Whamcloud - gitweb
LU-181 ptlrpc: reorganize ptlrpc_request
[fs/lustre-release.git] / lustre / ldlm / ldlm_pool.c
index 6cf50f2..a526f08 100644 (file)
 
 #define DEBUG_SUBSYSTEM S_LDLM
 
-#ifdef __KERNEL__
-# include <lustre_dlm.h>
-#else
-# include <liblustre.h>
-#endif
-
+#include <lustre_dlm.h>
 #include <cl_object.h>
-
 #include <obd_class.h>
 #include <obd_support.h>
 #include "ldlm_internal.h"
  */
 #define LDLM_POOL_SLV_SHIFT (10)
 
-#ifdef __KERNEL__
 extern struct proc_dir_entry *ldlm_ns_proc_dir;
-#endif
 
 static inline __u64 dru(__u64 val, __u32 shift, int round_up)
 {
@@ -223,6 +215,11 @@ static inline int ldlm_pool_t2gsp(unsigned int t)
                  (t >> LDLM_POOL_GSP_STEP_SHIFT));
 }
 
+static inline int ldlm_pool_granted(struct ldlm_pool *pl)
+{
+       return atomic_read(&pl->pl_granted);
+}
+
 /**
  * Recalculates next grant limit on passed \a pl.
  *
@@ -233,7 +230,7 @@ static void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
        int granted, grant_step, limit;
 
        limit = ldlm_pool_get_limit(pl);
-       granted = atomic_read(&pl->pl_granted);
+       granted = ldlm_pool_granted(pl);
 
        grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
        grant_step = ((limit - granted) * grant_step) / 100;
@@ -261,7 +258,7 @@ static void ldlm_pool_recalc_slv(struct ldlm_pool *pl)
        slv = pl->pl_server_lock_volume;
        grant_plan = pl->pl_grant_plan;
        limit = ldlm_pool_get_limit(pl);
-       granted = atomic_read(&pl->pl_granted);
+       granted = ldlm_pool_granted(pl);
        round_up = granted < limit;
 
         grant_usage = max_t(int, limit - (granted - grant_plan), 1);
@@ -297,7 +294,7 @@ static void ldlm_pool_recalc_stats(struct ldlm_pool *pl)
 {
        int grant_plan = pl->pl_grant_plan;
        __u64 slv = pl->pl_server_lock_volume;
-       int granted = atomic_read(&pl->pl_granted);
+       int granted = ldlm_pool_granted(pl);
        int grant_rate = atomic_read(&pl->pl_grant_rate);
        int cancel_rate = atomic_read(&pl->pl_cancel_rate);
 
@@ -386,7 +383,7 @@ static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
  * locks smaller in next 10h.
  */
 static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
-                               int nr, unsigned int gfp_mask)
+                               int nr,  gfp_t gfp_mask)
 {
        __u32 limit;
 
@@ -394,13 +391,13 @@ static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
         * VM is asking how many entries may be potentially freed.
         */
        if (nr == 0)
-               return atomic_read(&pl->pl_granted);
+               return ldlm_pool_granted(pl);
 
        /*
         * Client already canceled locks but server is already in shrinker
         * and can't cancel anything. Let's catch this race.
         */
-       if (atomic_read(&pl->pl_granted) == 0)
+       if (ldlm_pool_granted(pl) == 0)
                RETURN(0);
 
        spin_lock(&pl->pl_lock);
@@ -480,6 +477,7 @@ static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
 static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
 {
         time_t recalc_interval_sec;
+       int ret;
         ENTRY;
 
         recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
@@ -500,17 +498,13 @@ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
          * Make sure that pool knows last SLV and Limit from obd.
          */
         ldlm_cli_pool_pop_slv(pl);
-
-        pl->pl_recalc_time = cfs_time_current_sec();
-        lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
-                            recalc_interval_sec);
        spin_unlock(&pl->pl_lock);
 
         /*
          * Do not cancel locks in case lru resize is disabled for this ns.
          */
         if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
-                RETURN(0);
+               GOTO(out, ret = 0);
 
