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