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 #define LDLM_FL_MATCH_DATA     (1 << 16) /* see ldlm_lock_match */
  67
  68 /* The blocking callback is overloaded to perform two functions.  These flags
  69  * indicate which operation should be performed. */
  70 #define LDLM_CB_BLOCKING    1
  71 #define LDLM_CB_CANCELING   2
  72
  73 #define L2B(c) (1 << c)
  74
  75 /* compatibility matrix */
  76 #define LCK_COMPAT_EX  L2B(LCK_NL)
  77 #define LCK_COMPAT_PW  (LCK_COMPAT_EX | L2B(LCK_CR))
  78 #define LCK_COMPAT_PR  (LCK_COMPAT_PW | L2B(LCK_PR))
  79 #define LCK_COMPAT_CW  (LCK_COMPAT_PW | L2B(LCK_CW))
  80 #define LCK_COMPAT_CR  (LCK_COMPAT_CW | L2B(LCK_PR) | L2B(LCK_PW))
  81 #define LCK_COMPAT_NL  (LCK_COMPAT_CR | L2B(LCK_EX))
  82
  83 static ldlm_mode_t lck_compat_array[] = {
  84         [LCK_EX] LCK_COMPAT_EX,
  85         [LCK_PW] LCK_COMPAT_PW,
  86         [LCK_PR] LCK_COMPAT_PR,
  87         [LCK_CW] LCK_COMPAT_CW,
  88         [LCK_CR] LCK_COMPAT_CR,
  89         [LCK_NL] LCK_COMPAT_NL
  90 };
  91
  92 static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
  93 {
  94        if (exist < LCK_EX || exist > LCK_NL)
  95               LBUG();
  96        if (new < LCK_EX || new > LCK_NL)
  97               LBUG();
  98
  99        return (lck_compat_array[exist] & L2B(new));
 100 }
 101
 102 /*
 103  *
 104  * cluster name spaces
 105  *
 106  */
 107
 108 #define DLM_OST_NAMESPACE 1
 109 #define DLM_MDS_NAMESPACE 2
 110
 111 /* XXX
 112    - do we just separate this by security domains and use a prefix for
 113      multiple namespaces in the same domain?
 114    -
 115 */
 116
 117 struct ldlm_namespace {
 118         char                  *ns_name;
 119         __u32                  ns_client; /* is this a client-side lock tree? */
 120         struct list_head      *ns_hash; /* hash table for ns */
 121         __u32                  ns_refcount; /* count of resources in the hash */
 122         struct list_head       ns_root_list; /* all root resources in ns */
 123         struct lustre_lock     ns_lock; /* protects hash, refcount, list */
 124         struct list_head       ns_list_chain; /* position in global NS list */
 125         /*
 126         struct proc_dir_entry *ns_proc_dir;
 127         */
 128
 129         struct list_head       ns_unused_list; /* all root resources in ns */
 130         unsigned int           ns_nr_unused;
 131         unsigned int           ns_max_unused;
 132
 133         spinlock_t             ns_counter_lock;
 134         __u64                  ns_locks;
 135         __u64                  ns_resources;
 136 };
 137
 138 /*
 139  *
 140  * Resource hash table
 141  *
 142  */
 143
 144 #define RES_HASH_BITS 10
 145 #define RES_HASH_SIZE (1UL << RES_HASH_BITS)
 146 #define RES_HASH_MASK (RES_HASH_SIZE - 1)
 147
 148 struct ldlm_lock;
 149
 150 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
 151                                       struct ldlm_lock_desc *new, void *data,
 152                                       int flag);
 153 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, int flags,
 154                                         void *data);
 155
 156 struct ldlm_lock {
 157         struct portals_handle l_handle; // must be first in the structure
 158         atomic_t              l_refc;
 159         struct ldlm_resource *l_resource;
 160         struct ldlm_lock     *l_parent;
 161         struct list_head      l_children;
 162         struct list_head      l_childof;
 163         struct list_head      l_lru;
 164         struct list_head      l_res_link; // position in one of three res lists
 165         struct list_head      l_export_chain; // per-export chain of locks
 166         struct list_head      l_pending_chain; // locks with callbacks pending
 167         unsigned long         l_callback_timeout;
 168
 169         ldlm_mode_t           l_req_mode;
 170         ldlm_mode_t           l_granted_mode;
 171
 172         ldlm_completion_callback l_completion_ast;
 173         ldlm_blocking_callback   l_blocking_ast;
 174
 175         struct obd_export    *l_export;
 176         struct lustre_handle *l_connh;
 177         __u32                 l_flags;
 178         struct lustre_handle  l_remote_handle;
 179         void                 *l_data;
 180         struct ldlm_extent    l_extent;
 181         __u32                 l_version[RES_VERSION_SIZE];
 182
 183         __u32                 l_readers;
 184         __u32                 l_writers;
 185         __u8                  l_destroyed;
 186
 187         /* If the lock is granted, a process sleeps on this waitq to learn when
 188          * it's no longer in use.  If the lock is not granted, a process sleeps
 189          * on this waitq to learn when it becomes granted. */
 190         wait_queue_head_t     l_waitq;
