Whamcloud - gitweb
1. Bug fix for Bug #320, also needs repair of lconf, since it tries
[fs/lustre-release.git] / lustre / ldlm / ldlm_resource.c
index da263a7..53b835a 100644 (file)
 #define DEBUG_SUBSYSTEM S_LDLM
 
 #include <linux/lustre_dlm.h>
+#include <linux/obd_class.h>
 
 kmem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
 
-struct ldlm_namespace *ldlm_namespace_new(__u32 local)
+spinlock_t ldlm_namespace_lock = SPIN_LOCK_UNLOCKED;
+struct list_head ldlm_namespace_list = LIST_HEAD_INIT(ldlm_namespace_list);
+static struct proc_dir_entry *ldlm_ns_proc_dir = NULL;
+
+int ldlm_proc_setup(struct obd_device *obd)
+{
+        ENTRY;
+        LASSERT(ldlm_ns_proc_dir == NULL);
+        ldlm_ns_proc_dir=obd->obd_type->typ_procroot;
+        RETURN(0);
+}
+
+void ldlm_proc_cleanup(struct obd_device *obd)
+{
+        ldlm_ns_proc_dir = NULL;
+}
+
+#define MAX_STRING_SIZE 100
+void ldlm_proc_namespace(struct ldlm_namespace *ns)
+{
+        struct lprocfs_vars lock_vars[2];
+        char lock_names[MAX_STRING_SIZE];
+
+        memset(lock_vars, 0, sizeof(lock_vars));
+        snprintf(lock_names, MAX_STRING_SIZE, "%s/resource_count", 
+                 ns->ns_name);
+        lock_names[MAX_STRING_SIZE] = '\0';
+        lock_vars[0].name = lock_names;
+        lock_vars[0].read_fptr = lprocfs_ll_rd;
+        lock_vars[0].write_fptr = NULL;
+        lock_vars[0].data = &ns->ns_resources;
+        lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
+
+        memset(lock_vars, 0, sizeof(lock_vars));
+        snprintf(lock_names, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
+        lock_names[MAX_STRING_SIZE] = '\0';
+        lock_vars[0].name = lock_names;
+        lock_vars[0].read_fptr = lprocfs_ll_rd;
+        lock_vars[0].write_fptr = NULL;
+        lock_vars[0].data = &ns->ns_locks;
+        lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
+
+}
+#undef MAX_STRING_SIZE
+
+#define LDLM_MAX_UNUSED 20
+struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
 {
         struct ldlm_namespace *ns = NULL;
         struct list_head *bucket;
 
+       
+
         OBD_ALLOC(ns, sizeof(*ns));
         if (!ns) {
                 LBUG();
@@ -31,98 +80,153 @@ struct ldlm_namespace *ldlm_namespace_new(__u32 local)
                 LBUG();
                 GOTO(out, ns);
         }
+        obd_memory += sizeof(*ns->ns_hash) * RES_HASH_SIZE;
 
-        ptlrpc_init_client(NULL, NULL,
-                           LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
-                           &ns->ns_client);
+        OBD_ALLOC(ns->ns_name, strlen(name) + 1);
+        if (!ns->ns_name) {
+                LBUG();
+                GOTO(out, ns);
+        }
+        strcpy(ns->ns_name, name);
 
         INIT_LIST_HEAD(&ns->ns_root_list);
-        ns->ns_lock = SPIN_LOCK_UNLOCKED;
+        l_lock_init(&ns->ns_lock);
         ns->ns_refcount = 0;
-        ns->ns_local = local;
+        ns->ns_client = client;
+        spin_lock_init(&ns->ns_counter_lock);
+        ns->ns_locks = 0;
+        ns->ns_resources = 0;
 
         for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
              bucket--)
                 INIT_LIST_HEAD(bucket);
+
+        INIT_LIST_HEAD(&ns->ns_unused_list);
+        ns->ns_nr_unused = 0;
+        ns->ns_max_unused = LDLM_MAX_UNUSED;
+
+        spin_lock(&ldlm_namespace_lock);
+        list_add(&ns->ns_list_chain, &ldlm_namespace_list);
+        spin_unlock(&ldlm_namespace_lock);
+        ldlm_proc_namespace(ns);
         RETURN(ns);
 
- out: 
-        if (ns && ns->ns_hash)
+ out:
+        if (ns && ns->ns_hash) {
                 vfree(ns->ns_hash);
-        if (ns) 
+                obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
+        }
+        if (ns && ns->ns_name)
+                OBD_FREE(ns->ns_name, strlen(name) + 1);
+        if (ns)
                 OBD_FREE(ns, sizeof(*ns));
         return NULL;
 }
 
