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         struct portals_handle l_handle; // must be first in the structure
 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_namespace *, struct ldlm_lock *,
 187                                void *req_cookie, ldlm_mode_t mode, int flags,
 188                                void *data);
 189
 190 #define LDLM_PLAIN       10
 191 #define LDLM_EXTENT      11
 192
 193 #define LDLM_MIN_TYPE 10
 194 #define LDLM_MAX_TYPE 11
 195
 196 extern ldlm_res_compat ldlm_res_compat_table [];
 197 extern ldlm_res_policy ldlm_res_policy_table [];
 198
 199 struct ldlm_resource {
 200         struct ldlm_namespace *lr_namespace;
 201         struct list_head       lr_hash;
 202         struct ldlm_resource  *lr_parent;   /* 0 for a root resource */
 203         struct list_head       lr_children; /* list head for child resources */
 204         struct list_head       lr_childof;  /* part of ns_root_list if root res,
 205                                              * part of lr_children if child */
 206
 207         struct list_head       lr_granted;
 208         struct list_head       lr_converting;
 209         struct list_head       lr_waiting;
 210         ldlm_mode_t            lr_most_restr;
 211         __u32                  lr_type; /* LDLM_PLAIN or LDLM_EXTENT */
 212         struct ldlm_resource  *lr_root;
 213         __u64                  lr_name[RES_NAME_SIZE];
 214         __u32                  lr_version[RES_VERSION_SIZE];
 215         atomic_t               lr_refcount;
 216         void                  *lr_tmp;
 217 };
 218
 219 struct ldlm_ast_work {
 220         struct ldlm_lock *w_lock;
 221         int               w_blocking;
 222         struct ldlm_lock_desc w_desc;
 223         struct list_head   w_list;
 224         int w_flags;
 225         void *w_data;
 226         int w_datalen;
 227 };
 228
 229 /* Per-export ldlm state. */
 230 struct ldlm_export_data {
 231         struct list_head        led_held_locks;
 232         struct obd_import       led_import;
 233 };
 234
 235 static inline struct ldlm_extent *ldlm_res2extent(struct ldlm_resource *res)
 236 {
 237         return (struct ldlm_extent *)(res->lr_name);
 238 }
 239
 240 extern struct obd_ops ldlm_obd_ops;
 241
 242 extern char *ldlm_lockname[];
 243 extern char *ldlm_typename[];
 244 extern char *ldlm_it2str(int it);
 245
 246 #define LDLM_DEBUG(lock, format, a...)                                        \
 247 do {                                                                          \
 248         if (lock->l_resource == NULL) {                                       \
 249                 CDEBUG(D_DLMTRACE, "### " format                              \
 250                        " ns: \?\? lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "\
 251                        "res: \?\? rrc=\?\? type: \?\?\? remote: "LPX64")\n"   \
 252                        , ## a, lock, lock->l_handle.h_cookie,                 \
 253                        atomic_read(&lock->l_refc),                            \
 254                        lock->l_readers, lock->l_writers,                      \
 255                        ldlm_lockname[lock->l_granted_mode],                   \
 256                        ldlm_lockname[lock->l_req_mode],                       \
 257                        lock->l_remote_handle.cookie);                         \
 258                 break;                                                        \
 259         }                                                                     \
 260         if (lock->l_resource->lr_type == LDLM_EXTENT) {                       \
 261                 CDEBUG(D_DLMTRACE, "### " format                              \
 262                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 263                        "res: "LPU64"/"LPU64" rrc: %d type: %s ["LPU64"->"LPU64\
 264                        "] remote: "LPX64"\n" , ## a,                          \
 265                        lock->l_resource->lr_namespace->ns_name, lock,         \
 266                        lock->l_handle.h_cookie, atomic_read(&lock->l_refc),   \
 267                        lock->l_readers, lock->l_writers,                      \
 268                        ldlm_lockname[lock->l_granted_mode],                   \
 269                        ldlm_lockname[lock->l_req_mode],                       \
 270                        lock->l_resource->lr_name[0],                          \
 271                        lock->l_resource->lr_name[1],                          \
 272                        atomic_read(&lock->l_resource->lr_refcount),           \
 273                        ldlm_typename[lock->l_resource->lr_type],              \
 274                        lock->l_extent.start, lock->l_extent.end,              \
 275                        lock->l_remote_handle.cookie);                         \
