Whamcloud - gitweb
most of the work for lock replay:
[fs/lustre-release.git] / lustre / include / linux / lustre_dlm.h
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  */
4
5 #ifndef _LUSTRE_DLM_H__
6 #define _LUSTRE_DLM_H__
7
8 #ifdef __KERNEL__
9
10 #include <linux/proc_fs.h>
11 #include <linux/lustre_lib.h>
12 #include <linux/lustre_net.h>
13 #include <linux/lustre_import.h>
14
15 struct obd_ops;
16 struct obd_device;
17
18 #define OBD_LDLM_DEVICENAME  "ldlm"
19
20 typedef enum {
21         ELDLM_OK = 0,
22
23         ELDLM_LOCK_CHANGED = 300,
24         ELDLM_LOCK_ABORTED = 301,
25
26         ELDLM_NAMESPACE_EXISTS = 400,
27         ELDLM_BAD_NAMESPACE    = 401
28 } ldlm_error_t;
29
30 #define LDLM_NAMESPACE_SERVER 0
31 #define LDLM_NAMESPACE_CLIENT 1
32
33 #define LDLM_FL_LOCK_CHANGED   (1 << 0)
34 #define LDLM_FL_BLOCK_GRANTED  (1 << 1)
35 #define LDLM_FL_BLOCK_CONV     (1 << 2)
36 #define LDLM_FL_BLOCK_WAIT     (1 << 3)
37 #define LDLM_FL_CBPENDING      (1 << 4)
38 #define LDLM_FL_AST_SENT       (1 << 5)
39 #define LDLM_FL_DESTROYED      (1 << 6)
40 #define LDLM_FL_WAIT_NOREPROC  (1 << 7)
41 #define LDLM_FL_CANCEL         (1 << 8)
42 #define LDLM_FL_REPLAY         (1 << 9)
43
44 #define LDLM_CB_BLOCKING    1
45 #define LDLM_CB_CANCELING   2
46
47 #define L2B(c) (1 << c)
48
49 /* compatibility matrix */
50 #define LCK_COMPAT_EX  L2B(LCK_NL)
51 #define LCK_COMPAT_PW  (LCK_COMPAT_EX | L2B(LCK_CR))
52 #define LCK_COMPAT_PR  (LCK_COMPAT_PW | L2B(LCK_PR))
53 #define LCK_COMPAT_CW  (LCK_COMPAT_PW | L2B(LCK_CW))
54 #define LCK_COMPAT_CR  (LCK_COMPAT_CW | L2B(LCK_PR) | L2B(LCK_PW))
55 #define LCK_COMPAT_NL  (LCK_COMPAT_CR | L2B(LCK_EX))
56
57 static ldlm_mode_t lck_compat_array[] = {
58         [LCK_EX] LCK_COMPAT_EX,
59         [LCK_PW] LCK_COMPAT_PW,
60         [LCK_PR] LCK_COMPAT_PR,
61         [LCK_CW] LCK_COMPAT_CW,
62         [LCK_CR] LCK_COMPAT_CR,
63         [LCK_NL] LCK_COMPAT_NL
64 };
65
66 static inline int lockmode_compat(ldlm_mode_t exist, ldlm_mode_t new)
67 {
68        if (exist < LCK_EX || exist > LCK_NL)
69               LBUG();
70        if (new < LCK_EX || new > LCK_NL)
71               LBUG();
72
73        return (lck_compat_array[exist] & L2B(new));
74 }
75
76 /*
77  *
78  * cluster name spaces
79  *
80  */
81
82 #define DLM_OST_NAMESPACE 1
83 #define DLM_MDS_NAMESPACE 2
84
85 /* XXX
86    - do we just separate this by security domains and use a prefix for
87      multiple namespaces in the same domain?