-static int cleanup_resource(struct ldlm_resource *res, struct list_head *q)
+extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
+
+/* If 'local_only' is true, don't try to tell the server, just cleanup. */
+static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
+                             int local_only)
 {
         struct list_head *tmp, *pos;
-        int rc = 0, client = res->lr_namespace->ns_local;
+        int rc = 0, client = res->lr_namespace->ns_client;
         ENTRY;
 
         list_for_each_safe(tmp, pos, q) {
                 struct ldlm_lock *lock;
                 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
+                LDLM_LOCK_GET(lock);
+
+                /* At shutdown time, don't call the cancellation callback */
+                lock->l_flags |= LDLM_FL_CANCEL;
 
                 if (client) {
-                        rc = ldlm_cli_cancel(lock->l_client, lock);
-                        if (rc < 0) {
-                                CERROR("ldlm_cli_cancel: %d\n", rc);
-                                LBUG();
+                        struct lustre_handle lockh;
+                        ldlm_lock2handle(lock, &lockh);
+                        if (!local_only) {
+                                rc = ldlm_cli_cancel(&lockh);
+                                if (rc)
+                                        CERROR("ldlm_cli_cancel: %d\n", rc);
                         }
-                        if (rc == ELDLM_RESOURCE_FREED)
-                                rc = 1;
+                        /* Force local cleanup on errors, too. */
+                        if (local_only || rc != ELDLM_OK)
+                                ldlm_lock_cancel(lock);
                 } else {
-                        CERROR("Freeing a lock still held by a client node.\n");
-
-                        spin_lock(&lock->l_lock);
-                        ldlm_resource_del_lock(lock);
-                        ldlm_lock_free(lock);
+                        LDLM_DEBUG(lock, "Freeing a lock still held by a "
+                                   "client node.\n");
 
-                        rc = ldlm_resource_put(res);
+                        ldlm_resource_unlink_lock(lock);
+                        ldlm_lock_destroy(lock);
                 }
+                LDLM_LOCK_PUT(lock);
         }
-
-        RETURN(rc);
 }
 
-int ldlm_namespace_free(struct ldlm_namespace *ns)
+int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only)
 {
-        struct list_head *tmp, *pos;
-        int i, rc;
+        int i;
 
-        if (!ns)
-                RETURN(ELDLM_OK);
-        /* We should probably take the ns_lock, but then ldlm_resource_put
-         * couldn't take it.  Hmm. */
+        l_lock(&ns->ns_lock);
         for (i = 0; i < RES_HASH_SIZE; i++) {
+                struct list_head *tmp, *pos;
                 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
                         struct ldlm_resource *res;
                         res = list_entry(tmp, struct ldlm_resource, lr_hash);
-
-                        spin_lock(&res->lr_lock);
-                        rc = cleanup_resource(res, &res->lr_granted);
-                        if (!rc)
-                                rc = cleanup_resource(res, &res->lr_converting);
-                        if (!rc)
-                                rc = cleanup_resource(res, &res->lr_waiting);
-
-                        if (rc == 0) {
-                                CERROR("Resource refcount nonzero after lock "
-                                       "cleanup; forcing cleanup.\n");
-                                res->lr_refcount = 1;
-                                rc = ldlm_resource_put(res);
+                        ldlm_resource_getref(res);
+
+                        cleanup_resource(res, &res->lr_granted, local_only);
+                        cleanup_resource(res, &res->lr_converting, local_only);
+                        cleanup_resource(res, &res->lr_waiting, local_only);
+
+                        /* XXX this is a bit counter-intuitive and should
+                         * probably be cleaner: don't force cleanup if we're
+                         * local_only (which is only used by recovery).  We
+                         * probably still have outstanding lock refs which
+                         * reference these resources. -phil */
+                        if (!ldlm_resource_put(res) && !local_only) {
+                                CERROR("Resource refcount nonzero (%d) after "
+                                       "lock cleanup; forcing cleanup.\n",
+                                       atomic_read(&res->lr_refcount));
+                                ldlm_resource_dump(res);
+                                atomic_set(&res->lr_refcount, 1);
+                                ldlm_resource_put(res);
                         }
                 }
         }
+        l_unlock(&ns->ns_lock);
+
+        return ELDLM_OK;
+}
+
+/* Cleanup, but also free, the namespace */
+int ldlm_namespace_free(struct ldlm_namespace *ns)
+{
+        if (!ns)
+                RETURN(ELDLM_OK);
 