 276                 break;                                                        \
 277         }                                                                     \
 278         {                                                                     \
 279                 CDEBUG(D_DLMTRACE, "### " format                              \
 280                        " ns: %s lock: %p/"LPX64" lrc: %d/%d,%d mode: %s/%s "  \
 281                        "res: "LPU64"/"LPU64" rrc: %d type: %s remote: "LPX64  \
 282                        "\n" , ## a, lock->l_resource->lr_namespace->ns_name,  \
 283                        lock, lock->l_handle.h_cookie,                         \
 284                        atomic_read (&lock->l_refc),                           \
 285                        lock->l_readers, lock->l_writers,                      \
 286                        ldlm_lockname[lock->l_granted_mode],                   \
 287                        ldlm_lockname[lock->l_req_mode],                       \
 288                        lock->l_resource->lr_name[0],                          \
 289                        lock->l_resource->lr_name[1],                          \
 290                        atomic_read(&lock->l_resource->lr_refcount),           \
 291                        ldlm_typename[lock->l_resource->lr_type],              \
 292                        lock->l_remote_handle.cookie);                         \
 293         }                                                                     \
 294 } while (0)
 295
 296 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
 297         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
 298
 299 /*
 300  * Iterators.
 301  */
 302
 303 #define LDLM_ITER_CONTINUE 1 /* keep iterating */
 304 #define LDLM_ITER_STOP     0 /* stop iterating */
 305
 306 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
 307 typedef int (*ldlm_res_iterator_t)(struct ldlm_resource *, void *);
 308
 309 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 310                           void *closure);
 311 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 312                            void *closure);
 313 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 314                                ldlm_res_iterator_t iter, void *closure);
 315
 316 int ldlm_replay_locks(struct obd_import *imp);
 317
 318 /* ldlm_extent.c */
 319 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
 320 int ldlm_extent_policy(struct ldlm_namespace *, struct ldlm_lock *, void *,
 321                        ldlm_mode_t, int flags, void *);
 322
 323 /* ldlm_lockd.c */
 324 int ldlm_handle_enqueue(struct ptlrpc_request *req);
 325 int ldlm_handle_convert(struct ptlrpc_request *req);
 326 int ldlm_handle_cancel(struct ptlrpc_request *req);
 327 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
 328
 329 /* ldlm_lock.c */
 330 void ldlm_register_intent(ldlm_res_policy arg);
 331 void ldlm_unregister_intent(void);
 332 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
 333 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int flags);
 334 void ldlm_cancel_callback(struct ldlm_lock *);
 335 int ldlm_lock_set_data(struct lustre_handle *, void *data, int datalen);
 336 void ldlm_lock_remove_from_lru(struct ldlm_lock *);
 337
 338 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
 339 {
 340         return __ldlm_handle2lock(h, 0);
 341 }
 342
 343 #define LDLM_LOCK_PUT(lock)                     \
 344 do {                                            \
 345         /*LDLM_DEBUG(lock, "put");*/            \
 346         ldlm_lock_put(lock);                    \
 347 } while (0)
 348
 349 #define LDLM_LOCK_GET(lock)                     \
 350 ({                                              \
 351         ldlm_lock_get(lock);                    \
 352         /*LDLM_DEBUG(lock, "get");*/            \
 353         lock;                                   \
 354 })
 355
 356 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 357 void ldlm_lock_put(struct ldlm_lock *lock);
 358 void ldlm_lock_destroy(struct ldlm_lock *lock);
 359 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
 360 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 361 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
 362 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 363 void ldlm_grant_lock(struct ldlm_lock *lock);
 364 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
 365                     void *cookie, int cookielen, ldlm_mode_t mode,
 366                     struct lustre_handle *lockh);
 367 struct ldlm_lock *
 368 ldlm_lock_create(struct ldlm_namespace *ns,
 369                  struct lustre_handle *parent_lock_handle,
 370                  __u64 *res_id, __u32 type, ldlm_mode_t mode, void *data,
 371                  __u32 data_len);
 372 ldlm_error_t ldlm_lock_enqueue(struct ldlm_namespace *, struct ldlm_lock *,
 373                                void *cookie, int cookie_len, int *flags,