         /*
          * In the time of canceling locks on client we do not need to maintain
@@ -518,8 +512,20 @@ 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, LCF_ASYNC,
-                              LDLM_CANCEL_LRUR));
+       ret = ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
+                              LDLM_CANCEL_LRUR);
+
+out:
+       spin_lock(&pl->pl_lock);
+       /*
+        * Time of LRU resizing might be longer than period,
+        * so update after LRU resizing rather than before it.
+        */
+       pl->pl_recalc_time = cfs_time_current_sec();
+       lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
+                           recalc_interval_sec);
+       spin_unlock(&pl->pl_lock);
+       RETURN(ret);
 }
 
 /**
@@ -528,10 +534,10 @@ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
  * passed \a pl according to \a nr and \a gfp_mask.
  */
 static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
-                                int nr, unsigned int gfp_mask)
+                               int nr, gfp_t gfp_mask)
 {
         struct ldlm_namespace *ns;
-        int canceled = 0, unused;
+       int unused;
 
         ns = ldlm_pl2ns(pl);
 
@@ -550,18 +556,10 @@ static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
        unused = ns->ns_nr_unused;
        spin_unlock(&ns->ns_lock);
 
-        if (nr) {
-               canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC,
-                                          LDLM_CANCEL_SHRINK);
-        }
-#ifdef __KERNEL__
-        /*
-         * Return the number of potentially reclaimable locks.
-         */
-        return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure;
-#else
-        return unused - canceled;
-#endif
+       if (nr == 0)
+               return (unused / 100) * sysctl_vfs_cache_pressure;
+       else
+               return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
 }
 
 struct ldlm_pool_ops ldlm_srv_pool_ops = {
@@ -611,6 +609,10 @@ int ldlm_pool_recalc(struct ldlm_pool *pl)
         }
        recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
                              pl->pl_recalc_period;
+       if (recalc_interval_sec <= 0) {
+               /* Prevent too frequent recalculation. */
+               recalc_interval_sec = 1;
+       }
 
         return recalc_interval_sec;
 }
@@ -619,8 +621,7 @@ int ldlm_pool_recalc(struct ldlm_pool *pl)
  * Pool shrink wrapper. Will call either client or server pool recalc callback
  * depending what pool \a pl is used.
  */
-int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
-                     unsigned int gfp_mask)
+int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, gfp_t gfp_mask)
 {
         int cancel = 0;
 
@@ -655,7 +656,6 @@ int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
 }
 EXPORT_SYMBOL(ldlm_pool_setup);
 
-#ifdef __KERNEL__
 static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
 {
        int granted, grant_rate, cancel_rate, grant_step;
@@ -669,7 +669,7 @@ static int lprocfs_pool_state_seq_show(struct seq_file *m, void *unused)
        clv = pl->pl_client_lock_volume;
        limit = ldlm_pool_get_limit(pl);
        grant_plan = pl->pl_grant_plan;
-       granted = atomic_read(&pl->pl_granted);
+       granted = ldlm_pool_granted(pl);
        grant_rate = atomic_read(&pl->pl_grant_rate);
        cancel_rate = atomic_read(&pl->pl_cancel_rate);
        grant_speed = grant_rate - cancel_rate;
@@ -714,7 +714,8 @@ LPROC_SEQ_FOPS_RO(lprocfs_grant_plan);
 
 LDLM_POOL_PROC_READER_SEQ_SHOW(recalc_period, int);
 LDLM_POOL_PROC_WRITER(recalc_period, int);
-static ssize_t lprocfs_recalc_period_seq_write(struct file *file, const char *buf,
+static ssize_t lprocfs_recalc_period_seq_write(struct file *file,
+                                              const char __user *buf,
                                               size_t len, loff_t *off)
 {
        struct seq_file *seq = file->private_data;
@@ -840,10 +841,6 @@ static void ldlm_pool_proc_fini(struct ldlm_pool *pl)
                 pl->pl_proc_dir = NULL;
         }
 }
-#else /* !__KERNEL__*/
-#define ldlm_pool_proc_init(pl) (0)
-#define ldlm_pool_proc_fini(pl) while (0) {}
-#endif
 
