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         ELDLM_LOCK_REPLACED = 302,
  28
  29         ELDLM_NAMESPACE_EXISTS = 400,
  30         ELDLM_BAD_NAMESPACE    = 401
  31 } ldlm_error_t;
  32
  33 #define LDLM_NAMESPACE_SERVER 0
  34 #define LDLM_NAMESPACE_CLIENT 1
  35
  36 #define LDLM_FL_LOCK_CHANGED   (1 << 0) /* extent, mode, or resource changed */
  37
  38 /* If the server returns one of these flags, then the lock was put on that list.
  39  * If the client sends one of these flags (during recovery ONLY!), it wants the
  40  * lock added to the specified list, no questions asked. -p */
  41 #define LDLM_FL_BLOCK_GRANTED  (1 << 1)
  42 #define LDLM_FL_BLOCK_CONV     (1 << 2)
  43 #define LDLM_FL_BLOCK_WAIT     (1 << 3)
  44
  45 #define LDLM_FL_CBPENDING      (1 << 4) // this lock is being destroyed
  46 #define LDLM_FL_AST_SENT       (1 << 5) // blocking or cancel packet was sent
  47 #define LDLM_FL_WAIT_NOREPROC  (1 << 6)// not a real lock flag,not saved in lock
  48 #define LDLM_FL_CANCEL         (1 << 7) // cancellation callback already run
  49
  50 /* Lock is being replayed.  This could probably be implied by the fact that one
  51  * of BLOCK_{GRANTED,CONV,WAIT} is set, but that is pretty dangerous. */
  52 #define LDLM_FL_REPLAY         (1 << 8)
  53
  54 #define LDLM_FL_INTENT_ONLY    (1 << 9) /* don't grant lock, just do intent */
  55 #define LDLM_FL_LOCAL_ONLY     (1 << 10) /* see ldlm_cli_cancel_unused */
  56 #define LDLM_FL_NO_CALLBACK    (1 << 11) /* see ldlm_cli_cancel_unused */
  57 #define LDLM_FL_HAS_INTENT     (1 << 12) /* lock request has intent */
  58 #define LDLM_FL_CANCELING      (1 << 13) /* lock cancel has already been sent */
  59 #define LDLM_FL_LOCAL          (1 << 14) // a local lock (ie, no srv/cli split)
  60
  61 /* The blocking callback is overloaded to perform two functions.  These flags
  62  * indicate which operation should be performed. */
  63 #define LDLM_CB_BLOCKING    1
  64 #define LDLM_CB_CANCELING   2
  65
  66 #define L2B(c) (1 << c)
  67
  68 /* compatibility matrix */
  69 #define LCK_COMPAT_EX  L2B(LCK_NL)
  70 #define LCK_COMPAT_PW  (LCK_COMPAT_EX | L2B(LCK_CR))
  71 #define LCK_COMPAT_PR  (LCK_COMPAT_PW | L2B(LCK_PR))
  72 #define LCK_COMPAT_CW  (LCK_COMPAT_PW | L2B(LCK_CW))
  73 #define LCK_COMPAT_CR  (LCK_COMPAT_CW | L2B(LCK_PR) | L2B(LCK_PW))
  74 #define LCK_COMPAT_NL  (LCK_COMPAT_CR | L2B(LCK_EX))
  75
  76 static ldlm_mode_t lck_compat_array[] = {
  77         [LCK_EX] LCK_COMPAT_EX,
  78         [LCK_PW] LCK_COMPAT_PW,
  79         [LCK_PR] LCK_COMPAT_PR,
  80         [LCK_CW] LCK_COMPAT_CW,
  81         [LCK_CR] LCK_COMPAT_CR,
  82         [LCK_NL] LCK_COMPAT_NL
  83 };
  84
  85 static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
  86 {
  87        if (exist < LCK_EX || exist > LCK_NL)
  88               LBUG();
  89        if (new < LCK_EX || new > LCK_NL)
  90               LBUG();
  91
  92        return (lck_compat_array[exist] & L2B(new));
  93 }
  94
  95 /*
  96  *
  97  * cluster name spaces
  98  *
  99  */
 100
 101 #define DLM_OST_NAMESPACE 1
 102 #define DLM_MDS_NAMESPACE 2
 103
 104 /* XXX
 105    - do we just separate this by security domains and use a prefix for
 106      multiple namespaces in the same domain?
