Whamcloud - gitweb
b=13872
[fs/lustre-release.git] / lustre / include / lustre_dlm.h
index cbf2b59..393f966 100644 (file)
@@ -27,7 +27,7 @@ struct obd_device;
 
 #define OBD_LDLM_DEVICENAME  "ldlm"
 
-#define LDLM_DEFAULT_LRU_SIZE (100 * smp_num_cpus)
+#define LDLM_DEFAULT_LRU_SIZE (100 * num_online_cpus())
 #define LDLM_DEFAULT_MAX_ALIVE (cfs_time_seconds(36000))
 
 typedef enum {
@@ -42,8 +42,10 @@ typedef enum {
         ELDLM_BAD_NAMESPACE    = 401
 } ldlm_error_t;
 
-#define LDLM_NAMESPACE_SERVER 0
-#define LDLM_NAMESPACE_CLIENT 1
+typedef enum {
+        LDLM_NAMESPACE_SERVER = 1 << 0,
+        LDLM_NAMESPACE_CLIENT = 1 << 1
+} ldlm_side_t;
 
 #define LDLM_FL_LOCK_CHANGED   0x000001 /* extent, mode, or resource changed */
 
@@ -128,8 +130,21 @@ typedef enum {
  * w/o involving separate thread. in order to decrease cs rate */
 #define LDLM_FL_ATOMIC_CB      0x4000000
 
+/* It may happen that a client initiate 2 operations, e.g. unlink and mkdir,
+ * such that server send blocking ast for conflict locks to this client for
+ * the 1st operation, whereas the 2nd operation has canceled this lock and
+ * is waiting for rpc_lock which is taken by the 1st operation.
+ * LDLM_FL_BL_AST is to be set by ldlm_callback_handler() to the lock not allow
+ * ELC code to cancel it. 
+ * LDLM_FL_BL_DONE is to be set by ldlm_cancel_callback() when lock cache is
+ * droped to let ldlm_callback_handler() return EINVAL to the server. It is
+ * used when ELC rpc is already prepared and is waiting for rpc_lock, too late
+ * to send a separate CANCEL rpc. */
+#define LDLM_FL_BL_AST          0x10000000
+#define LDLM_FL_BL_DONE         0x20000000
+
 /* Cancel lock asynchronously. See ldlm_cli_cancel_unused_resource. */
-#define LDLM_FL_ASYNC           0x20000000
+#define LDLM_FL_ASYNC           0x40000000
 
 /* The blocking callback is overloaded to perform two functions.  These flags
  * indicate which operation should be performed. */
@@ -197,10 +212,74 @@ static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
  *
  */
 
+struct ldlm_pool;
 struct ldlm_lock;
 struct ldlm_resource;
 struct ldlm_namespace;
 