 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
                   int idx, ldlm_side_t client)
@@ -1033,53 +1030,43 @@ __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl)
 }
 EXPORT_SYMBOL(ldlm_pool_get_lvf);
 
-#ifdef __KERNEL__
-static unsigned int ldlm_pool_granted(struct ldlm_pool *pl)
-{
-       return atomic_read(&pl->pl_granted);
-}
-
 static struct ptlrpc_thread *ldlm_pools_thread;
 static struct shrinker *ldlm_pools_srv_shrinker;
 static struct shrinker *ldlm_pools_cli_shrinker;
 static struct completion ldlm_pools_comp;
 
 /*
- * Cancel \a nr locks from all namespaces (if possible). Returns number of
- * cached locks after shrink is finished. All namespaces are asked to
- * cancel approximately equal amount of locks to keep balancing.
- */
-static int ldlm_pools_shrink(ldlm_side_t client, int nr,
-                             unsigned int gfp_mask)
+* count locks from all namespaces (if possible). Returns number of
+* cached locks.
+*/
+static unsigned long ldlm_pools_count(ldlm_side_t client, gfp_t gfp_mask)
 {
-       unsigned int total = 0, cached = 0;
+       unsigned long total = 0;
        int nr_ns;
-        struct ldlm_namespace *ns;
+       struct ldlm_namespace *ns;
        struct ldlm_namespace *ns_old = NULL; /* loop detection */
-        void *cookie;
+       void *cookie;
 
-        if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
-            !(gfp_mask & __GFP_FS))
-                return -1;
+       if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
+               return 0;
 
-        CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
-               nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
+       CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
+              client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
 
-        cookie = cl_env_reenter();
+       cookie = cl_env_reenter();
 
-        /*
-         * Find out how many resources we may release.
-         */
+       /*
+        * Find out how many resources we may release.
+        */
        for (nr_ns = ldlm_namespace_nr_read(client);
-            nr_ns > 0; nr_ns--)
-        {
+            nr_ns > 0; nr_ns--) {
                mutex_lock(ldlm_namespace_lock(client));
-                if (cfs_list_empty(ldlm_namespace_list(client))) {
+               if (list_empty(ldlm_namespace_list(client))) {
                        mutex_unlock(ldlm_namespace_lock(client));
-                        cl_env_reexit(cookie);
-                        return 0;
-                }
-                ns = ldlm_namespace_first_locked(client);
+                       cl_env_reexit(cookie);
+                       return 0;
+               }
+               ns = ldlm_namespace_first_locked(client);
 
                if (ns == ns_old) {
                        mutex_unlock(ldlm_namespace_lock(client));
@@ -1095,57 +1082,117 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
                if (ns_old == NULL)
                        ns_old = ns;
 
-                ldlm_namespace_get(ns);
-                ldlm_namespace_move_to_active_locked(ns, client);
+               ldlm_namespace_get(ns);
+               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);
-        }
+               total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
+               ldlm_namespace_put(ns);
+       }
 
-        if (nr == 0 || total == 0) {
-                cl_env_reexit(cookie);
-                return total;
-        }
+       cl_env_reexit(cookie);
+       return total;
+}
 
-        /*
-         * Shrink at least ldlm_namespace_nr(client) namespaces.
-         */
-       for (nr_ns = ldlm_namespace_nr_read(client) - nr_ns;
-            nr_ns > 0; nr_ns--)
-        {
-               __u64 cancel;
-               unsigned int nr_locks;
+static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr,
+                                    gfp_t gfp_mask)
+{
+       unsigned long freed = 0;
+       int tmp, nr_ns;
+       struct ldlm_namespace *ns;
+       void *cookie;
 
