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 # include <linux/proc_fs.h>
  11 #endif
  12
  13 #include <linux/lustre_lib.h>
  14 #include <linux/lustre_net.h>
  15 #include <linux/lustre_import.h>
  16 #include <linux/lustre_handles.h>
  17 #include <linux/lustre_export.h> /* for obd_export, for LDLM_DEBUG */
  18
  19 struct obd_ops;
  20 struct obd_device;
  21
  22 #define OBD_LDLM_DEVICENAME  "ldlm"
  23
  24 typedef enum {
  25         ELDLM_OK = 0,
  26
  27         ELDLM_LOCK_CHANGED = 300,
  28         ELDLM_LOCK_ABORTED = 301,
  29         ELDLM_LOCK_REPLACED = 302,
  30
  31         ELDLM_NAMESPACE_EXISTS = 400,
  32         ELDLM_BAD_NAMESPACE    = 401
  33 } ldlm_error_t;
  34
  35 #define LDLM_NAMESPACE_SERVER 0
  36 #define LDLM_NAMESPACE_CLIENT 1
  37
  38 #define LDLM_FL_LOCK_CHANGED   (1 << 0) /* extent, mode, or resource changed */
  39
  40 /* If the server returns one of these flags, then the lock was put on that list.
  41  * If the client sends one of these flags (during recovery ONLY!), it wants the
  42  * lock added to the specified list, no questions asked. -p */
  43 #define LDLM_FL_BLOCK_GRANTED  (1 << 1)
  44 #define LDLM_FL_BLOCK_CONV     (1 << 2)
  45 #define LDLM_FL_BLOCK_WAIT     (1 << 3)
  46
  47 #define LDLM_FL_CBPENDING      (1 << 4) // this lock is being destroyed
  48 #define LDLM_FL_AST_SENT       (1 << 5) // blocking or cancel packet was sent
  49 #define LDLM_FL_WAIT_NOREPROC  (1 << 6)// not a real lock flag,not saved in lock
  50 #define LDLM_FL_CANCEL         (1 << 7) // cancellation callback already run
  51
  52 /* Lock is being replayed.  This could probably be implied by the fact that one
  53  * of BLOCK_{GRANTED,CONV,WAIT} is set, but that is pretty dangerous. */
  54 #define LDLM_FL_REPLAY         (1 << 8)
  55
  56 #define LDLM_FL_INTENT_ONLY    (1 << 9) /* don't grant lock, just do intent */
  57 #define LDLM_FL_LOCAL_ONLY     (1 << 10) /* see ldlm_cli_cancel_unused */
  58
  59 /* don't run the cancel callback under ldlm_cli_cancel_unused */
  60 #define LDLM_FL_NO_CALLBACK    (1 << 11)
  61
  62 #define LDLM_FL_HAS_INTENT     (1 << 12) /* lock request has intent */
  63 #define LDLM_FL_CANCELING      (1 << 13) /* lock cancel has already been sent */
  64 #define LDLM_FL_LOCAL          (1 << 14) // a local lock (ie, no srv/cli split)
  65 #define LDLM_FL_WARN           (1 << 15) /* see ldlm_cli_cancel_unused */
  66
  67 /* The blocking callback is overloaded to perform two functions.  These flags
  68  * indicate which operation should be performed. */
  69 #define LDLM_CB_BLOCKING    1
  70 #define LDLM_CB_CANCELING   2
  71
  72 #define L2B(c) (1 << c)
  73
  74 /* compatibility matrix */
  75 #define LCK_COMPAT_EX  L2B(LCK_NL)
  76 #define LCK_COMPAT_PW  (LCK_COMPAT_EX | L2B(LCK_CR))
  77 #define LCK_COMPAT_PR  (LCK_COMPAT_PW | L2B(LCK_PR))
  78 #define LCK_COMPAT_CW  (LCK_COMPAT_PW | L2B(LCK_CW))
  79 #define LCK_COMPAT_CR  (LCK_COMPAT_CW | L2B(LCK_PR) | L2B(LCK_PW))
  80 #define LCK_COMPAT_NL  (LCK_COMPAT_CR | L2B(LCK_EX))
  81
  82 static ldlm_mode_t lck_compat_array[] = {
  83         [LCK_EX] LCK_COMPAT_EX,
  84         [LCK_PW] LCK_COMPAT_PW,
  85         [LCK_PR] LCK_COMPAT_PR,
  86         [LCK_CW] LCK_COMPAT_CW,
  87         [LCK_CR] LCK_COMPAT_CR,
  88         [LCK_NL] LCK_COMPAT_NL
  89 };
  90
  91 static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
  92 {
  93        if (exist < LCK_EX || exist > LCK_NL)
  94               LBUG();
  95        if (new < LCK_EX || new > LCK_NL)
  96               LBUG();
  97
  98        return (lck_compat_array[exist] & L2B(new));
  99 }
 100
 101 /*
 102  *
 103  * cluster name spaces
 104  *
 105  */
 106
 107 #define DLM_OST_NAMESPACE 1
 108 #define DLM_MDS_NAMESPACE 2
 109
 110 /* XXX
 111    - do we just separate this by security domains and use a prefix for
 112      multiple namespaces in the same domain?