+typedef int (*ldlm_pool_recalc_t)(struct ldlm_pool *pl);
+                                  
+typedef int (*ldlm_pool_shrink_t)(struct ldlm_pool *pl,
+                                  int nr, unsigned int gfp_mask);
+
+enum {
+        LDLM_POOL_CTL_RECALC = 1 << 0, /* Pool recalc is enabled */
+        LDLM_POOL_CTL_SHRINK = 1 << 1, /* Pool shrink is enabled */
+        LDLM_POOL_CTL_FULL   = (LDLM_POOL_CTL_RECALC | LDLM_POOL_CTL_SHRINK)
+};
+
+/* One second for pools thread check interval. */
+#define LDLM_POOLS_THREAD_PERIOD (1)
+
+/* 5% margin for modest pools. See ldlm_pool.c for details. */
+#define LDLM_POOLS_MODEST_MARGIN (5)
+
+/* A change to SLV in % after which we want to wake up pools thread asap. */
+#define LDLM_POOLS_FAST_SLV_CHANGE (5)
+
+struct ldlm_pool {
+        /* Common pool fields */
+        cfs_proc_dir_entry_t  *pl_proc_dir;           /* Pool proc directory. */
+        char                   pl_name[100];          /* Pool name, should be long 
+                                                       * enough to contain complex 
+                                                       * proc entry name. */
+        spinlock_t             pl_lock;               /* Lock for protecting slv/clv 
+                                                       * updates. */
+        atomic_t               pl_limit;              /* Number of allowed locks in
+                                                       * in pool, both, client and 
+                                                       * server side. */
+        atomic_t               pl_granted;            /* Number of granted locks. */
+        atomic_t               pl_grant_rate;         /* Grant rate per T. */
+        atomic_t               pl_cancel_rate;        /* Cancel rate per T. */
+        atomic_t               pl_grant_speed;        /* Grant speed (GR - CR) per T. */
+        __u64                  pl_server_lock_volume; /* Server lock volume. Protected 
+                                                       * by pl_lock. */
+        cfs_time_t             pl_update_time;        /* Time when last slv from server 
+                                                       * was obtained. */
+        ldlm_pool_recalc_t     pl_recalc;             /* Recalc callback func pointer. */
+        ldlm_pool_shrink_t     pl_shrink;             /* Shrink callback func pointer. */
+        int                    pl_control;            /* Pool features mask */
+        
+        /* Server side pool fields */
+        atomic_t               pl_grant_plan;         /* Planned number of granted 
+                                                       * locks for next T. */
+        atomic_t               pl_grant_step;         /* Grant plan step for next T. */
+
+        /* Client side pool related fields */
+        atomic_t               pl_lock_volume_factor; /* Lock volume factor. */
+        struct lprocfs_stats  *pl_stats;              /* Pool statistics. */
+};
+
+static inline int pool_recalc_enabled(struct ldlm_pool *pl)
+{
+        return pl->pl_control & LDLM_POOL_CTL_RECALC;
+}
+
+static inline int pool_shrink_enabled(struct ldlm_pool *pl)
+{
+        return pl->pl_control & LDLM_POOL_CTL_SHRINK;
+}
+
 typedef int (*ldlm_res_policy)(struct ldlm_namespace *, struct ldlm_lock **,
                                void *req_cookie, ldlm_mode_t mode, int flags,
                                void *data);
@@ -211,9 +290,16 @@ struct ldlm_valblock_ops {
                            int buf_idx, int increase);
 };
 
+typedef enum {
+        LDLM_NAMESPACE_GREEDY = 1 << 0,
+        LDLM_NAMESPACE_MODEST = 1 << 1
+} ldlm_appetite_t;
+
 struct ldlm_namespace {
         char                  *ns_name;
-        __u32                  ns_client; /* is this a client-side lock tree? */
+        ldlm_side_t            ns_client; /* is this a client-side lock tree? */
+        __u64                  ns_connect_flags; /* client side connect flags 
+                                                  * supported by server */
         struct list_head      *ns_hash; /* hash table for ns */
         spinlock_t             ns_hash_lock;
         __u32                  ns_refcount; /* count of resources in the hash */
@@ -234,8 +320,36 @@ struct ldlm_namespace {
         struct ldlm_valblock_ops *ns_lvbo;
         void                  *ns_lvbp;
         cfs_waitq_t            ns_waitq;
+        struct ldlm_pool       ns_pool;
+        ldlm_appetite_t        ns_appetite;
 };
 