 191         struct timeval        l_enqueued_time;
 192 };
 193
 194 typedef int (*ldlm_res_compat)(struct ldlm_lock *child, struct ldlm_lock *new);
 195 typedef int (*ldlm_res_policy)(struct ldlm_namespace *, struct ldlm_lock **,
 196                                void *req_cookie, ldlm_mode_t mode, int flags,
 197                                void *data);
 198
 199 #define LDLM_PLAIN       10
 200 #define LDLM_EXTENT      11
 201
 202 #define LDLM_MIN_TYPE 10
 203 #define LDLM_MAX_TYPE 11
 204
 205 struct ldlm_resource {
 206         struct ldlm_namespace *lr_namespace;
 207         struct list_head       lr_hash;
 208         struct ldlm_resource  *lr_parent;   /* 0 for a root resource */
 209         struct list_head       lr_children; /* list head for child resources */
 210         struct list_head       lr_childof;  /* part of ns_root_list if root res,
 211                                              * part of lr_children if child */
 212
 213         struct list_head       lr_granted;
 214         struct list_head       lr_converting;
 215         struct list_head       lr_waiting;
 216         ldlm_mode_t            lr_most_restr;
 217         __u32                  lr_type; /* LDLM_PLAIN or LDLM_EXTENT */
 218         struct ldlm_resource  *lr_root;
 219         struct ldlm_res_id     lr_name;
 220         __u32                  lr_version[RES_VERSION_SIZE];
 221         atomic_t               lr_refcount;
 222
 223         /* lr_tmp holds a list head temporarily, during the building of a work
 224          * queue.  see ldlm_add_ast_work_item and ldlm_run_ast_work */
 225         void                  *lr_tmp;
 226 };
 227
 228 struct ldlm_ast_work {
 229         struct ldlm_lock *w_lock;
 230         int               w_blocking;
 231         struct ldlm_lock_desc w_desc;
 232         struct list_head   w_list;
 233         int w_flags;
 234         void *w_data;
 235         int w_datalen;
 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(level, lock, format, a...)                               \
 245 do {                                                                          \
 246         if (lock->l_resource == NULL) {                                       \
 247                 CDEBUG(level, "### " format                                   \
 248                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "\
 249                        "res: \?\? rrc=\?\? type: \?\?\? remote: "             \
 250                        LPX64"\n" , ## a, lock, lock->l_handle.h_cookie,       \
 251                        atomic_read(&lock->l_refc),                            \
 252                        lock->l_readers, lock->l_writers,                      \
 253                        ldlm_lockname[lock->l_granted_mode],                   \
 254                        ldlm_lockname[lock->l_req_mode],                       \
 255                        lock->l_remote_handle.cookie);                         \
 256                 break;                                                        \
 257         }                                                                     \
 258         if (lock->l_resource->lr_type == LDLM_EXTENT) {                       \
 259                 CDEBUG(level, "### " format                                   \
 260                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 261                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64\
 262                        "] remote: "LPX64"\n" , ## a,                          \
 263                        lock->l_resource->lr_namespace->ns_name, lock,         \
 264                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),   \
 265                        lock->l_readers, lock->l_writers,                      \
 266                        ldlm_lockname[lock->l_granted_mode],                   \
 267                        ldlm_lockname[lock->l_req_mode],                       \
 268                        lock->l_resource->lr_name.name[0],                     \
 269                        lock->l_resource->lr_name.name[1],                     \
 270                        atomic_read(&lock->l_resource->lr_refcount),           \
 271                        ldlm_typename[lock->l_resource->lr_type],              \
 272                        lock->l_extent.start, lock->l_extent.end,              \
 273                        lock->l_remote_handle.cookie);                         \
 274                 break;                                                        \
 275         }                                                                     \
 276         {                                                                     \
 277                 CDEBUG(level, "### " format                                   \
