lustre/include/linux/lustre_dlm.h

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * (visit-tags-table FILE)
   3  * vim:expandtab:shiftwidth=8:tabstop=8:
   4  */
   5
   6 #ifndef _LUSTRE_DLM_H__
   7 #define _LUSTRE_DLM_H__
   8
   9 #ifdef __KERNEL__
  10
  11 #include <linux/proc_fs.h>
  12 #include <linux/lustre_lib.h>
  13 #include <linux/lustre_net.h>
  14 #include <linux/lustre_import.h>
  15 #include <linux/handles.h>
  16
  17 struct obd_ops;
  18 struct obd_device;
  19
  20 #define OBD_LDLM_DEVICENAME  "ldlm"
  21
  22 typedef enum {
  23         ELDLM_OK = 0,
  24
  25         ELDLM_LOCK_CHANGED = 300,
  26         ELDLM_LOCK_ABORTED = 301,
  27
  28         ELDLM_NAMESPACE_EXISTS = 400,
  29         ELDLM_BAD_NAMESPACE    = 401
  30 } ldlm_error_t;
  31
  32 #define LDLM_NAMESPACE_SERVER 0
  33 #define LDLM_NAMESPACE_CLIENT 1
  34
  35 #define LDLM_FL_LOCK_CHANGED   (1 << 0) /* extent, mode, or resource changed */
  36
  37 /* If the server returns one of these flags, then the lock was put on that list.
  38  * If the client sends one of these flags (during recovery ONLY!), it wants the
  39  * lock added to the specified list, no questions asked. -p */
  40 #define LDLM_FL_BLOCK_GRANTED  (1 << 1)
  41 #define LDLM_FL_BLOCK_CONV     (1 << 2)
  42 #define LDLM_FL_BLOCK_WAIT     (1 << 3)
  43
  44 #define LDLM_FL_CBPENDING      (1 << 4) // this lock is being destroyed
  45 #define LDLM_FL_AST_SENT       (1 << 5) // blocking or cancel packet was sent
  46 #define LDLM_FL_WAIT_NOREPROC  (1 << 6)// not a real lock flag,not saved in lock
  47 #define LDLM_FL_CANCEL         (1 << 7) // cancellation callback already run
  48
  49 /* Lock is being replayed.  This could probably be implied by the fact that one
  50  * of BLOCK_{GRANTED,CONV,WAIT} is set, but that is pretty dangerous. */
  51 #define LDLM_FL_REPLAY         (1 << 8)
  52
  53 #define LDLM_FL_INTENT_ONLY    (1 << 9) /* don't grant lock, just do intent */
  54 #define LDLM_FL_LOCAL_ONLY     (1 << 10) /* see ldlm_cli_cancel_unused */
  55 #define LDLM_FL_NO_CALLBACK    (1 << 11) /* see ldlm_cli_cancel_unused */
  56 #define LDLM_FL_HAS_INTENT     (1 << 12) /* lock request has intent */
  57 #define LDLM_FL_CANCELING      (1 << 13) /* lock cancel has already been sent */
  58
  59 /* The blocking callback is overloaded to perform two functions.  These flags
  60  * indicate which operation should be performed. */
  61 #define LDLM_CB_BLOCKING    1
  62 #define LDLM_CB_CANCELING   2
  63
  64 #define L2B(c) (1 << c)
  65
  66 /* compatibility matrix */
  67 #define LCK_COMPAT_EX  L2B(LCK_NL)
  68 #define LCK_COMPAT_PW  (LCK_COMPAT_EX | L2B(LCK_CR))
  69 #define LCK_COMPAT_PR  (LCK_COMPAT_PW | L2B(LCK_PR))
  70 #define LCK_COMPAT_CW  (LCK_COMPAT_PW | L2B(LCK_CW))
  71 #define LCK_COMPAT_CR  (LCK_COMPAT_CW | L2B(LCK_PR) | L2B(LCK_PW))
  72 #define LCK_COMPAT_NL  (LCK_COMPAT_CR | L2B(LCK_EX))
  73
  74 static ldlm_mode_t lck_compat_array[] = {
  75         [LCK_EX] LCK_COMPAT_EX,
  76         [LCK_PW] LCK_COMPAT_PW,
  77         [LCK_PR] LCK_COMPAT_PR,
  78         [LCK_CW] LCK_COMPAT_CW,
  79         [LCK_CR] LCK_COMPAT_CR,
  80         [LCK_NL] LCK_COMPAT_NL
  81 };
  82
  83 static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
  84 {
  85        if (exist < LCK_EX || exist > LCK_NL)
  86               LBUG();
  87        if (new < LCK_EX || new > LCK_NL)
  88               LBUG();
  89
  90        return (lck_compat_array[exist] & L2B(new));
  91 }
  92
  93 /*
  94  *
  95  * cluster name spaces
  96  *
  97  */
  98
  99 #define DLM_OST_NAMESPACE 1
 100 #define DLM_MDS_NAMESPACE 2
 101
 102 /* XXX
 103    - do we just separate this by security domains and use a prefix for
 104      multiple namespaces in the same domain?