+static inline int ns_is_client(struct ldlm_namespace *ns)
+{
+        LASSERT(ns != NULL);
+        LASSERT(!(ns->ns_client & ~(LDLM_NAMESPACE_CLIENT | 
+                                    LDLM_NAMESPACE_SERVER)));
+        LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT ||
+                ns->ns_client == LDLM_NAMESPACE_SERVER);
+        return ns->ns_client == LDLM_NAMESPACE_CLIENT;
+}
+
+static inline int ns_is_server(struct ldlm_namespace *ns)
+{
+        LASSERT(ns != NULL);
+        LASSERT(!(ns->ns_client & ~(LDLM_NAMESPACE_CLIENT | 
+                                    LDLM_NAMESPACE_SERVER)));
+        LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT ||
+                ns->ns_client == LDLM_NAMESPACE_SERVER);
+        return ns->ns_client == LDLM_NAMESPACE_SERVER;
+}
+
+static inline int ns_connect_lru_resize(struct ldlm_namespace *ns)
+{
+        LASSERT(ns != NULL);
+        return !!(ns->ns_connect_flags & OBD_CONNECT_LRU_RESIZE);
+}
+
 /*
  *
  * Resource hash table
@@ -364,6 +478,16 @@ struct ldlm_ast_work {
         int w_datalen;
 };
 
+/* ldlm_enqueue parameters common */
+struct ldlm_enqueue_info {
+        __u32 ei_type;   /* Type of the lock being enqueued. */
+        __u32 ei_mode;   /* Mode of the lock being enqueued. */
+        void *ei_cb_bl;  /* Different callbacks for lock handling (blocking, */
+        void *ei_cb_cp;  /* completion, glimpse) */
+        void *ei_cb_gl;
+        void *ei_cbdata; /* Data to be passed into callbacks. */
+};
+
 extern struct obd_ops ldlm_obd_ops;
 
 extern char *ldlm_lockname[];
@@ -434,7 +558,6 @@ int ldlm_replay_locks(struct obd_import *imp);
 void ldlm_resource_iterate(struct ldlm_namespace *, const struct ldlm_res_id *,
                            ldlm_iterator_t iter, void *data);
 
-
 /* ldlm_flock.c */
 int ldlm_flock_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 
@@ -476,6 +599,18 @@ static inline struct ldlm_lock *ldlm_handle2lock(const struct lustre_handle *h)
         return __ldlm_handle2lock(h, 0);
 }
 
+static inline int ldlm_res_lvbo_update(struct ldlm_resource *res,
+                                       struct lustre_msg *m, int buf_idx,
+                                       int increase)
+{
+        if (res->lr_namespace->ns_lvbo &&
+            res->lr_namespace->ns_lvbo->lvbo_update) {
+                return res->lr_namespace->ns_lvbo->lvbo_update(res, m, buf_idx,
+                                                               increase);
+        }
+        return 0;
+}
+
 #define LDLM_LOCK_PUT(lock)                     \
 do {                                            \
         /*LDLM_DEBUG((lock), "put");*/          \
@@ -494,11 +629,12 @@ do {                                            \
         struct ldlm_lock *_lock, *_next;                        \
         int c = count;                                          \
         list_for_each_entry_safe(_lock, _next, head, member) {  \
+                if (c-- == 0)                                   \
+                        break;                                  \
                 list_del_init(&_lock->member);                  \
                 LDLM_LOCK_PUT(_lock);                           \
-                if (--c == 0)                                   \
-                        break;                                  \
         }                                                       \
+        LASSERT(c <= 0);                                        \
 })
 
 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
@@ -509,10 +645,10 @@ void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode);
 void ldlm_lock_allow_match(struct ldlm_lock *lock);
-int ldlm_lock_match(struct ldlm_namespace *ns, int flags,
-                    const struct ldlm_res_id *,
-                    ldlm_type_t type, ldlm_policy_data_t *, ldlm_mode_t mode,
-                    struct lustre_handle *);
+ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, int flags,
+                            const struct ldlm_res_id *, ldlm_type_t type,
+                            ldlm_policy_data_t *, ldlm_mode_t mode,
+                            struct lustre_handle *);
 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
                                         int *flags);
 void ldlm_lock_cancel(struct ldlm_lock *lock);
@@ -524,9 +660,16 @@ void ldlm_lock_dump_handle(int level, struct lustre_handle *);
 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req);
 
