LRUR));
+       RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
+                              LDLM_CANCEL_LRUR));
 }
 
 /**
@@ -551,8 +551,8 @@ static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
        spin_unlock(&ns->ns_lock);
 
         if (nr) {
-                canceled = ldlm_cancel_lru(ns, nr, LDLM_ASYNC,
-                                           LDLM_CANCEL_SHRINK);
+               canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC,
+                                          LDLM_CANCEL_SHRINK);
         }
 #ifdef __KERNEL__
         /*
@@ -589,7 +589,6 @@ int ldlm_pool_recalc(struct ldlm_pool *pl)
                 goto recalc;
 
        spin_lock(&pl->pl_lock);
-        recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
         if (recalc_interval_sec > 0) {
                 /*
                  * Update pool statistics every 1s.
@@ -609,12 +608,12 @@ int ldlm_pool_recalc(struct ldlm_pool *pl)
                 count = pl->pl_ops->po_recalc(pl);
                 lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
                                     count);
-                return count;
         }
+       recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
+                             pl->pl_recalc_period;
 
-        return 0;
+        return recalc_interval_sec;
 }
-EXPORT_SYMBOL(ldlm_pool_recalc);
 
 /**
  * Pool shrink wrapper. Will call either client or server pool recalc callback
@@ -1078,6 +1077,7 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
 {
         int total = 0, cached = 0, nr_ns;
         struct ldlm_namespace *ns;
+       struct ldlm_namespace *ns_old = NULL; /* loop detection */
         void *cookie;
 
         if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
@@ -1092,8 +1092,8 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
         /*
          * Find out how many resources we may release.
          */
-        for (nr_ns = cfs_atomic_read(ldlm_namespace_nr(client));
-             nr_ns > 0; nr_ns--)
+       for (nr_ns = ldlm_namespace_nr_read(client);
+            nr_ns > 0; nr_ns--)
         {
                mutex_lock(ldlm_namespace_lock(client));
                 if (cfs_list_empty(ldlm_namespace_list(client))) {
@@ -1102,8 +1102,23 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
                         return 0;
                 }
                 ns = ldlm_namespace_first_locked(client);
+
+               if (ns == ns_old) {
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       break;
+               }
+
+               if (ldlm_ns_empty(ns)) {
+                       ldlm_namespace_move_to_inactive_locked(ns, client);
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       continue;
+               }
+
+               if (ns_old == NULL)
+                       ns_old = ns;
+
                 ldlm_namespace_get(ns);
-                ldlm_namespace_move_locked(ns, client);
+                ldlm_namespace_move_to_active_locked(ns, client);
                mutex_unlock(ldlm_namespace_lock(client));
                 total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
                 ldlm_namespace_put(ns);
@@ -1117,8 +1132,8 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
         /*
          * Shrink at least ldlm_namespace_nr(client) namespaces.
          */
-        for (nr_ns = cfs_atomic_read(ldlm_namespace_nr(client));
-             nr_ns > 0; nr_ns--)
+       for (nr_ns = ldlm_namespace_nr_read(client) - nr_ns;
+            nr_ns > 0; nr_ns--)
         {
                 int cancel, nr_locks;
 
@@ -1138,7 +1153,7 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
                 }
                 ns = ldlm_namespace_first_locked(client);
                 ldlm_namespace_get(ns);
-                ldlm_namespace_move_locked(ns, client);
+                ldlm_namespace_move_to_active_locked(ns, client);
                mutex_unlock(ldlm_namespace_lock(client));
 
                 nr_locks = ldlm_pool_granted(&ns->ns_pool);
@@ -1167,11 +1182,13 @@ static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
                                  shrink_param(sc, gfp_mask));
 }
 
-void ldlm_pools_recalc(ldlm_side_t client)
+int ldlm_pools_recalc(ldlm_side_t client)
 {
         __u32 nr_l = 0, nr_p = 0, l;
         struct ldlm_namespace *ns;
+        struct ldlm_namespace *ns_old = NULL;
         int nr, equal = 0;
+       int time = 50; /* seconds of sleep if no active namespaces */
 
