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
  18 struct obd_ops;
  19 struct obd_device;
  20
  21 #define OBD_LDLM_DEVICENAME  "ldlm"
  22
  23 typedef enum {
  24         ELDLM_OK = 0,
  25
  26         ELDLM_LOCK_CHANGED = 300,
  27         ELDLM_LOCK_ABORTED = 301,
  28         ELDLM_LOCK_REPLACED = 302,
  29
  30         ELDLM_NAMESPACE_EXISTS = 400,
  31         ELDLM_BAD_NAMESPACE    = 401
  32 } ldlm_error_t;
  33
  34 #define LDLM_NAMESPACE_SERVER 0
  35 #define LDLM_NAMESPACE_CLIENT 1
  36
  37 #define LDLM_FL_LOCK_CHANGED   (1 << 0) /* extent, mode, or resource changed */
  38
  39 /* If the server returns one of these flags, then the lock was put on that list.
  40  * If the client sends one of these flags (during recovery ONLY!), it wants the
  41  * lock added to the specified list, no questions asked. -p */
  42 #define LDLM_FL_BLOCK_GRANTED  (1 << 1)
  43 #define LDLM_FL_BLOCK_CONV     (1 << 2)
  44 #define LDLM_FL_BLOCK_WAIT     (1 << 3)
  45
  46 #define LDLM_FL_CBPENDING      (1 << 4) // this lock is being destroyed
  47 #define LDLM_FL_AST_SENT       (1 << 5) // blocking or cancel packet was sent
  48 #define LDLM_FL_WAIT_NOREPROC  (1 << 6)// not a real lock flag,not saved in lock
  49 #define LDLM_FL_CANCEL         (1 << 7) // cancellation callback already run
  50
  51 /* Lock is being replayed.  This could probably be implied by the fact that one
  52  * of BLOCK_{GRANTED,CONV,WAIT} is set, but that is pretty dangerous. */
  53 #define LDLM_FL_REPLAY         (1 << 8)
  54
  55 #define LDLM_FL_INTENT_ONLY    (1 << 9) /* don't grant lock, just do intent */
  56 #define LDLM_FL_LOCAL_ONLY     (1 << 10) /* see ldlm_cli_cancel_unused */
  57 #define LDLM_FL_NO_CALLBACK    (1 << 11) /* see ldlm_cli_cancel_unused */
  58 #define LDLM_FL_HAS_INTENT     (1 << 12) /* lock request has intent */
  59 #define LDLM_FL_CANCELING      (1 << 13) /* lock cancel has already been sent */
  60 #define LDLM_FL_LOCAL          (1 << 14) // a local lock (ie, no srv/cli split)
  61
  62 /* The blocking callback is overloaded to perform two functions.  These flags
  63  * indicate which operation should be performed. */
  64 #define LDLM_CB_BLOCKING    1
  65 #define LDLM_CB_CANCELING   2
  66
  67 #define L2B(c) (1 << c)
  68
  69 /* compatibility matrix */
  70 #define LCK_COMPAT_EX  L2B(LCK_NL)
  71 #define LCK_COMPAT_PW  (LCK_COMPAT_EX | L2B(LCK_CR))
  72 #define LCK_COMPAT_PR  (LCK_COMPAT_PW | L2B(LCK_PR))
  73 #define LCK_COMPAT_CW  (LCK_COMPAT_PW | L2B(LCK_CW))
  74 #define LCK_COMPAT_CR  (LCK_COMPAT_CW | L2B(LCK_PR) | L2B(LCK_PW))
  75 #define LCK_COMPAT_NL  (LCK_COMPAT_CR | L2B(LCK_EX))
  76
  77 static ldlm_mode_t lck_compat_array[] = {
  78         [LCK_EX] LCK_COMPAT_EX,
  79         [LCK_PW] LCK_COMPAT_PW,
  80         [LCK_PR] LCK_COMPAT_PR,
  81         [LCK_CW] LCK_COMPAT_CW,
  82         [LCK_CR] LCK_COMPAT_CR,
  83         [LCK_NL] LCK_COMPAT_NL
  84 };
  85
  86 static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
  87 {
  88        if (exist < LCK_EX || exist > LCK_NL)
  89               LBUG();
  90        if (new < LCK_EX || new > LCK_NL)
  91               LBUG();
  92
  93        return (lck_compat_array[exist] & L2B(new));
  94 }
  95
  96 /*
  97  *
  98  * cluster name spaces
  99  *
 100  */
 101
 102 #define DLM_OST_NAMESPACE 1
 103 #define DLM_MDS_NAMESPACE 2
 104
 105 /* XXX
 106    - do we just separate this by security domains and use a prefix for
 107      multiple namespaces in the same domain?