88    -
89 */
90
91 struct ldlm_namespace {
92         char                  *ns_name;
93         __u32                  ns_client; /* is this a client-side lock tree? */
94         struct list_head      *ns_hash; /* hash table for ns */
95         __u32                  ns_refcount; /* count of resources in the hash */
96         struct list_head       ns_root_list; /* all root resources in ns */
97         struct lustre_lock     ns_lock; /* protects hash, refcount, list */
98         struct list_head       ns_list_chain; /* position in global NS list */
99         struct proc_dir_entry *ns_proc_dir;
100
101         spinlock_t             ns_counter_lock;
102         __u64                  ns_locks;
103         __u64                  ns_resources;
104 };
105
106 /*
107  *
108  * Resource hash table
109  *
110  */
111
112 #define RES_HASH_BITS 14
113 #define RES_HASH_SIZE (1UL << RES_HASH_BITS)
114 #define RES_HASH_MASK (RES_HASH_SIZE - 1)
115
116 struct ldlm_lock;
117
118 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
119                                       struct ldlm_lock_desc *new, void *data,
120                                       __u32 data_len, int flag);
121
122 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, int flags);
123
124 struct ldlm_lock {
125         __u64                 l_random;
126         int                   l_refc;
127         struct ldlm_resource *l_resource;
128         struct ldlm_lock     *l_parent;
129         struct list_head      l_children;
130         struct list_head      l_childof;
131         struct list_head      l_res_link; /*position in one of three res lists*/
132         struct list_head      l_export_chain; /* per-export chain of locks */
133         struct list_head      l_pending_chain; /* locks with callbacks pending*/
134         unsigned long         l_callback_timeout;
135
136         ldlm_mode_t           l_req_mode;
137         ldlm_mode_t           l_granted_mode;
138
139         ldlm_completion_callback    l_completion_ast;
140         ldlm_blocking_callback    l_blocking_ast;
141
142         struct obd_export    *l_export;
143         struct lustre_handle *l_connh;
144         __u32                 l_flags;
145         struct lustre_handle   l_remote_handle;
146         void                 *l_data;
147         __u32                 l_data_len;
148         struct ldlm_extent    l_extent;
149         __u32                 l_version[RES_VERSION_SIZE];
150
151         __u32                 l_readers;
152         __u32                 l_writers;
153
154         /* If the lock is granted, a process sleeps on this waitq to learn when
155          * it's no longer in use.  If the lock is not granted, a process sleeps
156          * on this waitq to learn when it becomes granted. */
157         wait_queue_head_t     l_waitq;
158 };
159
160 typedef int (*ldlm_res_compat)(struct ldlm_lock *child, struct ldlm_lock *new);
161 typedef int (*ldlm_res_policy)(struct ldlm_lock *lock, void *req_cookie,
162                                ldlm_mode_t mode, void *data);
163
164 #define LDLM_PLAIN       10
165 #define LDLM_EXTENT      11
166 #define LDLM_MDSINTENT   12
167
168 #define LDLM_MIN_TYPE 10
169 #define LDLM_MAX_TYPE 12
170
171 extern ldlm_res_compat ldlm_res_compat_table [];
172 extern ldlm_res_policy ldlm_res_policy_table [];
173
174 struct ldlm_resource {
175         struct ldlm_namespace *lr_namespace;
176         struct list_head       lr_hash;
177         struct ldlm_resource  *lr_parent;   /* 0 for a root resource */
178         struct list_head       lr_children; /* list head for child resources */
179         struct list_head       lr_childof;  /* part of ns_root_list if root res,
180                                              * part of lr_children if child */
181
182         struct list_head       lr_granted;
183         struct list_head       lr_converting;
184         struct list_head       lr_waiting;
185         ldlm_mode_t            lr_most_restr;
186         __u32                  lr_type; /* PLAIN, EXTENT, or MDSINTENT */
187         struct ldlm_resource  *lr_root;
188         __u64                  lr_name[RES_NAME_SIZE];
189         __u32                  lr_version[RES_VERSION_SIZE];
190         atomic_t               lr_refcount;
191         void                  *lr_tmp;
192 };
193
194 struct ldlm_ast_work {
195         struct ldlm_lock *w_lock;
196         int               w_blocking;
197         struct ldlm_lock_desc w_desc;
198         struct list_head   w_list;
199         int w_flags;
200         void *w_data;
201         int w_datalen;
202 };
203
204 /* Per-export ldlm state. */
205 struct ldlm_export_data {
206         struct list_head        led_held_locks;
207         struct obd_import       led_import;
208 };
209
210 static inline struct ldlm_extent *ldlm_res2extent(struct ldlm_resource *res)
211 {
212         return (struct ldlm_extent *)(res->lr_name);
213 }
214
215 extern struct obd_ops ldlm_obd_ops;
216
217 extern char *ldlm_lockname[];
218 extern char *ldlm_typename[];
219 extern char *ldlm_it2str(int it);
220
221 #define LDLM_DEBUG(lock, format, a...)                                        \
222 do {                                                                          \
223         if (lock->l_resource == NULL) {                                       \