-        vfree(ns->ns_hash /* , sizeof(struct list_head) * RES_HASH_SIZE */);
-        ptlrpc_cleanup_client(&ns->ns_client); 
+        spin_lock(&ldlm_namespace_lock);
+        list_del(&ns->ns_list_chain);
+
+        spin_unlock(&ldlm_namespace_lock);
+
+        ldlm_namespace_cleanup(ns, 0);
+
+        vfree(ns->ns_hash /* , sizeof(*ns->ns_hash) * RES_HASH_SIZE */);
+        obd_memory -= sizeof(*ns->ns_hash) * RES_HASH_SIZE;
+        OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
         OBD_FREE(ns, sizeof(*ns));
 
         return ELDLM_OK;
 }
 
+int ldlm_client_free(struct obd_export *exp)
+{
+        struct ldlm_export_data *led = &exp->exp_ldlm_data;
+        ptlrpc_cleanup_client(&led->led_import);
+        RETURN(0);
+}
+
 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, __u64 *name)
 {
         __u32 hash = 0;
@@ -153,8 +257,7 @@ static struct ldlm_resource *ldlm_resource_new(void)
         INIT_LIST_HEAD(&res->lr_converting);
         INIT_LIST_HEAD(&res->lr_waiting);
 
-        res->lr_lock = SPIN_LOCK_UNLOCKED;
-        res->lr_refcount = 1;
+        atomic_set(&res->lr_refcount, 1);
 
         return res;
 }
@@ -169,7 +272,7 @@ static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
         struct ldlm_resource *res;
         ENTRY;
 
-        if (type < 0 || type > LDLM_MAX_TYPE) {
+        if (type < LDLM_MIN_TYPE || type > LDLM_MAX_TYPE) {
                 LBUG();
                 RETURN(NULL);
         }
@@ -180,20 +283,23 @@ static struct ldlm_resource *ldlm_resource_add(struct ldlm_namespace *ns,
                 RETURN(NULL);
         }
 
+        spin_lock(&ns->ns_counter_lock);
+        ns->ns_resources++;
+        spin_unlock(&ns->ns_counter_lock);
+
         memcpy(res->lr_name, name, sizeof(res->lr_name));
         res->lr_namespace = ns;
         ns->ns_refcount++;
 
-        res->lr_type = type; 
+        res->lr_type = type;
         res->lr_most_restr = LCK_NL;
 
         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
         list_add(&res->lr_hash, bucket);
 
-        if (parent == NULL) {
-                res->lr_parent = res;
-                list_add(&res->lr_rootlink, &ns->ns_root_list);
-        } else {
+        if (parent == NULL)
+                list_add(&res->lr_childof, &ns->ns_root_list);
+        else {
                 res->lr_parent = parent;
                 list_add(&res->lr_childof, &parent->lr_children);
         }
@@ -218,7 +324,7 @@ struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
                 RETURN(NULL);
         }
 
-        spin_lock(&ns->ns_lock);
+        l_lock(&ns->ns_lock);
         bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
 
         list_for_each(tmp, bucket) {
@@ -227,9 +333,7 @@ struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 
                 if (memcmp(chk->lr_name, name, sizeof(chk->lr_name)) == 0) {
                         res = chk;
-                        spin_lock(&res->lr_lock);
-                        res->lr_refcount++;
-                        spin_unlock(&res->lr_lock);
+                        atomic_inc(&res->lr_refcount);
                         EXIT;
                         break;
                 }
@@ -237,29 +341,31 @@ struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 
         if (res == NULL && create)
                 res = ldlm_resource_add(ns, parent, name, type);
-        spin_unlock(&ns->ns_lock);
+        l_unlock(&ns->ns_lock);
 
         RETURN(res);
 }
 
-/* Args: locked resource
- * Locks: takes and releases res->lr_lock
- *        takes and releases ns->ns_lock iff res->lr_refcount falls to 0
- */
+struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
+{
+        atomic_inc(&res->lr_refcount);
+        return res;
+}
+
+/* Returns 1 if the resource was freed, 0 if it remains. */
 int ldlm_resource_put(struct ldlm_resource *res)
 {
         int rc = 0;
 
-        if (res->lr_refcount == 1) {
+        if (atomic_dec_and_test(&res->lr_refcount)) {
                 struct ldlm_namespace *ns = res->lr_namespace;
                 ENTRY;
 