 108    -
 109 */
 110
 111 struct ldlm_namespace {
 112         char                  *ns_name;
 113         __u32                  ns_client; /* is this a client-side lock tree? */
 114         struct list_head      *ns_hash; /* hash table for ns */
 115         __u32                  ns_refcount; /* count of resources in the hash */
 116         struct list_head       ns_root_list; /* all root resources in ns */
 117         struct lustre_lock     ns_lock; /* protects hash, refcount, list */
 118         struct list_head       ns_list_chain; /* position in global NS list */
 119         /*
 120         struct proc_dir_entry *ns_proc_dir;
 121         */
 122
 123         struct list_head       ns_unused_list; /* all root resources in ns */
 124         unsigned int           ns_nr_unused;
 125         unsigned int           ns_max_unused;
 126
 127         spinlock_t             ns_counter_lock;
 128         __u64                  ns_locks;
 129         __u64                  ns_resources;
 130 };
 131
 132 /*
 133  *
 134  * Resource hash table
 135  *
 136  */
 137
 138 #define RES_HASH_BITS 14
 139 #define RES_HASH_SIZE (1UL << RES_HASH_BITS)
 140 #define RES_HASH_MASK (RES_HASH_SIZE - 1)
 141
 142 struct ldlm_lock;
 143
 144 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
 145                                       struct ldlm_lock_desc *new, void *data,
 146                                       int flag);
 147 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, int flags, void *data);
 148 typedef int (*ldlm_granted_callback)(struct ldlm_lock *,
 149                                      struct lustre_msg *, int offset);
 150
 151 struct ldlm_lock {
 152         struct portals_handle l_handle; // must be first in the structure
 153         atomic_t              l_refc;
 154         struct ldlm_resource *l_resource;
 155         struct ldlm_lock     *l_parent;
 156         struct list_head      l_children;
 157         struct list_head      l_childof;
 158         struct list_head      l_lru;
 159         struct list_head      l_res_link; // position in one of three res lists
 160         struct list_head      l_export_chain; // per-export chain of locks
 161         struct list_head      l_pending_chain; // locks with callbacks pending
 162         unsigned long         l_callback_timeout;
 163
 164         ldlm_mode_t           l_req_mode;
 165         ldlm_mode_t           l_granted_mode;
 166
 167         ldlm_completion_callback l_completion_ast;
 168         ldlm_blocking_callback   l_blocking_ast;
 169         ldlm_granted_callback l_granted_cb;
 170
 171         struct obd_export    *l_export;
 172         struct lustre_handle *l_connh;
 173         __u32                 l_flags;
 174         struct lustre_handle  l_remote_handle;
 175         void                 *l_data;
 176         void                 *l_cp_data;
 177         struct ldlm_extent    l_extent;
 178         __u32                 l_version[RES_VERSION_SIZE];
 179
 180         __u32                 l_readers;
 181         __u32                 l_writers;
 182         __u8                  l_destroyed;
 183
 184         /* If the lock is granted, a process sleeps on this waitq to learn when
 185          * it's no longer in use.  If the lock is not granted, a process sleeps
 186          * on this waitq to learn when it becomes granted. */
 187         wait_queue_head_t     l_waitq;
 188 };
 189
 190 typedef int (*ldlm_res_compat)(struct ldlm_lock *child, struct ldlm_lock *new);
 191 typedef int (*ldlm_res_policy)(struct ldlm_namespace *, struct ldlm_lock **,
 192                                void *req_cookie, ldlm_mode_t mode, int flags,
 193                                void *data);
 194
 195 #define LDLM_PLAIN       10
 196 #define LDLM_EXTENT      11
 197
 198 #define LDLM_MIN_TYPE 10
 199 #define LDLM_MAX_TYPE 11
 200
 201 struct ldlm_resource {
 202         struct ldlm_namespace *lr_namespace;
 203         struct list_head       lr_hash;
 204         struct ldlm_resource  *lr_parent;   /* 0 for a root resource */
 205         struct list_head       lr_children; /* list head for child resources */
 206         struct list_head       lr_childof;  /* part of ns_root_list if root res,
 207                                              * part of lr_children if child */
 208
 209         struct list_head       lr_granted;
 210         struct list_head       lr_converting;
 211         struct list_head       lr_waiting;
 212         ldlm_mode_t            lr_most_restr;
 213         __u32                  lr_type; /* LDLM_PLAIN or LDLM_EXTENT */
 214         struct ldlm_resource  *lr_root;
 215         struct ldlm_res_id     lr_name;
 216         __u32                  lr_version[RES_VERSION_SIZE];
 217         atomic_t               lr_refcount;
 218
 219         /* lr_tmp holds a list head temporarily, during the building of a work
 220          * queue.  see ldlm_add_ast_work_item and ldlm_run_ast_work */
 221         void                  *lr_tmp;
 222 };
 223
 224 struct ldlm_ast_work {