224                 CDEBUG(D_DLMTRACE, "### " format                              \
225                        " (UNKNOWN: lock %p(rc=%d/%d,%d) mode %s/%s on "       \
226                        "res \?\? (rc=\?\?) type \?\?\? remote "LPX64")\n"     \
227                        , ## a, lock, lock->l_refc, lock->l_readers,           \
228                        lock->l_writers,                                       \
229                        ldlm_lockname[lock->l_granted_mode],                   \
230                        ldlm_lockname[lock->l_req_mode],                       \
231                        lock->l_remote_handle.addr);                           \
232                 break;                                                        \
233         }                                                                     \
234         if (lock->l_resource->lr_type == LDLM_EXTENT) {                       \
235                 CDEBUG(D_DLMTRACE, "### " format                              \
236                        " (%s: lock %p(rc=%d/%d,%d) mode %s/%s on res "        \
237                        LPU64" (rc=%d) type %s ["LPU64"->"LPU64"] remote "     \
238                        LPX64")\n" , ## a,                                     \
239                        lock->l_resource->lr_namespace->ns_name, lock,         \
240                        lock->l_refc, lock->l_readers, lock->l_writers,        \
241                        ldlm_lockname[lock->l_granted_mode],                   \
242                        ldlm_lockname[lock->l_req_mode],                       \
243                        lock->l_resource->lr_name[0],                          \
244                        atomic_read(&lock->l_resource->lr_refcount),           \
245                        ldlm_typename[lock->l_resource->lr_type],              \
246                        lock->l_extent.start, lock->l_extent.end,              \
247                        lock->l_remote_handle.addr);                           \
248                 break;                                                        \
249         }                                                                     \
250         {                                                                     \
251                 CDEBUG(D_DLMTRACE, "### " format                              \
252                        " (%s: lock %p(rc=%d/%d,%d) mode %s/%s on res "        \
253                        LPU64" (rc=%d) type %s remote "LPX64")\n" , ## a,      \
254                        lock->l_resource->lr_namespace->ns_name, lock,         \
255                        lock->l_refc, lock->l_readers, lock->l_writers,        \
256                        ldlm_lockname[lock->l_granted_mode],                   \
257                        ldlm_lockname[lock->l_req_mode],                       \
258                        lock->l_resource->lr_name[0],                          \
259                        atomic_read(&lock->l_resource->lr_refcount),           \
260                        ldlm_typename[lock->l_resource->lr_type],              \
261                        lock->l_remote_handle.addr);                           \
262         }                                                                     \
263 } while (0)
264
265 #define LDLM_DEBUG_NOLOCK(format, a...)                 \
266         CDEBUG(D_DLMTRACE, "### " format "\n" , ## a)
267
268 /* ldlm_extent.c */
269 int ldlm_extent_compat(struct ldlm_lock *, struct ldlm_lock *);
270 int ldlm_extent_policy(struct ldlm_lock *, void *, ldlm_mode_t, void *);
271
272 /* ldlm_lockd.c */
273 int ldlm_handle_enqueue(struct ptlrpc_request *req);
274 int ldlm_handle_convert(struct ptlrpc_request *req);
275 int ldlm_handle_cancel(struct ptlrpc_request *req);
276 int ldlm_del_waiting_lock(struct ldlm_lock *lock);
277
278 /* ldlm_lock.c */
279 void ldlm_register_intent(int (*arg)(struct ldlm_lock *lock, void *req_cookie,
280                                      ldlm_mode_t mode, void *data));
281 void ldlm_unregister_intent(void);
282 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
283 struct ldlm_lock *__ldlm_handle2lock(struct lustre_handle *, int strict);
284 void ldlm_lock2handle(struct ldlm_lock *lock, struct lustre_handle *lockh);
285 void ldlm_cancel_callback(struct ldlm_lock *lock);
286
287 static inline struct ldlm_lock *ldlm_handle2lock(struct lustre_handle *h)
288 {
289         return __ldlm_handle2lock(h, 1);
290 }
291
292 #define LDLM_LOCK_PUT(lock)                     \
293 do {                                            \
294         /*LDLM_DEBUG(lock, "put");*/            \
295         ldlm_lock_put(lock);                    \
296 } while (0)
297
298 #define LDLM_LOCK_GET(lock)                     \
299 ({                                              \
300         ldlm_lock_get(lock);                    \
301         /*LDLM_DEBUG(lock, "get");*/            \
302         lock;                                   \
303 })
304
305 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
306 void ldlm_lock_put(struct ldlm_lock *lock);
307 void ldlm_lock_destroy(struct ldlm_lock *lock);
308 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
309 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
310 void ldlm_lock_addref_internal(struct ldlm_lock *, __u32 mode);
311 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