-                spin_unlock(&res->lr_lock);
-                spin_lock(&ns->ns_lock);
-                spin_lock(&res->lr_lock);
+                l_lock(&ns->ns_lock);
 
-                if (res->lr_refcount != 1) {
-                        spin_unlock(&ns->ns_lock);
+                if (atomic_read(&res->lr_refcount) != 0) {
+                        /* We lost the race. */
+                        l_unlock(&ns->ns_lock);
                         goto out;
                 }
 
@@ -277,47 +383,46 @@ int ldlm_resource_put(struct ldlm_resource *res)
 
                 ns->ns_refcount--;
                 list_del(&res->lr_hash);
-                list_del(&res->lr_rootlink);
                 list_del(&res->lr_childof);
 
                 kmem_cache_free(ldlm_resource_slab, res);
-                spin_unlock(&ns->ns_lock);
+                l_unlock(&ns->ns_lock);
+
+                spin_lock(&ns->ns_counter_lock);
+                ns->ns_resources--;
+                spin_unlock(&ns->ns_counter_lock);
+
                 rc = 1;
         } else {
                 ENTRY;
         out:
-                res->lr_refcount--;
-                if (res->lr_refcount < 0)
+                if (atomic_read(&res->lr_refcount) < 0)
                         LBUG();
         }
 
-        RETURN(rc); 
+        RETURN(rc);
 }
 
-/* Must be called with resource->lr_lock taken */
 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
                             struct ldlm_lock *lock)
 {
-        ldlm_resource_dump(res); 
-        ldlm_lock_dump(lock); 
+        l_lock(&res->lr_namespace->ns_lock);
+
+        ldlm_resource_dump(res);
+        ldlm_lock_dump(lock);
+
         if (!list_empty(&lock->l_res_link))
                 LBUG();
 
         list_add(&lock->l_res_link, head);
-        res->lr_refcount++;
+        l_unlock(&res->lr_namespace->ns_lock);
 }
 
-/* Must be called with resource->lr_lock taken */
-void ldlm_resource_del_lock(struct ldlm_lock *lock)
+void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
 {
+        l_lock(&lock->l_resource->lr_namespace->ns_lock);
         list_del_init(&lock->l_res_link);
-        lock->l_resource->lr_refcount--;
-}
-
-int ldlm_get_resource_handle(struct ldlm_resource *res, struct lustre_handle *h)
-{
-        LBUG();
-        return 0;
+        l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 }
 
 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
@@ -327,6 +432,40 @@ void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
         memcpy(desc->lr_version, res->lr_version, sizeof(desc->lr_version));
 }
 
+void ldlm_dump_all_namespaces(void)
+{
+        struct list_head *tmp;
+
+        spin_lock(&ldlm_namespace_lock);
+
+        list_for_each(tmp, &ldlm_namespace_list) {
+                struct ldlm_namespace *ns;
+                ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
+                ldlm_namespace_dump(ns);
+        }
+
+        spin_unlock(&ldlm_namespace_lock);
+}
+
+void ldlm_namespace_dump(struct ldlm_namespace *ns)
+{
+        struct list_head *tmp;
+
+        l_lock(&ns->ns_lock);
+        CDEBUG(D_OTHER, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
+               ns->ns_refcount, ns->ns_client);
+
+        list_for_each(tmp, &ns->ns_root_list) {
+                struct ldlm_resource *res;
+                res = list_entry(tmp, struct ldlm_resource, lr_childof);
+
+                /* Once we have resources with children, this should really dump
+                 * them recursively. */
+                ldlm_resource_dump(res);
+        }
+        l_unlock(&ns->ns_lock);
+}
+
 void ldlm_resource_dump(struct ldlm_resource *res)
 {
         struct list_head *tmp;
@@ -340,8 +479,10 @@ void ldlm_resource_dump(struct ldlm_resource *res)
                  (unsigned long long)res->lr_name[1],
                  (unsigned long long)res->lr_name[2]);
 
-        CDEBUG(D_OTHER, "--- Resource: %p (%s)\n", res, name);
-        CDEBUG(D_OTHER, "Namespace: %p\n", res->lr_namespace);
+        CDEBUG(D_OTHER, "--- Resource: %p (%s) (rc: %d)\n", res, name,
+               atomic_read(&res->lr_refcount));
+        CDEBUG(D_OTHER, "Namespace: %p (%s)\n", res->lr_namespace,
+               res->lr_namespace->ns_name);
         CDEBUG(D_OTHER, "Parent: %p, root: %p\n", res->lr_parent, res->lr_root);
 
         CDEBUG(D_OTHER, "Granted locks:\n");