 225         struct ldlm_lock *w_lock;
 226         int               w_blocking;
 227         struct ldlm_lock_desc w_desc;
 228         struct list_head   w_list;
 229         int w_flags;
 230         void *w_data;
 231         int w_datalen;
 232 };
 233
 234 /* Per-export ldlm state. */
 235 struct ldlm_export_data {
 236         struct list_head       led_held_locks; /* protected by namespace lock */
 237         struct obd_import      led_import;
 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(level, lock, format, a...)                               \
 247 do {                                                                          \
 248         if (lock->l_resource == NULL) {                                       \
 249                 CDEBUG(level, "### " 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(level, "### " 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.name[0],                     \
 271                        lock->l_resource->lr_name.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(level, "### " 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.name[0],                     \
 289                        lock->l_resource->lr_name.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(lock, format, a...) __LDLM_DEBUG(D_DLMTRACE, lock, \
 297                                                     format, ## a)
 298 #define LDLM_ERROR(lock, format, a...) __LDLM_DEBUG(D_ERROR, lock, format, ## a)
 299
 300 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
 301         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
 302
 303 /*
 304  * Iterators.
 305  */
 306
 307 #define LDLM_ITER_CONTINUE 1 /* keep iterating */
 308 #define LDLM_ITER_STOP     0 /* stop iterating */
 309
 310 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
 311 typedef int (*ldlm_res_iterator_t)(struct ldlm_resource *, void *);
 312
 313 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
 314                           void *closure);
 315 int ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
 316                            void *closure);
 317 int ldlm_namespace_foreach_res(struct ldlm_namespace *ns,
 318                                ldlm_res_iterator_t iter, void *closure);
 319
 320 int ldlm_replay_locks(struct obd_import *imp);
 321
 322 /* ldlm_extent.c */
 323 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
 324 int ldlm_extent_policy(struct ldlm_namespace *, struct ldlm_lock **, void *,
 325                        ldlm_mode_t, int flags, void *);
 326
 327 /* ldlm_lockd.c */
 328 int ldlm_server_blocking_ast(struct ldlm_lock *, struct ldlm_lock_desc *,
 329                              void *data, int flag);
 330 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 331 int ldlm_handle_enqueue(struct ptlrpc_request *req, ldlm_completion_callback,
 332                         ldlm_blocking_callback);
 333 int ldlm_handle_convert(struct ptlrpc_request *req);
 334 int ldlm_handle_cancel(struct ptlrpc_request *req);
 335 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
 336
 337 /* ldlm_lock.c */
 338 void ldlm_register_intent(ldlm_res_policy arg);
 339 void ldlm_unregister_intent(void);
 340 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
 341 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int flags);
 342 void ldlm_cancel_callback(struct ldlm_lock *);
 343 int ldlm_lock_set_data(struct lustre_handle *, void *data, void *cp_data);
 344 void ldlm_lock_remove_from_lru(struct ldlm_lock *);
 345
 346 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
 347 {
 348         return __ldlm_handle2lock(h, 0);
 349 }
 350
 351 #define LDLM_LOCK_PUT(lock)                     \
 352 do {                                            \
 353         /*LDLM_DEBUG((lock), "put");*/          \
 354         ldlm_lock_put(lock);                    \
 355 } while (0)
 356
 357 #define LDLM_LOCK_GET(lock)                     \
 358 ({                                              \
 359         ldlm_lock_get(lock);                    \
 360         /*LDLM_DEBUG((lock), "get");*/          \
 361         lock;                                   \
 362 })
 363
 364 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
 365 void ldlm_lock_put(struct ldlm_lock *lock);
 366 void ldlm_lock_destroy(struct ldlm_lock *lock);
 367 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
 368 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
 369 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
 370 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
 371 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode);
 372 void ldlm_grant_lock(struct ldlm_lock *lock, void *data, int datalen);
 373 int ldlm_lock_match(struct ldlm_namespace *ns, int flags, struct ldlm_res_id *,
 374                     __u32 type, void *cookie, int cookielen, ldlm_mode_t mode,
 375                     struct lustre_handle *);
 376 struct ldlm_lock *
 377 ldlm_lock_create(struct ldlm_namespace *ns,
 378                  struct lustre_handle *parent_lock_handle, struct ldlm_res_id,