-                /*
-                 * Do not call shrink under ldlm_namespace_lock(client)
-                 */
+       if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
+               return -1;
+
+       cookie = cl_env_reenter();
+
+       /*
+        * Shrink at least ldlm_namespace_nr_read(client) namespaces.
+        */
+       for (tmp = nr_ns = ldlm_namespace_nr_read(client);
+            tmp > 0; tmp--) {
+               int cancel, nr_locks;
+
+               /*
+                * Do not call shrink under ldlm_namespace_lock(client)
+               */
                mutex_lock(ldlm_namespace_lock(client));
-                if (cfs_list_empty(ldlm_namespace_list(client))) {
+               if (list_empty(ldlm_namespace_list(client))) {
                        mutex_unlock(ldlm_namespace_lock(client));
-                        /*
-                         * If list is empty, we can't return any @cached > 0,
-                         * that probably would cause needless shrinker
-                         * call.
-                         */
-                        cached = 0;
-                        break;
-                }
-                ns = ldlm_namespace_first_locked(client);
-                ldlm_namespace_get(ns);
-                ldlm_namespace_move_to_active_locked(ns, client);
+                       break;
+               }
+               ns = ldlm_namespace_first_locked(client);
+               ldlm_namespace_get(ns);
+               ldlm_namespace_move_to_active_locked(ns, client);
                mutex_unlock(ldlm_namespace_lock(client));
 
-                nr_locks = ldlm_pool_granted(&ns->ns_pool);
-               cancel = (__u64)nr_locks * nr;
-               do_div(cancel, total);
-               ldlm_pool_shrink(&ns->ns_pool, 1 + cancel, gfp_mask);
-                cached += ldlm_pool_granted(&ns->ns_pool);
-                ldlm_namespace_put(ns);
-        }
-        cl_env_reexit(cookie);
-        /* we only decrease the SLV in server pools shrinker, return -1 to
-         * kernel to avoid needless loop. LU-1128 */
-        return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached;
+               nr_locks = ldlm_pool_granted(&ns->ns_pool);
+               /*
+                * We use to shrink propotionally but with new shrinker API,
+                * we lost the total number of freeable locks.
+                */
+               cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
+               freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
+               ldlm_namespace_put(ns);
+       }
+       cl_env_reexit(cookie);
+       /*
+        * we only decrease the SLV in server pools shrinker, return
+        * SHRINK_STOP to kernel to avoid needless loop. LU-1128
+        */
+       return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
+}
+
+#ifdef HAVE_SHRINKER_COUNT
+static unsigned long ldlm_pools_srv_count(struct shrinker *s,
+                                         struct shrink_control *sc)
+{
+       return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
+}
+
+static unsigned long ldlm_pools_srv_scan(struct shrinker *s,
+                                        struct shrink_control *sc)
+{
+       return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
+                              sc->gfp_mask);
+}
+
+static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
+{
+       return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
+}
+
+static unsigned long ldlm_pools_cli_scan(struct shrinker *s,
+                                        struct shrink_control *sc)
+{
+       return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
+                              sc->gfp_mask);
+}
+
+#else
+/*
+ * Cancel \a nr locks from all namespaces (if possible). Returns number of
+ * cached locks after shrink is finished. All namespaces are asked to
+ * cancel approximately equal amount of locks to keep balancing.
+ */
+static int ldlm_pools_shrink(ldlm_side_t client, int nr,
+                            gfp_t gfp_mask)
+{
+       unsigned long total = 0;
+
+       if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
+           !(gfp_mask & __GFP_FS))
+               return -1;
+
+       CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
+              nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
+
+       total = ldlm_pools_count(client, gfp_mask);
+
+       if (nr == 0 || total == 0)
+               return total;
+
+       return ldlm_pools_scan(client, nr, gfp_mask);
 }
 
 static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
@@ -1162,27 +1209,29 @@ static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
                                  shrink_param(sc, gfp_mask));
 }
 