 107    -
 108 */
 109
 110 struct ldlm_namespace {
 111         char                  *ns_name;
 112         __u32                  ns_client; /* is this a client-side lock tree? */
 113         struct list_head      *ns_hash; /* hash table for ns */
 114         __u32                  ns_refcount; /* count of resources in the hash */
 115         struct list_head       ns_root_list; /* all root resources in ns */
 116         struct lustre_lock     ns_lock; /* protects hash, refcount, list */
 117         struct list_head       ns_list_chain; /* position in global NS list */
 118         /*
 119         struct proc_dir_entry *ns_proc_dir;
 120         */
 121
 122         struct list_head       ns_unused_list; /* all root resources in ns */
 123         unsigned int           ns_nr_unused;
 124         unsigned int           ns_max_unused;
 125
 126         spinlock_t             ns_counter_lock;
 127         __u64                  ns_locks;
 128         __u64                  ns_resources;
 129 };
 130
 131 /*
 132  *
 133  * Resource hash table
 134  *
 135  */
 136
 137 #define RES_HASH_BITS 14
 138 #define RES_HASH_SIZE (1UL << RES_HASH_BITS)
 139 #define RES_HASH_MASK (RES_HASH_SIZE - 1)
 140
 141 struct ldlm_lock;
 142
 143 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
 144                                       struct ldlm_lock_desc *new, void *data,
 145                                       int flag);
 146 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, int flags, void *data);
 147 typedef int (*ldlm_granted_callback)(struct ldlm_lock *,
 148                                      struct lustre_msg *, int offset);
 149
 150 struct ldlm_lock {
 151         struct portals_handle l_handle; // must be first in the structure
 152         atomic_t              l_refc;
 153         struct ldlm_resource *l_resource;
 154         struct ldlm_lock     *l_parent;
 155         struct list_head      l_children;
 156         struct list_head      l_childof;
 157         struct list_head      l_lru;
 158         struct list_head      l_res_link; // position in one of three res lists
 159         struct list_head      l_export_chain; // per-export chain of locks
 160         struct list_head      l_pending_chain; // locks with callbacks pending
 161         unsigned long         l_callback_timeout;
 162
 163         ldlm_mode_t           l_req_mode;
 164         ldlm_mode_t           l_granted_mode;
 165
 166         ldlm_completion_callback l_completion_ast;
 167         ldlm_blocking_callback   l_blocking_ast;
 168         ldlm_granted_callback l_granted_cb;
 169
 170         struct obd_export    *l_export;
 171         struct lustre_handle *l_connh;
 172         __u32                 l_flags;
 173         struct lustre_handle  l_remote_handle;
 174         void                 *l_data;
 175         void                 *l_cp_data;
 176         struct ldlm_extent    l_extent;
 177         __u32                 l_version[RES_VERSION_SIZE];
 178
 179         __u32                 l_readers;
 180         __u32                 l_writers;
 181         __u8                  l_destroyed;
 182
 183         /* If the lock is granted, a process sleeps on this waitq to learn when
 184          * it's no longer in use.  If the lock is not granted, a process sleeps
 185          * on this waitq to learn when it becomes granted. */
 186         wait_queue_head_t     l_waitq;
 187 };
 188
 189 typedef int (*ldlm_res_compat)(struct ldlm_lock *child, struct ldlm_lock *new);
 190 typedef int (*ldlm_res_policy)(struct ldlm_namespace *, struct ldlm_lock **,
 191                                void *req_cookie, ldlm_mode_t mode, int flags,
 192                                void *data);
 193
 194 #define LDLM_PLAIN       10
 195 #define LDLM_EXTENT      11
 196
 197 #define LDLM_MIN_TYPE 10
 198 #define LDLM_MAX_TYPE 11
 199
 200 extern ldlm_res_compat ldlm_res_compat_table [];
 201 extern ldlm_res_policy ldlm_res_policy_table [];
 202
 203 struct ldlm_resource {
 204         struct ldlm_namespace *lr_namespace;
 205         struct list_head       lr_hash;
 206         struct ldlm_resource  *lr_parent;   /* 0 for a root resource */
 207         struct list_head       lr_children; /* list head for child resources */
 208         struct list_head       lr_childof;  /* part of ns_root_list if root res,
 209                                              * part of lr_children if child */
 210
 211         struct list_head       lr_granted;
 212         struct list_head       lr_converting;
 213         struct list_head       lr_waiting;
 214         ldlm_mode_t            lr_most_restr;
 215         __u32                  lr_type; /* LDLM_PLAIN or LDLM_EXTENT */
 216         struct ldlm_resource  *lr_root;
 217         struct ldlm_res_id     lr_name;
 218         __u32                  lr_version[RES_VERSION_SIZE];
 219         atomic_t               lr_refcount;
 220
 221         /* lr_tmp holds a list head temporarily, during the building of a work
 222          * queue.  see ldlm_add_ast_work_item and ldlm_run_ast_work */
 223         void                  *lr_tmp;
 224 };
 225