 105    -
 106 */
 107
 108 struct ldlm_namespace {
 109         char                  *ns_name;
 110         __u32                  ns_client; /* is this a client-side lock tree? */
 111         struct list_head      *ns_hash; /* hash table for ns */
 112         __u32                  ns_refcount; /* count of resources in the hash */
 113         struct list_head       ns_root_list; /* all root resources in ns */
 114         struct lustre_lock     ns_lock; /* protects hash, refcount, list */
 115         struct list_head       ns_list_chain; /* position in global NS list */
 116         /*
 117         struct proc_dir_entry *ns_proc_dir;
 118         */
 119
 120         struct list_head       ns_unused_list; /* all root resources in ns */
 121         unsigned int           ns_nr_unused;
 122         unsigned int           ns_max_unused;
 123
 124         spinlock_t             ns_counter_lock;
 125         __u64                  ns_locks;
 126         __u64                  ns_resources;
 127 };
 128
 129 /*
 130  *
 131  * Resource hash table
 132  *
 133  */
 134
 135 #define RES_HASH_BITS 14
 136 #define RES_HASH_SIZE (1UL << RES_HASH_BITS)
 137 #define RES_HASH_MASK (RES_HASH_SIZE - 1)
 138
 139 struct ldlm_lock;
 140
 141 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
 142                                       struct ldlm_lock_desc *new, void *data,
 143                                       __u32 data_len, int flag);
 144
 145 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, int flags);
 146
 147 struct ldlm_lock {
 148         __u64                 l_random;
 149         atomic_t              l_refc;
 150         struct ldlm_resource *l_resource;
 151         struct ldlm_lock     *l_parent;
 152         struct list_head      l_children;
 153         struct list_head      l_childof;
 154         struct list_head      l_lru;
 155         struct list_head      l_res_link; // position in one of three res lists
 156         struct list_head      l_export_chain; // per-export chain of locks
 157         struct list_head      l_pending_chain; // locks with callbacks pending
 158         unsigned long         l_callback_timeout;
 159
 160         ldlm_mode_t           l_req_mode;
 161         ldlm_mode_t           l_granted_mode;
 162
 163         ldlm_completion_callback l_completion_ast;
 164         ldlm_blocking_callback   l_blocking_ast;
 165
 166         struct obd_export    *l_export;
 167         struct lustre_handle *l_connh;
 168         __u32                 l_flags;
 169         struct lustre_handle  l_remote_handle;
 170         void                 *l_data;
 171         __u32                 l_data_len;
 172         struct ldlm_extent    l_extent;
 173         __u32                 l_version[RES_VERSION_SIZE];
 174
 175         __u32                 l_readers;
 176         __u32                 l_writers;
 177         __u8                  l_destroyed;
 178
 179         /* If the lock is granted, a process sleeps on this waitq to learn when
 180          * it's no longer in use.  If the lock is not granted, a process sleeps
 181          * on this waitq to learn when it becomes granted. */
 182         wait_queue_head_t     l_waitq;
 183 };
 184
 185 typedef int (*ldlm_res_compat)(struct ldlm_lock *child, struct ldlm_lock *new);
 186 typedef int (*ldlm_res_policy)(struct ldlm_lock *lock, void *req_cookie,
 187                                ldlm_mode_t mode, int flags, void *data);
 188
 189 #define LDLM_PLAIN       10
 190 #define LDLM_EXTENT      11
 191
 192 #define LDLM_MIN_TYPE 10
 193 #define LDLM_MAX_TYPE 11
 194
 195 extern ldlm_res_compat ldlm_res_compat_table [];
 196 extern ldlm_res_policy ldlm_res_policy_table [];
 197
 198 struct ldlm_resource {
 199         struct ldlm_namespace *lr_namespace;
 200         struct list_head       lr_hash;
 201         struct ldlm_resource  *lr_parent;   /* 0 for a root resource */
 202         struct list_head       lr_children; /* list head for child resources */
 203         struct list_head       lr_childof;  /* part of ns_root_list if root res,
 204                                              * part of lr_children if child */
 205
 206         struct list_head       lr_granted;
 207         struct list_head       lr_converting;
 208         struct list_head       lr_waiting;
 209         ldlm_mode_t            lr_most_restr;