         /*
          * No need to setup pool limit for client pools.
@@ -1229,16 +1246,14 @@ void ldlm_pools_recalc(ldlm_side_t client)
                                  * for _all_ pools.
                                  */
                                 l = LDLM_POOL_HOST_L /
-                                        cfs_atomic_read(
-                                                ldlm_namespace_nr(client));
+                                       ldlm_namespace_nr_read(client);
                         } else {
                                 /*
                                  * All the rest of greedy pools will have
                                  * all locks in equal parts.
                                  */
                                 l = (LDLM_POOL_HOST_L - nr_l) /
-                                        (cfs_atomic_read(
-                                                ldlm_namespace_nr(client)) -
+                                       (ldlm_namespace_nr_read(client) -
                                          nr_p);
                         }
                         ldlm_pool_setup(&ns->ns_pool, l);
@@ -1249,7 +1264,7 @@ void ldlm_pools_recalc(ldlm_side_t client)
         /*
          * Recalc at least ldlm_namespace_nr(client) namespaces.
          */
-        for (nr = cfs_atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) {
+       for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
                 int     skip;
                 /*
                  * Lock the list, get first @ns in the list, getref, move it
@@ -1265,6 +1280,30 @@ void ldlm_pools_recalc(ldlm_side_t client)
                }
                ns = ldlm_namespace_first_locked(client);
 
+               if (ns_old == ns) { /* Full pass complete */
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       break;
+               }
+
+               /* We got an empty namespace, need to move it back to inactive
+                * list.
+                * The race with parallel resource creation is fine:
+                * - If they do namespace_get before our check, we fail the
+                *   check and they move this item to the end of the list anyway
+                * - If we do the check and then they do namespace_get, then
+                *   we move the namespace to inactive and they will move
+                *   it back to active (synchronised by the lock, so no clash
+                *   there).
+                */
+               if (ldlm_ns_empty(ns)) {
+                       ldlm_namespace_move_to_inactive_locked(ns, client);
+                       mutex_unlock(ldlm_namespace_lock(client));
+                       continue;
+               }
+
+               if (ns_old == NULL)
+                       ns_old = ns;
+
                spin_lock(&ns->ns_lock);
                /*
                 * skip ns which is being freed, and we don't want to increase
@@ -1278,17 +1317,22 @@ void ldlm_pools_recalc(ldlm_side_t client)
                }
                spin_unlock(&ns->ns_lock);
 
-               ldlm_namespace_move_locked(ns, client);
+               ldlm_namespace_move_to_active_locked(ns, client);
                mutex_unlock(ldlm_namespace_lock(client));
 
-                /*
-                 * After setup is done - recalc the pool.
-                 */
-                if (!skip) {
-                        ldlm_pool_recalc(&ns->ns_pool);
-                        ldlm_namespace_put(ns);
-                }
+               /*
+                * After setup is done - recalc the pool.
+                */
+               if (!skip) {
+                       int ttime = ldlm_pool_recalc(&ns->ns_pool);
+
+                       if (ttime < time)
+                               time = ttime;
+
+                       ldlm_namespace_put(ns);
+               }
         }
+       return time;
 }
 EXPORT_SYMBOL(ldlm_pools_recalc);
 
@@ -1296,6 +1340,7 @@ static int ldlm_pools_thread_main(void *arg)
 {
         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
         char *t_name = "ldlm_poold";
+       int s_time, c_time;
         ENTRY;
 
         cfs_daemonize(t_name);
@@ -1308,18 +1353,18 @@ static int ldlm_pools_thread_main(void *arg)
         while (1) {
                 struct l_wait_info lwi;
 
-                /*
-                 * Recal all pools on this tick.
-                 */
-                ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
-                ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
+               /*
+                * Recal all pools on this tick.
+                */
+               s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
+               c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
 
-                /*
-                 * Wait until the next check time, or until we're
-                 * stopped.
-                 */
-                lwi = LWI_TIMEOUT(cfs_time_seconds(LDLM_POOLS_THREAD_PERIOD),
-                                  NULL, NULL);
+               /*
+                * Wait until the next check time, or until we're
+                * stopped.
+                */
+               lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
+                                 NULL, NULL);
                 l_wait_event(thread->t_ctl_waitq,
                              thread_is_stopping(thread) ||
                              thread_is_event(thread),
@@ -1529,9 +1574,9 @@ void ldlm_pools_fini(void)
 }
 EXPORT_SYMBOL(ldlm_pools_fini);
 
-void ldlm_pools_recalc(ldlm_side_t client)
+int ldlm_pools_recalc(ldlm_side_t client)
 {
-        return;
+        return 0;
 }
 EXPORT_SYMBOL(ldlm_pools_recalc);
 #endif /* HAVE_LRU_RESIZE_SUPPORT */