 278                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 279                        "res: "LPU64"/"LPU64" rrc: %d type: %s remote: "LPX64  \
 280                        "\n" , ## a,                                           \
 281                        lock->l_resource->lr_namespace->ns_name,               \
 282                        lock, lock->l_handle.h_cookie,                         \
 283                        atomic_read (&lock->l_refc),                           \
 284                        lock->l_readers, lock->l_writers,                      \
 285                        ldlm_lockname[lock->l_granted_mode],                   \
 286                        ldlm_lockname[lock->l_req_mode],                       \
 287                        lock->l_resource->lr_name.name[0],                     \
 288                        lock->l_resource->lr_name.name[1],                     \
 289                        atomic_read(&lock->l_resource->lr_refcount),           \
 290                        ldlm_typename[lock->l_resource->lr_type],              \
 291                        lock->l_remote_handle.cookie);                         \
 292         }                                                                     \
 293 } while (0)
 294
 295 #define LDLM_DEBUG(lock, format, a...) __LDLM_DEBUG(D_DLMTRACE, lock, \
 296                                                     format, ## a)
 297 #define LDLM_ERROR(lock, format, a...) __LDLM_DEBUG(D_ERROR, lock, format, ## a)
 298
 299 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
 300         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
 301
 302 /*
 303  * Iterators.
 304  */
 305
 306 #define LDLM_ITER_CONTINUE 1 /* keep iterating */
 307 #define LDLM_ITER_STOP     0 /* stop iterating */
 308
 309 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
 310 typedef int (*ldlm_res_iterator_t)(struct ldlm_resource *, void *);
 311
 312 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 313                           void *closure);
 314 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 315                            void *closure);
 316 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 317                                ldlm_res_iterator_t iter, void *closure);
 318
 319 int ldlm_replay_locks(struct obd_import *imp);
 320 void ldlm_change_cbdata(struct ldlm_namespace *, struct ldlm_res_id *,
 321                         ldlm_iterator_t iter, void *data);
 322
 323 /* ldlm_extent.c */
 324 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
 325 int ldlm_extent_policy(struct ldlm_namespace *, struct ldlm_lock **, void *,
 326                        ldlm_mode_t, int flags, void *);
 327
 328 /* ldlm_lockd.c */
 329 int ldlm_server_blocking_ast(struct ldlm_lock *, struct ldlm_lock_desc *,
 330                              void *data, int flag);
 331 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 332 int ldlm_handle_enqueue(struct ptlrpc_request *req, ldlm_completion_callback,
 333                         ldlm_blocking_callback);
 334 int ldlm_handle_convert(struct ptlrpc_request *req);
 335 int ldlm_handle_cancel(struct ptlrpc_request *req);
 336 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
 337
 338 /* ldlm_lock.c */
 339 void ldlm_register_intent(ldlm_res_policy arg);
 340 void ldlm_unregister_intent(void);
 341 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
 342 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int flags);
 343 void ldlm_cancel_callback(struct ldlm_lock *);
 344 int ldlm_lock_set_data(struct lustre_handle *, void *data);
 345 void ldlm_lock_remove_from_lru(struct ldlm_lock *);
 346 struct ldlm_lock *ldlm_handle2lock_ns(struct ldlm_namespace *,
 347                                       struct lustre_handle *);
 348
 349 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
 350 {
 351         return __ldlm_handle2lock(h, 0);
 352 }
 353
 354 #define LDLM_LOCK_PUT(lock)                     \
 355 do {                                            \
 356         /*LDLM_DEBUG((lock), "put");*/          \
 357         ldlm_lock_put(lock);                    \
 358 } while (0)
 359
 360 #define LDLM_LOCK_GET(lock)                     \
 361 ({                                              \
 362         ldlm_lock_get(lock);                    \
 363         /*LDLM_DEBUG((lock), "get");*/          \
 364         lock;                                   \
 365 })
 366
 367 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 368 void ldlm_lock_put(struct ldlm_lock *lock);
 369 void ldlm_lock_destroy(struct ldlm_lock *lock);
 370 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
 371 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 372 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