312 void ldlm_grant_lock(struct ldlm_lock *lock);
313 int ldlm_lock_match(struct ldlm_namespace *ns, __u64 *res_id, __u32 type,
314                     void *cookie, int cookielen, ldlm_mode_t mode,
315                     struct lustre_handle *lockh);
316 struct ldlm_lock *
317 ldlm_lock_create(struct ldlm_namespace *ns,
318                  struct lustre_handle *parent_lock_handle,
319                  __u64 *res_id, __u32 type, ldlm_mode_t mode, void *data,
320                  __u32 data_len);
321 ldlm_error_t ldlm_lock_enqueue(struct ldlm_lock *lock, void *cookie,
322                                int cookie_len, int *flags,
323                                ldlm_completion_callback completion,
324                                ldlm_blocking_callback blocking);
325 struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock, int new_mode,
326                                         int *flags);
327 void ldlm_lock_cancel(struct ldlm_lock *lock);
328 void ldlm_cancel_locks_for_export(struct obd_export *export);
329 void ldlm_run_ast_work(struct list_head *rpc_list);
330 void ldlm_reprocess_all(struct ldlm_resource *res);
331 void ldlm_lock_dump(struct ldlm_lock *lock);
332
333 /* ldlm_test.c */
334 int ldlm_test(struct obd_device *device, struct lustre_handle *connh);
335 int ldlm_regression_start(struct obd_device *obddev,
336                           struct lustre_handle *connh,
337                           unsigned int threads, unsigned int max_locks_in,
338                           unsigned int num_resources_in,
339                           unsigned int num_extents_in);
340 int ldlm_regression_stop(void);
341
342
343 /* resource.c */
344 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 local);
345 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int local_only);
346 int ldlm_namespace_free(struct ldlm_namespace *ns);
347 int ldlm_proc_setup(struct obd_device *obd);
348 void ldlm_proc_cleanup(struct obd_device *obd);
349
350 /* resource.c - internal */
351 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
352                                         struct ldlm_resource *parent,
353                                         __u64 *name, __u32 type, int create);
354 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
355 int ldlm_resource_put(struct ldlm_resource *res);
356 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
357                             struct ldlm_lock *lock);
358 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
359 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
360 void ldlm_dump_all_namespaces(void);
361 void ldlm_namespace_dump(struct ldlm_namespace *);
362 void ldlm_resource_dump(struct ldlm_resource *);
363 int ldlm_lock_change_resource(struct ldlm_lock *, __u64 new_resid[3]);
364
365 /* ldlm_request.c */
366 int ldlm_completion_ast(struct ldlm_lock *lock, int flags);
367 int ldlm_cli_enqueue(struct lustre_handle *conn,
368                      struct ptlrpc_request *req,
369                      struct ldlm_namespace *ns,
370                      struct lustre_handle *parent_lock_handle,
371                      __u64 *res_id,
372                      __u32 type,
373                      void *cookie, int cookielen,
374                      ldlm_mode_t mode,
375                      int *flags,
376                      ldlm_completion_callback completion,
377                      ldlm_blocking_callback callback,
378                      void *data,
379                      __u32 data_len,
380                      struct lustre_handle *lockh);
381 int ldlm_match_or_enqueue(struct lustre_handle *connh,
382                           struct ptlrpc_request *req,
383                           struct ldlm_namespace *ns,
384                           struct lustre_handle *parent_lock_handle,
385                           __u64 *res_id,
386                           __u32 type,
387                           void *cookie, int cookielen,
388                           ldlm_mode_t mode,
389                           int *flags,
390                           ldlm_completion_callback completion,
391                           ldlm_blocking_callback callback,
392                           void *data,
393                           __u32 data_len,
394                           struct lustre_handle *lockh);
395 int ldlm_server_ast(struct lustre_handle *lockh, struct ldlm_lock_desc *new,
396                     void *data, __u32 data_len);
397 int ldlm_cli_convert(struct lustre_handle *, int new_mode, int *flags);
398 int ldlm_cli_cancel(struct lustre_handle *lockh);
399 int ldlm_cli_cancel_unused(struct ldlm_namespace *, __u64 *, int local_only);
400
401 /* mds/handler.c */
402 /* This has to be here because recurisve inclusion sucks. */
403 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
404                      void *data, __u32 data_len, int flag);
405
406 #endif /* __KERNEL__ */
407
408 /* ioctls for trying requests */
409 #define IOC_LDLM_TYPE                   'f'
410 #define IOC_LDLM_MIN_NR                 40
411
412 #define IOC_LDLM_TEST                   _IOWR('f', 40, long)
413 #define IOC_LDLM_DUMP                   _IOWR('f', 41, long)
414 #define IOC_LDLM_REGRESS_START          _IOWR('f', 42, long)
415 #define IOC_LDLM_REGRESS_STOP           _IOWR('f', 43, long)
416 #define IOC_LDLM_MAX_NR                 43
417
418 #endif