Whamcloud - gitweb
b=13872
[fs/lustre-release.git] / lustre / ldlm / ldlm_resource.c
index 0bce589..4f68aed 100644 (file)
 
 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
 
-struct semaphore ldlm_namespace_lock;
-struct list_head ldlm_namespace_list = CFS_LIST_HEAD_INIT(ldlm_namespace_list);
+atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
+atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
+
+struct semaphore ldlm_srv_namespace_lock;
+struct list_head ldlm_srv_namespace_list = 
+        CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list);
+
+struct semaphore ldlm_cli_namespace_lock;
+struct list_head ldlm_cli_namespace_list = 
+        CFS_LIST_HEAD_INIT(ldlm_cli_namespace_list);
+
 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
@@ -46,7 +55,8 @@ cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
                              unsigned long count, void *data)
 {
-        ldlm_dump_all_namespaces(D_DLMTRACE);
+        ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
+        ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
         RETURN(count);
 }
 
@@ -111,23 +121,19 @@ void ldlm_proc_cleanup(void)
                 lprocfs_remove(&ldlm_type_proc_dir);
 }
 
-static int lprocfs_uint_rd(char *page, char **start, off_t off,
-                           int count, int *eof, void *data)
-{
-        unsigned int *temp = (unsigned int *)data;
-        return snprintf(page, count, "%u\n", *temp);
-}
-
-static int lprocfs_read_lru_size(char *page, char **start, off_t off,
-                                 int count, int *eof, void *data)
+static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
+                               int count, int *eof, void *data)
 {
         struct ldlm_namespace *ns = data;
-        return snprintf(page, count, "%u\n", ns->ns_max_unused);
+        __u32 *nr = &ns->ns_max_unused;
+
+        if (ns_connect_lru_resize(ns))
+                nr = &ns->ns_nr_unused;
+        return lprocfs_rd_uint(page, start, off, count, eof, nr);
 }
 
-#define MAX_STRING_SIZE 128
-static int lprocfs_write_lru_size(struct file *file, const char *buffer,
-                                  unsigned long count, void *data)
+static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
+                               unsigned long count, void *data)
 {
         struct ldlm_namespace *ns = data;
         char dummy[MAX_STRING_SIZE + 1], *end;
@@ -141,24 +147,46 @@ static int lprocfs_write_lru_size(struct file *file, const char *buffer,
                 CDEBUG(D_DLMTRACE,
                        "dropping all unused locks from namespace %s\n",
                        ns->ns_name);
-                tmp = ns->ns_max_unused;
-                ns->ns_max_unused = 0;
-                ldlm_cancel_lru(ns, LDLM_SYNC);
-                ns->ns_max_unused = tmp;
+                if (ns_connect_lru_resize(ns)) {
+                        int canceled, unused  = ns->ns_nr_unused;
+                        
+                        /* Try to cancel all @ns_nr_unused locks. */
+                        canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC);
+                        if (canceled < unused) {
+                                CERROR("not all requested locks are canceled, "
+                                       "requested: %d, canceled: %d\n", unused, 
+                                       canceled);
+                                return -EINVAL;
+                        }
+                } else {
+                        tmp = ns->ns_max_unused;
+                        ns->ns_max_unused = 0;
+                        ldlm_cancel_lru(ns, 0, LDLM_SYNC);
+                        ns->ns_max_unused = tmp;
+                }
                 return count;
         }
 
         tmp = simple_strtoul(dummy, &end, 0);
-        if (tmp == 0 && *end) {
+        if (dummy == end) {
                 CERROR("invalid value written\n");
                 return -EINVAL;
         }
 
-        CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
-               ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
-        ns->ns_max_unused = (unsigned int)tmp;
-
-        ldlm_cancel_lru(ns, LDLM_ASYNC);
+        if (ns_connect_lru_resize(ns)) {
+                if (tmp > ns->ns_nr_unused)
+                        tmp = ns->ns_nr_unused;
+                tmp = ns->ns_nr_unused - tmp;
+                
+                CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n", 
+                       ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
+                ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC);
+        } else {
+                CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
+                       ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
+                ns->ns_max_unused = (unsigned int)tmp;
+                ldlm_cancel_lru(ns, 0, LDLM_ASYNC);
+        }
 
         return count;
 }
@@ -186,19 +214,27 @@ void ldlm_proc_namespace(struct ldlm_namespace *ns)
         lock_vars[0].read_fptr = lprocfs_rd_atomic;
         lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 
-        if (ns->ns_client) {
+        if (ns_is_client(ns)) {
                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
                          ns->ns_name);
                 lock_vars[0].data = &ns->ns_nr_unused;
-                lock_vars[0].read_fptr = lprocfs_uint_rd;
+                lock_vars[0].read_fptr = lprocfs_rd_uint;
                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
 