 /* resource.c */
-struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
+struct ldlm_namespace *ldlm_namespace_new(char *name, ldlm_side_t client, 
+                                          ldlm_appetite_t apt);
 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags);
 int ldlm_namespace_free(struct ldlm_namespace *ns, int force);
+void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client);
+struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client);
+void ldlm_namespace_get(struct ldlm_namespace *ns);
+void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup);
+void ldlm_namespace_get_nolock(struct ldlm_namespace *ns);
+void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup);
 int ldlm_proc_setup(void);
 #ifdef LPROCFS
 void ldlm_proc_cleanup(void);
@@ -545,7 +688,7 @@ void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
                             struct ldlm_lock *lock);
 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
-void ldlm_dump_all_namespaces(int level);
+void ldlm_dump_all_namespaces(ldlm_side_t client, int level);
 void ldlm_namespace_dump(int level, struct ldlm_namespace *);
 void ldlm_resource_dump(int level, struct ldlm_resource *);
 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
@@ -564,13 +707,10 @@ int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 int ldlm_glimpse_ast(struct ldlm_lock *lock, void *reqp);
 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
+                     struct ldlm_enqueue_info *einfo,
                      const struct ldlm_res_id *res_id,
-                     ldlm_type_t type, ldlm_policy_data_t *policy,
-                     ldlm_mode_t mode, int *flags,
-                     ldlm_blocking_callback blocking,
-                     ldlm_completion_callback completion,
-                     ldlm_glimpse_callback glimpse,
-                     void *data, void *lvb, __u32 lvb_len, void *lvb_swabber,
+                     ldlm_policy_data_t *policy, int *flags,
+                     void *lvb, __u32 lvb_len, void *lvb_swabber,
                      struct lustre_handle *lockh, int async);
 struct ptlrpc_request *ldlm_prep_enqueue_req(struct obd_export *exp,
                                              int bufcount, int *size,
@@ -595,6 +735,7 @@ int ldlm_cli_enqueue_local(struct ldlm_namespace *ns,
 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
                     void *data, __u32 data_len);
 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
+int ldlm_cli_update_pool(struct ptlrpc_request *req);
 int ldlm_handle_convert0(struct ptlrpc_request *req,
                          const struct ldlm_request *dlm_req);
 int ldlm_cli_cancel(struct lustre_handle *lockh);
@@ -603,7 +744,7 @@ int ldlm_cli_cancel_unused(struct ldlm_namespace *, const struct ldlm_res_id *,
 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
                                     const struct ldlm_res_id *res_id,
                                     ldlm_policy_data_t *policy,
-                                    int mode, int flags, void *opaque);
+                                    ldlm_mode_t mode, int flags, void *opaque);
 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *head,
                         int count, int flags);
 int ldlm_cli_join_lru(struct ldlm_namespace *,
@@ -650,4 +791,24 @@ static inline void check_res_locked(struct ldlm_resource *res)
 struct ldlm_resource * lock_res_and_lock(struct ldlm_lock *lock);
 void unlock_res_and_lock(struct ldlm_lock *lock);
 
+/* ldlm_pool.c */
+void ldlm_pools_recalc(ldlm_side_t client);
+int ldlm_pools_init(void);
+void ldlm_pools_fini(void);
+void ldlm_pools_wakeup(void);
+
+int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, 
+                   int idx, ldlm_side_t client);
+int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, 
+                     unsigned int gfp_mask);
+void ldlm_pool_fini(struct ldlm_pool *pl);
+int ldlm_pool_setup(struct ldlm_pool *pl, __u32 limit);
+int ldlm_pool_recalc(struct ldlm_pool *pl);
+__u64 ldlm_pool_get_slv(struct ldlm_pool *pl);
+__u32 ldlm_pool_get_limit(struct ldlm_pool *pl);
+void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv);
+void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit);
+void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock);
+void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock);
 #endif
+