 226 struct ldlm_ast_work {
 227         struct ldlm_lock *w_lock;
 228         int               w_blocking;
 229         struct ldlm_lock_desc w_desc;
 230         struct list_head   w_list;
 231         int w_flags;
 232         void *w_data;
 233         int w_datalen;
 234 };
 235
 236 /* Per-export ldlm state. */
 237 struct ldlm_export_data {
 238         struct list_head        led_held_locks;
 239         struct obd_import       led_import;
 240 };
 241
 242 extern struct obd_ops ldlm_obd_ops;
 243
 244 extern char *ldlm_lockname[];
 245 extern char *ldlm_typename[];
 246 extern char *ldlm_it2str(int it);
 247
 248 #define __LDLM_DEBUG(level, lock, format, a...)                               \
 249 do {                                                                          \
 250         if (lock->l_resource == NULL) {                                       \
 251                 CDEBUG(level, "### " format                                   \
 252                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "\
 253                        "res: \?\? rrc=\?\? type: \?\?\? remote: "LPX64")\n"   \
 254                        , ## a, lock, lock->l_handle.h_cookie,                 \
 255                        atomic_read(&lock->l_refc),                            \
 256                        lock->l_readers, lock->l_writers,                      \
 257                        ldlm_lockname[lock->l_granted_mode],                   \
 258                        ldlm_lockname[lock->l_req_mode],                       \
 259                        lock->l_remote_handle.cookie);                         \
 260                 break;                                                        \
 261         }                                                                     \
 262         if (lock->l_resource->lr_type == LDLM_EXTENT) {                       \
 263                 CDEBUG(level, "### " format                                   \
 264                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 265                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64\
 266                        "] remote: "LPX64"\n" , ## a,                          \
 267                        lock->l_resource->lr_namespace->ns_name, lock,         \
 268                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),   \
 269                        lock->l_readers, lock->l_writers,                      \
 270                        ldlm_lockname[lock->l_granted_mode],                   \
 271                        ldlm_lockname[lock->l_req_mode],                       \
 272                        lock->l_resource->lr_name.name[0],                     \
 273                        lock->l_resource->lr_name.name[1],                     \
 274                        atomic_read(&lock->l_resource->lr_refcount),           \
 275                        ldlm_typename[lock->l_resource->lr_type],              \
 276                        lock->l_extent.start, lock->l_extent.end,              \
 277                        lock->l_remote_handle.cookie);                         \
 278                 break;                                                        \
 279         }                                                                     \
 280         {                                                                     \
 281                 CDEBUG(level, "### " format                                   \
 282                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 283                        "res: "LPU64"/"LPU64" rrc: %d type: %s remote: "LPX64  \
 284                        "\n" , ## a, lock->l_resource->lr_namespace->ns_name,  \
 285                        lock, lock->l_handle.h_cookie,                         \
 286                        atomic_read (&lock->l_refc),                           \
 287                        lock->l_readers, lock->l_writers,                      \
 288                        ldlm_lockname[lock->l_granted_mode],                   \
 289                        ldlm_lockname[lock->l_req_mode],                       \
 290                        lock->l_resource->lr_name.name[0],                     \
 291                        lock->l_resource->lr_name.name[1],                     \
 292                        atomic_read(&lock->l_resource->lr_refcount),           \
 293                        ldlm_typename[lock->l_resource->lr_type],              \
 294                        lock->l_remote_handle.cookie);                         \
 295         }                                                                     \
 296 } while (0)
 297
 298 /* I hate hate hate hate hate this.  This cannot stay.  bug 850. -phil */
 299 #define LDLM_DEBUG0(lock, format) __LDLM_DEBUG(D_DLMTRACE, lock, format"%s","")
 300
 301 #define LDLM_DEBUG(lock, format, a...) __LDLM_DEBUG(D_DLMTRACE, lock, format, a)
 302 #define LDLM_ERROR(lock, format, a...) __LDLM_DEBUG(D_ERROR, lock, format, a)
 303
 304 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
 305         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
 306
 307 /*
 308  * Iterators.