                 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
                          ns->ns_name);
                 lock_vars[0].data = ns;
-                lock_vars[0].read_fptr = lprocfs_read_lru_size;
-                lock_vars[0].write_fptr = lprocfs_write_lru_size;
+                lock_vars[0].read_fptr = lprocfs_rd_lru_size;
+                lock_vars[0].write_fptr = lprocfs_wr_lru_size;
+                lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
+
+                snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
+                         ns->ns_name);
+                lock_vars[0].data = &ns->ns_max_age;
+                lock_vars[0].read_fptr = lprocfs_rd_uint;
+                lock_vars[0].write_fptr = lprocfs_wr_uint;
                 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
+
         }
 }
 #undef MAX_STRING_SIZE
@@ -206,11 +242,12 @@ void ldlm_proc_namespace(struct ldlm_namespace *ns)
 #define ldlm_proc_namespace(ns) do {} while (0)
 #endif /* LPROCFS */
 
-struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
+struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client, 
+                                          ldlm_appetite_t apt)
 {
         struct ldlm_namespace *ns = NULL;
         struct list_head *bucket;
-        int rc;
+        int rc, idx, namelen;
         ENTRY;
 
         rc = ldlm_get_ref();
@@ -219,7 +256,7 @@ struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
                 RETURN(NULL);
         }
 
-        OBD_ALLOC(ns, sizeof(*ns));
+        OBD_ALLOC_PTR(ns);
         if (!ns)
                 GOTO(out_ref, NULL);
 
@@ -227,7 +264,9 @@ struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
         if (!ns->ns_hash)
                 GOTO(out_ns, NULL);
 
-        OBD_ALLOC(ns->ns_name, strlen(name) + 1);
+        ns->ns_appetite = apt;
+        namelen = strlen(name);
+        OBD_ALLOC(ns->ns_name, namelen + 1);
         if (!ns->ns_name)
                 GOTO(out_hash, NULL);
 
@@ -248,19 +287,31 @@ struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
         CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
         ns->ns_nr_unused = 0;
         ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
+        ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
         spin_lock_init(&ns->ns_unused_lock);
-
-        mutex_down(&ldlm_namespace_lock);
-        list_add(&ns->ns_list_chain, &ldlm_namespace_list);
-        mutex_up(&ldlm_namespace_lock);
+        ns->ns_connect_flags = 0;
         ldlm_proc_namespace(ns);
-        RETURN(ns);
 
+        idx = atomic_read(ldlm_namespace_nr(client));
+        rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
+        if (rc) {
+                CERROR("Can't initialize lock pool, rc %d\n", rc);
+                GOTO(out_proc, rc);
+        }
+
+        mutex_down(ldlm_namespace_lock(client));
+        list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
+        atomic_inc(ldlm_namespace_nr(client));
+        mutex_up(ldlm_namespace_lock(client));
+
+        RETURN(ns);
+out_proc:
+        ldlm_namespace_cleanup(ns, 0);
+        OBD_FREE(ns->ns_name, namelen + 1);
 out_hash:
-        POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
 out_ns:
-        OBD_FREE(ns, sizeof(*ns));
+        OBD_FREE_PTR(ns);
 out_ref:
         ldlm_put_ref(0);
         RETURN(NULL);
@@ -277,7 +328,7 @@ static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
                              int flags)
 {
         struct list_head *tmp;
-        int rc = 0, client = res->lr_namespace->ns_client;
+        int rc = 0, client = ns_is_client(res->lr_namespace);
         int local_only = (flags & LDLM_FL_LOCAL_ONLY);
         ENTRY;
 
@@ -310,13 +361,15 @@ static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
                 lock->l_flags |= LDLM_FL_FAILED;
                 lock->l_flags |= flags;
 
+                /* ... without sending a CANCEL message for local_only. */
+                if (local_only)
+                        lock->l_flags |= LDLM_FL_LOCAL_ONLY;
+
                 if (local_only && (lock->l_readers || lock->l_writers)) {
                         /* This is a little bit gross, but much better than the
                          * alternative: pretend that we got a blocking AST from
                          * the server, so that when the lock is decref'd, it
                          * will go away ... */
-                        /* ... without sending a CANCEL message. */
-                        lock->l_flags |= LDLM_FL_LOCAL_ONLY;
                         unlock_res(res);
                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
                         if (lock->l_completion_ast)
@@ -330,14 +383,9 @@ static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
 