 113    -
 114 */
 115
 116 struct ldlm_namespace {
 117         char                  *ns_name;
 118         __u32                  ns_client; /* is this a client-side lock tree? */
 119         struct list_head      *ns_hash; /* hash table for ns */
 120         __u32                  ns_refcount; /* count of resources in the hash */
 121         struct list_head       ns_root_list; /* all root resources in ns */
 122         struct lustre_lock     ns_lock; /* protects hash, refcount, list */
 123         struct list_head       ns_list_chain; /* position in global NS list */
 124         /*
 125         struct proc_dir_entry *ns_proc_dir;
 126         */
 127
 128         struct list_head       ns_unused_list; /* all root resources in ns */
 129         unsigned int           ns_nr_unused;
 130         unsigned int           ns_max_unused;
 131
 132         spinlock_t             ns_counter_lock;
 133         __u64                  ns_locks;
 134         __u64                  ns_resources;
 135 };
 136
 137 /*
 138  *
 139  * Resource hash table
 140  *
 141  */
 142
 143 #define RES_HASH_BITS 10
 144 #define RES_HASH_SIZE (1UL << RES_HASH_BITS)
 145 #define RES_HASH_MASK (RES_HASH_SIZE - 1)
 146
 147 struct ldlm_lock;
 148
 149 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
 150                                       struct ldlm_lock_desc *new, void *data,
 151                                       int flag);
 152 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, int flags,
 153                                         void *data);
 154
 155 struct ldlm_lock {
 156         struct portals_handle l_handle; // must be first in the structure
 157         atomic_t              l_refc;
 158         struct ldlm_resource *l_resource;
 159         struct ldlm_lock     *l_parent;
 160         struct list_head      l_children;
 161         struct list_head      l_childof;
 162         struct list_head      l_lru;
 163         struct list_head      l_res_link; // position in one of three res lists
 164         struct list_head      l_export_chain; // per-export chain of locks
 165         struct list_head      l_pending_chain; // locks with callbacks pending
 166         unsigned long         l_callback_timeout;
 167
 168         ldlm_mode_t           l_req_mode;
 169         ldlm_mode_t           l_granted_mode;
 170
 171         ldlm_completion_callback l_completion_ast;
 172         ldlm_blocking_callback   l_blocking_ast;
 173
 174         struct obd_export    *l_export;
 175         struct lustre_handle *l_connh;
 176         __u32                 l_flags;
 177         struct lustre_handle  l_remote_handle;
 178         void                 *l_data;
 179         struct ldlm_extent    l_extent;
 180         __u32                 l_version[RES_VERSION_SIZE];
 181
 182         __u32                 l_readers;
 183         __u32                 l_writers;
 184         __u8                  l_destroyed;
 185
 186         /* If the lock is granted, a process sleeps on this waitq to learn when
 187          * it's no longer in use.  If the lock is not granted, a process sleeps
 188          * on this waitq to learn when it becomes granted. */
 189         wait_queue_head_t     l_waitq;
 190 };
 191
 192 typedef int (*ldlm_res_compat)(struct ldlm_lock *child, struct ldlm_lock *new);
 193 typedef int (*ldlm_res_policy)(struct ldlm_namespace *, struct ldlm_lock **,
 194                                void *req_cookie, ldlm_mode_t mode, int flags,
 195                                void *data);
 196
 197 #define LDLM_PLAIN       10
 198 #define LDLM_EXTENT      11
 199
 200 #define LDLM_MIN_TYPE 10
 201 #define LDLM_MAX_TYPE 11
 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 extern struct obd_ops ldlm_obd_ops;
 237
 238 extern char *ldlm_lockname[];
 239 extern char *ldlm_typename[];
 240 extern char *ldlm_it2str(int it);
 241
 242 #define __LDLM_DEBUG(level, lock, format, a...)                               \
 243 do {                                                                          \
 244         if (lock->l_resource == NULL) {                                       \
 245                 CDEBUG(level, "### " format                                   \
 246                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "\
 247                        "res: \?\? rrc=\?\? type: \?\?\? remote: "             \
 248                        LPX64"\n" , ## a, lock, lock->l_handle.h_cookie,       \
 249                        atomic_read(&lock->l_refc),                            \
 250                        lock->l_readers, lock->l_writers,                      \
 251                        ldlm_lockname[lock->l_granted_mode],                   \
 252                        ldlm_lockname[lock->l_req_mode],                       \
 253                        lock->l_remote_handle.cookie);                         \
 254                 break;                                                        \
 255         }                                                                     \
 256         if (lock->l_resource->lr_type == LDLM_EXTENT) {                       \
 257                 CDEBUG(level, "### " format                                   \
 258                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 259                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64\
 260                        "] remote: "LPX64"\n" , ## a,                          \