 210         __u32                  lr_type; /* LDLM_PLAIN or LDLM_EXTENT */
 211         struct ldlm_resource  *lr_root;
 212         __u64                  lr_name[RES_NAME_SIZE];
 213         __u32                  lr_version[RES_VERSION_SIZE];
 214         atomic_t               lr_refcount;
 215         void                  *lr_tmp;
 216 };
 217
 218 struct ldlm_ast_work {
 219         struct ldlm_lock *w_lock;
 220         int               w_blocking;
 221         struct ldlm_lock_desc w_desc;
 222         struct list_head   w_list;
 223         int w_flags;
 224         void *w_data;
 225         int w_datalen;
 226 };
 227
 228 /* Per-export ldlm state. */
 229 struct ldlm_export_data {
 230         struct list_head        led_held_locks;
 231         struct obd_import       led_import;
 232 };
 233
 234 static inline struct ldlm_extent *ldlm_res2extent(struct ldlm_resource *res)
 235 {
 236         return (struct ldlm_extent *)(res->lr_name);
 237 }
 238
 239 extern struct obd_ops ldlm_obd_ops;
 240
 241 extern char *ldlm_lockname[];
 242 extern char *ldlm_typename[];
 243 extern char *ldlm_it2str(int it);
 244
 245 #define LDLM_DEBUG(lock, format, a...)                                        \
 246 do {                                                                          \
 247         if (lock->l_resource == NULL) {                                       \
 248                 CDEBUG(D_DLMTRACE, "### " format                              \
 249                        " ns: \?\? lock: %p lrc: %d/%d,%d mode: %s/%s "        \
 250                        "res: \?\? rrc=\?\? type: \?\?\? remote: "LPX64")\n"   \
 251                        , ## a, lock, lock->l_refc, lock->l_readers,           \
 252                        lock->l_writers,                                       \
 253                        ldlm_lockname[lock->l_granted_mode],                   \
 254                        ldlm_lockname[lock->l_req_mode],                       \
 255                        lock->l_remote_handle.addr);                           \
 256                 break;                                                        \
 257         }                                                                     \
 258         if (lock->l_resource->lr_type == LDLM_EXTENT) {                       \
 259                 CDEBUG(D_DLMTRACE, "### " format                              \
 260                        " ns: %s lock: %p lrc: %d/%d,%d mode: %s/%s res: "LPU64   \
 261                        "/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64"] remote: "  \
 262                        LPX64"\n" , ## a,                                     \
 263                        lock->l_resource->lr_namespace->ns_name, lock,         \
 264                        lock->l_refc, lock->l_readers, lock->l_writers,        \
 265                        ldlm_lockname[lock->l_granted_mode],                   \
 266                        ldlm_lockname[lock->l_req_mode],                       \
 267                        lock->l_resource->lr_name[0],                          \
 268                        lock->l_resource->lr_name[1],                          \
 269                        atomic_read(&lock->l_resource->lr_refcount),           \
 270                        ldlm_typename[lock->l_resource->lr_type],              \
 271                        lock->l_extent.start, lock->l_extent.end,              \
 272                        lock->l_remote_handle.addr);                           \
 273                 break;                                                        \
 274         }                                                                     \
 275         {                                                                     \
 276                 CDEBUG(D_DLMTRACE, "### " format                              \
 277                        " ns: %s lock: %p lrc: %d/%d,%d mode: %s/%s res: "LPU64   \
 278                        "/"LPU64" rrc: %d type: %s remote: "LPX64"\n" , ## a,   \
 279                        lock->l_resource->lr_namespace->ns_name, lock,         \
 280                        lock->l_refc, lock->l_readers, lock->l_writers,        \
 281                        ldlm_lockname[lock->l_granted_mode],                   \
 282                        ldlm_lockname[lock->l_req_mode],                       \