                         unlock_res(res);
                         ldlm_lock2handle(lock, &lockh);
-                        if (!local_only) {
-                                rc = ldlm_cli_cancel(&lockh);
-                                if (rc)
-                                        CERROR("ldlm_cli_cancel: %d\n", rc);
-                        }
-                        /* Force local cleanup on errors, too. */
-                        if (local_only || rc != ELDLM_OK)
-                                ldlm_lock_cancel(lock);
+                        rc = ldlm_cli_cancel(&lockh);
+                        if (rc)
+                                CERROR("ldlm_cli_cancel: %d\n", rc);
                 } else {
                         ldlm_resource_unlink_lock(lock);
                         unlock_res(res);
@@ -401,9 +449,11 @@ int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
         if (!ns)
                 RETURN(ELDLM_OK);
 
-        mutex_down(&ldlm_namespace_lock);
+        mutex_down(ldlm_namespace_lock(ns->ns_client));
         list_del(&ns->ns_list_chain);
-        mutex_up(&ldlm_namespace_lock);
+        atomic_dec(ldlm_namespace_nr(ns->ns_client));
+        ldlm_pool_fini(&ns->ns_pool);
+        mutex_up(ldlm_namespace_lock(ns->ns_client));
 
         /* At shutdown time, don't call the cancellation callback */
         ldlm_namespace_cleanup(ns, 0);
@@ -441,16 +491,59 @@ int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
                        "dlm namespace %s free done waiting\n", ns->ns_name);
         }
 
-        POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
         OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
         OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
-        OBD_FREE(ns, sizeof(*ns));
+        OBD_FREE_PTR(ns);
 
         ldlm_put_ref(force);
 
         RETURN(ELDLM_OK);
 }
 
+void ldlm_namespace_get_nolock(struct ldlm_namespace *ns)
+{
+        LASSERT(ns->ns_refcount >= 0);
+        ns->ns_refcount++;
+}
+
+void ldlm_namespace_get(struct ldlm_namespace *ns)
+{
+        spin_lock(&ns->ns_hash_lock);
+        ldlm_namespace_get_nolock(ns);
+        spin_unlock(&ns->ns_hash_lock);
+}
+
+void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup)
+{
+        LASSERT(ns->ns_refcount > 0);
+        ns->ns_refcount--;
+        if (ns->ns_refcount == 0 && wakeup)
+                wake_up(&ns->ns_waitq);
+}
+
+void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
+{
+        spin_lock(&ns->ns_hash_lock);
+        ldlm_namespace_put_nolock(ns, wakeup);
+        spin_unlock(&ns->ns_hash_lock);
+}
+
+/* Should be called under ldlm_namespace_lock(client) taken */
+void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client)
+{
+        LASSERT(!list_empty(&ns->ns_list_chain));
+        LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
+        list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
+}
+
+/* Should be called under ldlm_namespace_lock(client) taken */
+struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client)
+{
+        LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
+        LASSERT(!list_empty(ldlm_namespace_list(client)));
+        return container_of(ldlm_namespace_list(client)->next, 
+                struct ldlm_namespace, ns_list_chain);
+}
 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
                           const struct ldlm_res_id *name)
 {
@@ -551,7 +644,7 @@ ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
         bucket = ns->ns_hash + hash;
         list_add(&res->lr_hash, bucket);
         ns->ns_resources++;
-        ns->ns_refcount++;
+        ldlm_namespace_get_nolock(ns);
 
         if (parent == NULL) {
                 list_add(&res->lr_childof, &ns->ns_root_list);
@@ -648,7 +741,9 @@ void __ldlm_resource_putref_final(struct ldlm_resource *res)
                 LBUG();
         }
 
-        ns->ns_refcount--;
+        /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few 
+         * lines below when all children are freed. */
+        ldlm_namespace_put_nolock(ns, 0);
         list_del_init(&res->lr_hash);
         list_del_init(&res->lr_childof);
 
@@ -760,22 +855,22 @@ void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
         desc->lr_name = res->lr_name;
 }
 
-void ldlm_dump_all_namespaces(int level)
+void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
 {
         struct list_head *tmp;
 
         if (!((libcfs_debug | D_ERROR) & level))
                 return;
 
-        mutex_down(&ldlm_namespace_lock);
+        mutex_down(ldlm_namespace_lock(client));
 
-        list_for_each(tmp, &ldlm_namespace_list) {
+        list_for_each(tmp, ldlm_namespace_list(client)) {
                 struct ldlm_namespace *ns;
                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
                 ldlm_namespace_dump(level, ns);
         }
 
-        mutex_up(&ldlm_namespace_lock);
+        mutex_up(ldlm_namespace_lock(client));
 }
 
 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
@@ -785,8 +880,9 @@ void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
         if (!((libcfs_debug | D_ERROR) & level))
                 return;
 
-        CDEBUG(level, "--- Namespace: %s (rc: %d, client: %d)\n",
-                  ns->ns_name, ns->ns_refcount, ns->ns_client);
+        CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n", 
+               ns->ns_name, ns->ns_refcount, 
+               ns_is_client(ns) ? "client" : "server");
 
         if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
                 return;