 261                        lock->l_resource->lr_namespace->ns_name, lock,         \
 262                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),   \
 263                        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.name[0],                     \
 267                        lock->l_resource->lr_name.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.cookie);                         \
 272                 break;                                                        \
 273         }                                                                     \
 274         {                                                                     \
 275                 CDEBUG(level, "### " format                                   \
 276                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 277                        "res: "LPU64"/"LPU64" rrc: %d type: %s remote: "LPX64  \
 278                        "\n" , ## a,                                           \
 279                        lock->l_resource->lr_namespace->ns_name,               \
 280                        lock, lock->l_handle.h_cookie,                         \
 281                        atomic_read (&lock->l_refc),                           \
 282                        lock->l_readers, lock->l_writers,                      \
 283                        ldlm_lockname[lock->l_granted_mode],                   \
 284                        ldlm_lockname[lock->l_req_mode],                       \
 285                        lock->l_resource->lr_name.name[0],                     \
 286                        lock->l_resource->lr_name.name[1],                     \
 287                        atomic_read(&lock->l_resource->lr_refcount),           \
 288                        ldlm_typename[lock->l_resource->lr_type],              \
 289                        lock->l_remote_handle.cookie);                         \
 290         }                                                                     \
 291 } while (0)
 292
 293 #define LDLM_DEBUG(lock, format, a...) __LDLM_DEBUG(D_DLMTRACE, lock, \
 294                                                     format, ## a)
 295 #define LDLM_ERROR(lock, format, a...) __LDLM_DEBUG(D_ERROR, lock, format, ## a)
 296
 297 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
 298         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
 299
 300 /*
 301  * Iterators.
 302  */
 303
 304 #define LDLM_ITER_CONTINUE 1 /* keep iterating */
 305 #define LDLM_ITER_STOP     0 /* stop iterating */
 306
 307 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
 308 typedef int (*ldlm_res_iterator_t)(struct ldlm_resource *, void *);
 309
 310 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 311                           void *closure);
 312 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 313                            void *closure);
 314 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 315                                ldlm_res_iterator_t iter, void *closure);
 316
 317 int ldlm_replay_locks(struct obd_import *imp);
 318
 319 /* ldlm_extent.c */
 320 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
 321 int ldlm_extent_policy(struct ldlm_namespace *, struct ldlm_lock **, void *,
 322                        ldlm_mode_t, int flags, void *);
 323
 324 /* ldlm_lockd.c */
 325 int ldlm_server_blocking_ast(struct ldlm_lock *, struct ldlm_lock_desc *,
 326                              void *data, int flag);
 327 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 328 int ldlm_handle_enqueue(struct ptlrpc_request *req, ldlm_completion_callback,
 329                         ldlm_blocking_callback);
 330 int ldlm_handle_convert(struct ptlrpc_request *req);
 331 int ldlm_handle_cancel(struct ptlrpc_request *req);
 332 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
 333
 334 /* ldlm_lock.c */
 335 void ldlm_register_intent(ldlm_res_policy arg);
 336 void ldlm_unregister_intent(void);
 337 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
 338 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int flags);
 339 void ldlm_cancel_callback(struct ldlm_lock *);
 340 int ldlm_lock_set_data(struct lustre_handle *, void *data);
 341 void ldlm_lock_remove_from_lru(struct ldlm_lock *);
 342 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *,
 343                                       struct lustre_handle *);
 344
 345 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
 346 {
 347         return __ldlm_handle2lock(h, 0);
 348 }
 349
 350 #define LDLM_LOCK_PUT(lock)                     \
 351 do {                                            \
 352         /*LDLM_DEBUG((lock), "put");*/          \
 353         ldlm_lock_put(lock);                    \
 354 } while (0)
 355
 356 #define LDLM_LOCK_GET(lock)                     \
 357 ({                                              \
 358         ldlm_lock_get(lock);                    \
 359         /*LDLM_DEBUG((lock), "get");*/          \
 360         lock;                                   \
 361 })
 362
 363 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 364 void ldlm_lock_put(struct ldlm_lock *lock);
 365 void ldlm_lock_destroy(struct ldlm_lock *lock);
 366 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
 367 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 368 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
 369 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 370 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode);