+#endif /* HAVE_SHRINKER_COUNT */
+
 int ldlm_pools_recalc(ldlm_side_t client)
 {
-        __u32 nr_l = 0, nr_p = 0, l;
+       unsigned long 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.
-         */
-        if (client == LDLM_NAMESPACE_SERVER) {
-                /*
-                 * Check all modest namespaces first.
-                 */
+       /*
+        * No need to setup pool limit for client pools.
+        */
+       if (client == LDLM_NAMESPACE_SERVER) {
+               /*
+                * Check all modest namespaces first.
+                */
                mutex_lock(ldlm_namespace_lock(client));
-                cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
-                                        ns_list_chain)
-                {
-                        if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
-                                continue;
+               list_for_each_entry(ns, ldlm_namespace_list(client),
+                                   ns_list_chain)
+               {
+                       if (ns->ns_appetite != LDLM_NAMESPACE_MODEST)
+                               continue;
 
                         l = ldlm_pool_granted(&ns->ns_pool);
                         if (l == 0)
@@ -1204,20 +1253,20 @@ int ldlm_pools_recalc(ldlm_side_t client)
                  */
                 if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) {
                         CWARN("\"Modest\" pools eat out 2/3 of server locks "
-                              "limit (%d of %lu). This means that you have too "
+                              "limit (%lu of %lu). This means that you have too "
                               "many clients for this amount of server RAM. "
                               "Upgrade server!\n", nr_l, LDLM_POOL_HOST_L);
                         equal = 1;
                 }
 
-                /*
-                 * The rest is given to greedy namespaces.
-                 */
-                cfs_list_for_each_entry(ns, ldlm_namespace_list(client),
-                                        ns_list_chain)
-                {
-                        if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
-                                continue;
+               /*
+                * The rest is given to greedy namespaces.
+                */
+               list_for_each_entry(ns, ldlm_namespace_list(client),
+                                   ns_list_chain)
+               {
+                       if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY)
+                               continue;
 
                         if (equal) {
                                 /*
@@ -1254,7 +1303,7 @@ int ldlm_pools_recalc(ldlm_side_t client)
                  * locks synchronously.
                  */
                mutex_lock(ldlm_namespace_lock(client));
-               if (cfs_list_empty(ldlm_namespace_list(client))) {
+               if (list_empty(ldlm_namespace_list(client))) {
                        mutex_unlock(ldlm_namespace_lock(client));
                        break;
                }
@@ -1418,16 +1467,18 @@ static void ldlm_pools_thread_stop(void)
 int ldlm_pools_init(void)
 {
        int rc;
+       DEF_SHRINKER_VAR(shsvar, ldlm_pools_srv_shrink,
+                        ldlm_pools_srv_count, ldlm_pools_srv_scan);
+       DEF_SHRINKER_VAR(shcvar, ldlm_pools_cli_shrink,
+                        ldlm_pools_cli_count, ldlm_pools_cli_scan);
        ENTRY;
 
        rc = ldlm_pools_thread_start();
        if (rc == 0) {
                ldlm_pools_srv_shrinker =
-                       set_shrinker(DEFAULT_SEEKS,
-                                        ldlm_pools_srv_shrink);
+                       set_shrinker(DEFAULT_SEEKS, &shsvar);
                ldlm_pools_cli_shrinker =
-                       set_shrinker(DEFAULT_SEEKS,
-                                        ldlm_pools_cli_shrink);
+                       set_shrinker(DEFAULT_SEEKS, &shcvar);
        }
        RETURN(rc);
 }
@@ -1446,7 +1497,6 @@ void ldlm_pools_fini(void)
        ldlm_pools_thread_stop();
 }
 EXPORT_SYMBOL(ldlm_pools_fini);
-#endif /* __KERNEL__ */
 
 #else /* !HAVE_LRU_RESIZE_SUPPORT */
 int ldlm_pool_setup(struct ldlm_pool *pl, int limit)
@@ -1462,7 +1512,7 @@ int ldlm_pool_recalc(struct ldlm_pool *pl)
 EXPORT_SYMBOL(ldlm_pool_recalc);
 
 int ldlm_pool_shrink(struct ldlm_pool *pl,
-                     int nr, unsigned int gfp_mask)
+                    int nr, gfp_t gfp_mask)
 {
         return 0;
 }