 309  */
 310
 311 #define LDLM_ITER_CONTINUE 1 /* keep iterating */
 312 #define LDLM_ITER_STOP     0 /* stop iterating */
 313
 314 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
 315 typedef int (*ldlm_res_iterator_t)(struct ldlm_resource *, void *);
 316
 317 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 318                           void *closure);
 319 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 320                            void *closure);
 321 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 322                                ldlm_res_iterator_t iter, void *closure);
 323
 324 int ldlm_replay_locks(struct obd_import *imp);
 325
 326 /* ldlm_extent.c */
 327 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
 328 int ldlm_extent_policy(struct ldlm_namespace *, struct ldlm_lock **, void *,
 329                        ldlm_mode_t, int flags, void *);
 330
 331 /* ldlm_lockd.c */
 332 int ldlm_server_blocking_ast(struct ldlm_lock *, struct ldlm_lock_desc *,
 333                              void *data, int flag);
 334 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 335 int ldlm_handle_enqueue(struct ptlrpc_request *req, ldlm_completion_callback,
 336                         ldlm_blocking_callback);
 337 int ldlm_handle_convert(struct ptlrpc_request *req);
 338 int ldlm_handle_cancel(struct ptlrpc_request *req);
 339 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
 340
 341 /* ldlm_lock.c */
 342 void ldlm_register_intent(ldlm_res_policy arg);
 343 void ldlm_unregister_intent(void);
 344 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
 345 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int flags);
 346 void ldlm_cancel_callback(struct ldlm_lock *);
 347 int ldlm_lock_set_data(struct lustre_handle *, void *data, void *cp_data);
 348 void ldlm_lock_remove_from_lru(struct ldlm_lock *);
 349
 350 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
 351 {
 352         return __ldlm_handle2lock(h, 0);
 353 }
 354
 355 #define LDLM_LOCK_PUT(lock)                     \
 356 do {                                            \
 357         /*LDLM_DEBUG((lock), "put");*/          \
 358         ldlm_lock_put(lock);                    \
 359 } while (0)
 360
 361 #define LDLM_LOCK_GET(lock)                     \
 362 ({                                              \
 363         ldlm_lock_get(lock);                    \
 364         /*LDLM_DEBUG((lock), "get");*/          \
 365         lock;                                   \
 366 })
 367
 368 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 369 void ldlm_lock_put(struct ldlm_lock *lock);
 370 void ldlm_lock_destroy(struct ldlm_lock *lock);
 371 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
 372 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 373 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
 374 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 375 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode);
 376 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen);
 377 int ldlm_lock_match(struct ldlm_namespace *ns, int flags, struct ldlm_res_id *,
 378                     __u32 type, void *cookie, int cookielen, ldlm_mode_t mode,
 379                     struct lustre_handle *);
 380 struct ldlm_lock *
 381 ldlm_lock_create(struct ldlm_namespace *ns,
 382                  struct lustre_handle *parent_lock_handle, struct ldlm_res_id,
 383                  __u32 type, ldlm_mode_t mode, void *data, void *cp_data);
 384 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *, struct ldlm_lock **,
 385                                void *cookie, int cookie_len, int *flags,
 386                                ldlm_completion_callback completion,
 387                                ldlm_blocking_callback blocking);
 388 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 389                                         int *flags);
 390 void ldlm_lock_cancel(struct ldlm_lock *lock);
 391 void ldlm_cancel_locks_for_export(struct obd_export *export);
 392 int ldlm_run_ast_work(struct list_head *rpc_list);
 393 void ldlm_reprocess_all(struct ldlm_resource *res);
 394 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns);
 395 void ldlm_lock_dump(int level, struct ldlm_lock *lock);
 396 void ldlm_lock_dump_handle(int level, struct lustre_handle *);
 397