 371 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen);
 372 int ldlm_lock_match(struct ldlm_namespace *ns, int flags, struct ldlm_res_id *,
 373                     __u32 type, void *cookie, int cookielen, ldlm_mode_t mode,
 374                     struct lustre_handle *);
 375 struct ldlm_lock *
 376 ldlm_lock_create(struct ldlm_namespace *ns,
 377                  struct lustre_handle *parent_lock_handle, struct ldlm_res_id,
 378                  __u32 type, ldlm_mode_t, ldlm_blocking_callback,
 379                  void *data);
 380 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *, struct ldlm_lock **,
 381                                void *cookie, int cookie_len, int *flags,
 382                                ldlm_completion_callback completion);
 383 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 384                                         int *flags);
 385 void ldlm_lock_cancel(struct ldlm_lock *lock);
 386 void ldlm_cancel_locks_for_export(struct obd_export *export);
 387 int ldlm_run_ast_work(struct list_head *rpc_list);
 388 void ldlm_reprocess_all(struct ldlm_resource *res);
 389 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns);
 390 void ldlm_lock_dump(int level, struct ldlm_lock *lock);
 391 void ldlm_lock_dump_handle(int level, struct lustre_handle *);
 392
 393 /* ldlm_test.c */
 394 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
 395 int ldlm_regression_start(struct obd_device *obddev,
 396                           struct lustre_handle *connh,
 397                           unsigned int threads, unsigned int max_locks_in,
 398                           unsigned int num_resources_in,
 399                           unsigned int num_extents_in);
 400 int ldlm_regression_stop(void);
 401
 402
 403 /* resource.c */
 404 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
 405 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
 406 int ldlm_namespace_free(struct ldlm_namespace *ns);
 407 int ldlm_proc_setup(struct obd_device *obd);
 408 void ldlm_proc_cleanup(struct obd_device *obd);
 409
 410 /* resource.c - internal */
 411 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 412                                         struct ldlm_resource *parent,
 413                                         struct ldlm_res_id, __u32 type,
 414                                         int create);
 415 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
 416 int ldlm_resource_putref(struct ldlm_resource *res);
 417 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 418                             struct ldlm_lock *lock);
 419 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 420 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
 421 void ldlm_dump_all_namespaces(void);
 422 void ldlm_namespace_dump(struct ldlm_namespace *);
 423 void ldlm_resource_dump(struct ldlm_resource *);
 424 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
 425                               struct ldlm_res_id);
 426
 427 /* ldlm_request.c */
 428 int ldlm_expired_completion_wait(void *data);
 429 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 430 int ldlm_cli_enqueue(struct lustre_handle *conn,
 431                      struct ptlrpc_request *req,
 432                      struct ldlm_namespace *ns,
 433                      struct lustre_handle *parent_lock_handle,
 434                      struct ldlm_res_id,
 435                      __u32 type,
 436                      void *cookie, int cookielen,
 437                      ldlm_mode_t mode,
 438                      int *flags,
 439                      ldlm_completion_callback completion,
 440                      ldlm_blocking_callback callback,
 441                      void *data,
 442                      struct lustre_handle *lockh);
 443 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 444                           struct ptlrpc_request *req,
 445                           struct ldlm_namespace *ns,
 446                           struct lustre_handle *parent_lock_handle,
 447                           struct ldlm_res_id,
 448                           __u32 type,
 449                           void *cookie, int cookielen,
 450                           ldlm_mode_t mode,
 451                           int *flags,
 452                           ldlm_completion_callback completion,
 453                           ldlm_blocking_callback callback,
 454                           void *data,
 455                           struct lustre_handle *lockh);
 456 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
 457                     void *data, __u32 data_len);
 458 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
 459 int ldlm_cli_cancel(struct lustre_handle *lockh);
 460 int ldlm_cli_cancel_unused(struct ldlm_namespace *, struct ldlm_res_id *,
 461                            int flags, void *opaque);
 462
 463 /* mds/handler.c */
 464 /* This has to be here because recurisve inclusion sucks. */
 465 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 466                      void *data, int flag);
 467
 468
 469 /* ioctls for trying requests */
 470 #define IOC_LDLM_TYPE                   'f'
 471 #define IOC_LDLM_MIN_NR                 40
 472
 473 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
 474 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
 475 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
 476 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
 477 #define IOC_LDLM_MAX_NR                 43
 478
 479 #endif