 379                  __u32 type, ldlm_mode_t mode, void *data, void *cp_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                                ldlm_blocking_callback blocking);
 384 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
 385                                         int *flags);
 386 void ldlm_lock_cancel(struct ldlm_lock *lock);
 387 void ldlm_cancel_locks_for_export(struct obd_export *export);
 388 int ldlm_run_ast_work(struct list_head *rpc_list);
 389 void ldlm_reprocess_all(struct ldlm_resource *res);
 390 void ldlm_reprocess_all_ns(struct ldlm_namespace *ns);
 391 void ldlm_lock_dump(int level, struct ldlm_lock *lock);
 392 void ldlm_lock_dump_handle(int level, struct lustre_handle *);
 393
 394 /* ldlm_test.c */
 395 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
 396 int ldlm_regression_start(struct obd_device *obddev,
 397                           struct lustre_handle *connh,
 398                           unsigned int threads, unsigned int max_locks_in,
 399                           unsigned int num_resources_in,
 400                           unsigned int num_extents_in);
 401 int ldlm_regression_stop(void);
 402
 403
 404 /* resource.c */
 405 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
 406 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
 407 int ldlm_namespace_free(struct ldlm_namespace *ns);
 408 int ldlm_proc_setup(struct obd_device *obd);
 409 void ldlm_proc_cleanup(struct obd_device *obd);
 410
 411 /* resource.c - internal */
 412 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
 413                                         struct ldlm_resource *parent,
 414                                         struct ldlm_res_id, __u32 type,
 415                                         int create);
 416 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
 417 int ldlm_resource_putref(struct ldlm_resource *res);
 418 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
 419                             struct ldlm_lock *lock);
 420 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
 421 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
 422 void ldlm_dump_all_namespaces(void);
 423 void ldlm_namespace_dump(struct ldlm_namespace *);
 424 void ldlm_resource_dump(struct ldlm_resource *);
 425 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
 426                               struct ldlm_res_id);
 427
 428 /* ldlm_request.c */
 429 int ldlm_expired_completion_wait(void *data);
 430 int ldlm_completion_ast(struct ldlm_lock *lock, int flags, void *data);
 431 int ldlm_cli_enqueue(struct lustre_handle *conn,
 432                      struct ptlrpc_request *req,
 433                      struct ldlm_namespace *ns,
 434                      struct lustre_handle *parent_lock_handle,
 435                      struct ldlm_res_id,
 436                      __u32 type,
 437                      void *cookie, int cookielen,
 438                      ldlm_mode_t mode,
 439                      int *flags,
 440                      ldlm_completion_callback completion,
 441                      ldlm_blocking_callback callback,
 442                      void *data,
 443                      void *cp_data,
 444                      struct lustre_handle *lockh);
 445 int ldlm_match_or_enqueue(struct lustre_handle *connh,
 446                           struct ptlrpc_request *req,
 447                           struct ldlm_namespace *ns,
 448                           struct lustre_handle *parent_lock_handle,
 449                           struct ldlm_res_id,
 450                           __u32 type,
 451                           void *cookie, int cookielen,
 452                           ldlm_mode_t mode,
 453                           int *flags,
 454                           ldlm_completion_callback completion,
 455                           ldlm_blocking_callback callback,
 456                           void *data,
 457                           void *cp_data,
 458                           struct lustre_handle *lockh);
 459 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
 460                     void *data, __u32 data_len);
 461 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
 462 int ldlm_cli_cancel(struct lustre_handle *lockh);
 463 int ldlm_cli_cancel_unused(struct ldlm_namespace *, struct ldlm_res_id *,
 464                            int flags);
 465 int ldlm_cancel_lru(struct ldlm_namespace *ns);
 466
 467 /* mds/handler.c */
 468 /* This has to be here because recurisve inclusion sucks. */
 469 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 470                      void *data, int flag);
 471
 472
 473 /* ioctls for trying requests */
 474 #define IOC_LDLM_TYPE                   'f'
 475 #define IOC_LDLM_MIN_NR                 40
 476
 477 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
 478 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
 479 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
 480 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
 481 #define IOC_LDLM_MAX_NR                 43
 482
 483 #endif