 398 /* ldlm_test.c */
 399 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
 400 int ldlm_regression_start(struct obd_device *obddev,
 401                           struct lustre_handle *connh,
 402                           unsigned int threads, unsigned int max_locks_in,
 403                           unsigned int num_resources_in,
 404                           unsigned int num_extents_in);
 405 int ldlm_regression_stop(void);
 406
 407
 408 /* resource.c */
 409 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
 410 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
 411 int ldlm_namespace_free(struct ldlm_namespace *ns);
 412 int ldlm_proc_setup(struct obd_device *obd);
 413 void ldlm_proc_cleanup(struct obd_device *obd);
 414
 415 /* resource.c - internal */
 416 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 417                                         struct ldlm_resource *parent,
 418                                         struct ldlm_res_id, __u32 type,
 419                                         int create);
 420 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
 421 int ldlm_resource_putref(struct ldlm_resource *res);
 422 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 423                             struct ldlm_lock *lock);
 424 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 425 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
 426 void ldlm_dump_all_namespaces(void);
 427 void ldlm_namespace_dump(struct ldlm_namespace *);
 428 void ldlm_resource_dump(struct ldlm_resource *);
 429 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
 430                               struct ldlm_res_id);
 431
 432 /* ldlm_request.c */
 433 int ldlm_expired_completion_wait(void *data);
 434 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 435 int ldlm_cli_enqueue(struct lustre_handle *conn,
 436                      struct ptlrpc_request *req,
 437                      struct ldlm_namespace *ns,
 438                      struct lustre_handle *parent_lock_handle,
 439                      struct ldlm_res_id,
 440                      __u32 type,
 441                      void *cookie, int cookielen,
 442                      ldlm_mode_t mode,
 443                      int *flags,
 444                      ldlm_completion_callback completion,
 445                      ldlm_blocking_callback callback,
 446                      void *data,
 447                      void *cp_data,
 448                      struct lustre_handle *lockh);
 449 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 450                           struct ptlrpc_request *req,
 451                           struct ldlm_namespace *ns,
 452                           struct lustre_handle *parent_lock_handle,
 453                           struct ldlm_res_id,
 454                           __u32 type,
 455                           void *cookie, int cookielen,
 456                           ldlm_mode_t mode,
 457                           int *flags,
 458                           ldlm_completion_callback completion,
 459                           ldlm_blocking_callback callback,
 460                           void *data,
 461                           void *cp_data,
 462                           struct lustre_handle *lockh);
 463 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
 464                     void *data, __u32 data_len);
 465 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
 466 int ldlm_cli_cancel(struct lustre_handle *lockh);
 467 int ldlm_cli_cancel_unused(struct ldlm_namespace *, struct ldlm_res_id *,
 468                            int flags);
 469 int ldlm_cancel_lru(struct ldlm_namespace *ns);
 470
 471 /* mds/handler.c */
 472 /* This has to be here because recurisve inclusion sucks. */
 473 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 474                      void *data, int flag);
 475
 476 #endif /* __KERNEL__ */
 477
 478 /* ioctls for trying requests */
 479 #define IOC_LDLM_TYPE                   'f'
 480 #define IOC_LDLM_MIN_NR                 40
 481
 482 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
 483 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
 484 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
 485 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
 486 #define IOC_LDLM_MAX_NR                 43
 487
 488 #endif