 374                                ldlm_completion_callback completion,
 375                                ldlm_blocking_callback blocking);
 376 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 377                                         int *flags);
 378 void ldlm_lock_cancel(struct ldlm_lock *lock);
 379 void ldlm_cancel_locks_for_export(struct obd_export *export);
 380 int ldlm_run_ast_work(struct list_head *rpc_list);
 381 void ldlm_reprocess_all(struct ldlm_resource *res);
 382 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns);
 383 void ldlm_lock_dump(int level, struct ldlm_lock *lock);
 384 void ldlm_lock_dump_handle(int level, struct lustre_handle *);
 385
 386 /* ldlm_test.c */
 387 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
 388 int ldlm_regression_start(struct obd_device *obddev,
 389                           struct lustre_handle *connh,
 390                           unsigned int threads, unsigned int max_locks_in,
 391                           unsigned int num_resources_in,
 392                           unsigned int num_extents_in);
 393 int ldlm_regression_stop(void);
 394
 395
 396 /* resource.c */
 397 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
 398 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
 399 int ldlm_namespace_free(struct ldlm_namespace *ns);
 400 int ldlm_proc_setup(struct obd_device *obd);
 401 void ldlm_proc_cleanup(struct obd_device *obd);
 402
 403 /* resource.c - internal */
 404 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 405                                         struct ldlm_resource *parent,
 406                                         __u64 *name, __u32 type, int create);
 407 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
 408 int ldlm_resource_putref(struct ldlm_resource *res);
 409 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 410                             struct ldlm_lock *lock);
 411 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 412 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
 413 void ldlm_dump_all_namespaces(void);
 414 void ldlm_namespace_dump(struct ldlm_namespace *);
 415 void ldlm_resource_dump(struct ldlm_resource *);
 416 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
 417                               __u64 new_resid[3]);
 418
 419 /* ldlm_request.c */
 420 int ldlm_expired_completion_wait(void *data);
 421 int ldlm_completion_ast(struct ldlm_lock *lock, int flags);
 422 int ldlm_cli_enqueue(struct lustre_handle *conn,
 423                      struct ptlrpc_request *req,
 424                      struct ldlm_namespace *ns,
 425                      struct lustre_handle *parent_lock_handle,
 426                      __u64 *res_id,
 427                      __u32 type,
 428                      void *cookie, int cookielen,
 429                      ldlm_mode_t mode,
 430                      int *flags,
 431                      ldlm_completion_callback completion,
 432                      ldlm_blocking_callback callback,
 433                      void *data,
 434                      __u32 data_len,
 435                      struct lustre_handle *lockh);
 436 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 437                           struct ptlrpc_request *req,
 438                           struct ldlm_namespace *ns,
 439                           struct lustre_handle *parent_lock_handle,
 440                           __u64 *res_id,
 441                           __u32 type,
 442                           void *cookie, int cookielen,
 443                           ldlm_mode_t mode,
 444                           int *flags,
 445                           ldlm_completion_callback completion,
 446                           ldlm_blocking_callback callback,
 447                           void *data,
 448                           __u32 data_len,
 449                           struct lustre_handle *lockh);
 450 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
 451                     void *data, __u32 data_len);
 452 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
 453 int ldlm_cli_cancel(struct lustre_handle *lockh);
 454 int ldlm_cli_cancel_unused(struct ldlm_namespace *, __u64 *, int flags);
 455 int ldlm_cancel_lru(struct ldlm_namespace *ns);
 456
 457 /* mds/handler.c */
 458 /* This has to be here because recurisve inclusion sucks. */
 459 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 460                      void *data, __u32 data_len, int flag);
 461
 462 #endif /* __KERNEL__ */
 463
 464 /* ioctls for trying requests */
 465 #define IOC_LDLM_TYPE                   'f'
 466 #define IOC_LDLM_MIN_NR                 40
 467
 468 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
 469 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
 470 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
 471 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
 472 #define IOC_LDLM_MAX_NR                 43
 473
 474 #endif