 373 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 374 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode);
 375 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen);
 376 int ldlm_lock_match(struct ldlm_namespace *ns, int flags, struct ldlm_res_id *,
 377                     __u32 type, void *cookie, int cookielen, ldlm_mode_t mode,
 378                     void *data, struct lustre_handle *);
 379 struct ldlm_lock *
 380 ldlm_lock_create(struct ldlm_namespace *ns,
 381                  struct lustre_handle *parent_lock_handle, struct ldlm_res_id,
 382                  __u32 type, ldlm_mode_t, ldlm_blocking_callback,
 383                  void *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 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 388                                         int *flags);
 389 void ldlm_lock_cancel(struct ldlm_lock *lock);
 390 void ldlm_cancel_locks_for_export(struct obd_export *export);
 391 int ldlm_run_ast_work(struct list_head *rpc_list);
 392 void ldlm_reprocess_all(struct ldlm_resource *res);
 393 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns);
 394 void ldlm_lock_dump(int level, struct ldlm_lock *lock);
 395 void ldlm_lock_dump_handle(int level, struct lustre_handle *);
 396
 397 /* ldlm_test.c */
 398 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
 399 int ldlm_regression_start(struct obd_device *obddev,
 400                           struct lustre_handle *connh,
 401                           unsigned int threads, unsigned int max_locks_in,
 402                           unsigned int num_resources_in,
 403                           unsigned int num_extents_in);
 404 int ldlm_regression_stop(void);
 405
 406
 407 /* resource.c */
 408 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
 409 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
 410 int ldlm_namespace_free(struct ldlm_namespace *ns);
 411 int ldlm_proc_setup(struct obd_device *obd);
 412 void ldlm_proc_cleanup(struct obd_device *obd);
 413
 414 /* resource.c - internal */
 415 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 416                                         struct ldlm_resource *parent,
 417                                         struct ldlm_res_id, __u32 type,
 418                                         int create);
 419 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
 420 int ldlm_resource_putref(struct ldlm_resource *res);
 421 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 422                             struct ldlm_lock *lock);
 423 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 424 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
 425 void ldlm_dump_all_namespaces(void);
 426 void ldlm_namespace_dump(struct ldlm_namespace *);
 427 void ldlm_resource_dump(struct ldlm_resource *);
 428 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
 429                               struct ldlm_res_id);
 430
 431 /* ldlm_request.c */
 432 int ldlm_expired_completion_wait(void *data);
 433 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 434 int ldlm_cli_enqueue(struct lustre_handle *conn,
 435                      struct ptlrpc_request *req,
 436                      struct ldlm_namespace *ns,
 437                      struct lustre_handle *parent_lock_handle,
 438                      struct ldlm_res_id,
 439                      __u32 type,
 440                      void *cookie, int cookielen,
 441                      ldlm_mode_t mode,
 442                      int *flags,
 443                      ldlm_completion_callback completion,
 444                      ldlm_blocking_callback callback,
 445                      void *data,
 446                      struct lustre_handle *lockh);
 447 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
 448                     void *data, __u32 data_len);
 449 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
 450 int ldlm_cli_cancel(struct lustre_handle *lockh);
 451 int ldlm_cli_cancel_unused(struct ldlm_namespace *, struct ldlm_res_id *,
 452                            int flags, void *opaque);
 453
 454 /* mds/handler.c */
 455 /* This has to be here because recurisve inclusion sucks. */
 456 int intent_disposition(struct ldlm_reply *rep, int flag);
 457 void intent_set_disposition(struct ldlm_reply *rep, int flag);
 458 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 459                      void *data, int flag);
 460
 461
 462 /* ioctls for trying requests */
 463 #define IOC_LDLM_TYPE                   'f'
 464 #define IOC_LDLM_MIN_NR                 40
 465
 466 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
 467 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
 468 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
 469 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
 470 #define IOC_LDLM_MAX_NR                 43
 471
 472 #endif