 283                        lock->l_resource->lr_name[0],                          \
 284                        lock->l_resource->lr_name[1],                          \
 285                        atomic_read(&lock->l_resource->lr_refcount),           \
 286                        ldlm_typename[lock->l_resource->lr_type],              \
 287                        lock->l_remote_handle.addr);                           \
 288         }                                                                     \
 289 } while (0)
 290
 291 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
 292         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
 293
 294 /*
 295  * Iterators.
 296  */
 297
 298 #define LDLM_ITER_CONTINUE 0 /* keep iterating */
 299 #define LDLM_ITER_STOP     1 /* stop iterating */
 300
 301 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
 302
 303 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 304                           void *closure);
 305 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 306                            void *closure);
 307
 308 int ldlm_replay_locks(struct obd_import *imp);
 309
 310 /* ldlm_extent.c */
 311 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
 312 int ldlm_extent_policy(struct ldlm_lock *, void *, ldlm_mode_t, int flags,
 313                        void *);
 314
 315 /* ldlm_lockd.c */
 316 int ldlm_handle_enqueue(struct ptlrpc_request *req);
 317 int ldlm_handle_convert(struct ptlrpc_request *req);
 318 int ldlm_handle_cancel(struct ptlrpc_request *req);
 319 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
 320
 321 /* ldlm_lock.c */
 322 void ldlm_register_intent(int (*arg)(struct ldlm_lock *lock, void *req_cookie,
 323                                      ldlm_mode_t mode, int flags, void *data));
 324 void ldlm_unregister_intent(void);
 325 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
 326 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int strict,
 327                                      int flags);
 328 void ldlm_cancel_callback(struct ldlm_lock *);
 329 int ldlm_lock_set_data(struct lustre_handle *, void *data, int datalen);
 330 void ldlm_lock_remove_from_lru(struct ldlm_lock *);
 331
 332 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
 333 {
 334         return __ldlm_handle2lock(h, 1, 0);
 335 }
 336
 337 #define LDLM_LOCK_PUT(lock)                     \
 338 do {                                            \
 339         /*LDLM_DEBUG(lock, "put");*/            \
 340         ldlm_lock_put(lock);                    \
 341 } while (0)
 342
 343 #define LDLM_LOCK_GET(lock)                     \
 344 ({                                              \
 345         ldlm_lock_get(lock);                    \
 346         /*LDLM_DEBUG(lock, "get");*/            \
 347         lock;                                   \
 348 })
 349
 350 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 351 void ldlm_lock_put(struct ldlm_lock *lock);
 352 void ldlm_lock_destroy(struct ldlm_lock *lock);
 353 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
 354 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 355 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
 356 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 357 void ldlm_grant_lock(struct ldlm_lock *lock);
 358 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 359                     void *cookie, int cookielen, ldlm_mode_t mode,
 360                     struct lustre_handle *lockh);
 361 struct ldlm_lock *
 362 ldlm_lock_create(struct ldlm_namespace *ns,
 363                  struct lustre_handle *parent_lock_handle,
 364                  __u64 *res_id, __u32 type, ldlm_mode_t mode, void *data,
 365                  __u32 data_len);
 366 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock *lock, void *cookie,
 367                                int cookie_len, int *flags,
 368                                ldlm_completion_callback completion,
 369                                ldlm_blocking_callback blocking);
 370 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 371                                         int *flags);
 372 void ldlm_lock_cancel(struct ldlm_lock *lock);
 373 void ldlm_cancel_locks_for_export(struct obd_export *export);
 374 void ldlm_run_ast_work(struct list_head *rpc_list);
 375 void ldlm_reprocess_all(struct ldlm_resource *res);
 376 void ldlm_lock_dump(struct ldlm_lock *lock);
 377 void ldlm_lock_dump_handle(struct lustre_handle *);
 378
 379 /* ldlm_test.c */
 380 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
 381 int ldlm_regression_start(struct obd_device *obddev,
 382                           struct lustre_handle *connh,
 383                           unsigned int threads, unsigned int max_locks_in,
 384                           unsigned int num_resources_in,
 385                           unsigned int num_extents_in);
 386 int ldlm_regression_stop(void);
 387
 388
 389 /* resource.c */
 390 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
 391 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
 392 int ldlm_namespace_free(struct ldlm_namespace *ns);
 393 int ldlm_proc_setup(struct obd_device *obd);
 394 void ldlm_proc_cleanup(struct obd_device *obd);
 395
 396 /* resource.c - internal */
 397 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 398                                         struct ldlm_resource *parent,
 399                                         __u64 *name, __u32 type, int create);
 400 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
 401 int ldlm_resource_putref(struct ldlm_resource *res);
 402 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 403                             struct ldlm_lock *lock);
 404 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 405 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
 406 void ldlm_dump_all_namespaces(void);
 407 void ldlm_namespace_dump(struct ldlm_namespace *);
 408 void ldlm_resource_dump(struct ldlm_resource *);
 409 int ldlm_lock_change_resource(struct ldlm_lock *, __u64 new_resid[3]);
 410
 411 /* ldlm_request.c */
 412 int ldlm_completion_ast(struct ldlm_lock *lock, int flags);
 413 int ldlm_cli_enqueue(struct lustre_handle *conn,
 414                      struct ptlrpc_request *req,
 415                      struct ldlm_namespace *ns,
 416                      struct lustre_handle *parent_lock_handle,
 417                      __u64 *res_id,
 418                      __u32 type,
 419                      void *cookie, int cookielen,
 420                      ldlm_mode_t mode,
 421                      int *flags,
 422                      ldlm_completion_callback completion,
 423                      ldlm_blocking_callback callback,
 424                      void *data,
 425                      __u32 data_len,
 426                      struct lustre_handle *lockh);
 427 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 428                           struct ptlrpc_request *req,
 429                           struct ldlm_namespace *ns,
 430                           struct lustre_handle *parent_lock_handle,
 431                           __u64 *res_id,
 432                           __u32 type,
 433                           void *cookie, int cookielen,
 434                           ldlm_mode_t mode,
 435                           int *flags,
 436                           ldlm_completion_callback completion,
 437                           ldlm_blocking_callback callback,
 438                           void *data,
 439                           __u32 data_len,
 440                           struct lustre_handle *lockh);
 441 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
 442                     void *data, __u32 data_len);
 443 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
 444 int ldlm_cli_cancel(struct lustre_handle *lockh);
 445 int ldlm_cli_cancel_unused(struct ldlm_namespace *, __u64 *, int flags);
 446 int ldlm_cancel_lru(struct ldlm_namespace *ns);
 447
 448 /* mds/handler.c */
 449 /* This has to be here because recurisve inclusion sucks. */
 450 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 451                      void *data, __u32 data_len, int flag);
 452
 453 #endif /* __KERNEL__ */
 454
 455 /* ioctls for trying requests */
 456 #define IOC_LDLM_TYPE                   'f'
 457 #define IOC_LDLM_MIN_NR                 40
 458
 459 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
 460 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
 461 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
 462 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
 463 #define IOC_LDLM_MAX_NR                 43
 464
 465 #endif