Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / staging / lustre / lustre / ldlm / ldlm_lockd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2012, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/ldlm/ldlm_lockd.c
37  *
38  * Author: Peter Braam <braam@clusterfs.com>
39  * Author: Phil Schwan <phil@clusterfs.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_LDLM
43
44 #include "../../include/linux/libcfs/libcfs.h"
45 #include "../include/lustre_dlm.h"
46 #include "../include/obd_class.h"
47 #include <linux/list.h>
48 #include "ldlm_internal.h"
49
50 static int ldlm_num_threads;
51 module_param(ldlm_num_threads, int, 0444);
52 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
53
54 static char *ldlm_cpts;
55 module_param(ldlm_cpts, charp, 0444);
56 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
57
58 static struct mutex     ldlm_ref_mutex;
59 static int ldlm_refcount;
60
61 struct ldlm_cb_async_args {
62         struct ldlm_cb_set_arg *ca_set_arg;
63         struct ldlm_lock       *ca_lock;
64 };
65
66 /* LDLM state */
67
68 static struct ldlm_state *ldlm_state;
69
70 inline unsigned long round_timeout(unsigned long timeout)
71 {
72         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
73 }
74
75 /* timeout for initial callback (AST) reply (bz10399) */
76 static inline unsigned int ldlm_get_rq_timeout(void)
77 {
78         /* Non-AT value */
79         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
80
81         return timeout < 1 ? 1 : timeout;
82 }
83
84 #define ELT_STOPPED   0
85 #define ELT_READY     1
86 #define ELT_TERMINATE 2
87
88 struct ldlm_bl_pool {
89         spinlock_t              blp_lock;
90
91         /*
92          * blp_prio_list is used for callbacks that should be handled
93          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
94          * see bug 13843
95          */
96         struct list_head              blp_prio_list;
97
98         /*
99          * blp_list is used for all other callbacks which are likely
100          * to take longer to process.
101          */
102         struct list_head              blp_list;
103
104         wait_queue_head_t            blp_waitq;
105         struct completion       blp_comp;
106         atomic_t            blp_num_threads;
107         atomic_t            blp_busy_threads;
108         int                  blp_min_threads;
109         int                  blp_max_threads;
110 };
111
112 struct ldlm_bl_work_item {
113         struct list_head              blwi_entry;
114         struct ldlm_namespace  *blwi_ns;
115         struct ldlm_lock_desc   blwi_ld;
116         struct ldlm_lock       *blwi_lock;
117         struct list_head              blwi_head;
118         int                  blwi_count;
119         struct completion       blwi_comp;
120         ldlm_cancel_flags_t     blwi_flags;
121         int                  blwi_mem_pressure;
122 };
123
124
125 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
126 {
127         return 0;
128 }
129
130 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
131 {
132         return 0;
133 }
134
135
136
137 /**
138  * Callback handler for receiving incoming blocking ASTs.
139  *
140  * This can only happen on client side.
141  */
142 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
143                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
144 {
145         int do_ast;
146
147         LDLM_DEBUG(lock, "client blocking AST callback handler");
148
149         lock_res_and_lock(lock);
150         lock->l_flags |= LDLM_FL_CBPENDING;
151
152         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
153                 lock->l_flags |= LDLM_FL_CANCEL;
154
155         do_ast = !lock->l_readers && !lock->l_writers;
156         unlock_res_and_lock(lock);
157
158         if (do_ast) {
159                 CDEBUG(D_DLMTRACE,
160                        "Lock %p already unused, calling callback (%p)\n", lock,
161                        lock->l_blocking_ast);
162                 if (lock->l_blocking_ast != NULL)
163                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
164                                              LDLM_CB_BLOCKING);
165         } else {
166                 CDEBUG(D_DLMTRACE,
167                        "Lock %p is referenced, will be cancelled later\n",
168                        lock);
169         }
170
171         LDLM_DEBUG(lock, "client blocking callback handler END");
172         LDLM_LOCK_RELEASE(lock);
173 }
174
175 /**
176  * Callback handler for receiving incoming completion ASTs.
177  *
178  * This only can happen on client side.
179  */
180 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
181                                     struct ldlm_namespace *ns,
182                                     struct ldlm_request *dlm_req,
183                                     struct ldlm_lock *lock)
184 {
185         int lvb_len;
186         LIST_HEAD(ast_list);
187         int rc = 0;
188
189         LDLM_DEBUG(lock, "client completion callback handler START");
190
191         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
192                 int to = cfs_time_seconds(1);
193
194                 while (to > 0) {
195                         set_current_state(TASK_INTERRUPTIBLE);
196                         schedule_timeout(to);
197                         if (lock->l_granted_mode == lock->l_req_mode ||
198                             lock->l_flags & LDLM_FL_DESTROYED)
199                                 break;
200                 }
201         }
202
203         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
204         if (lvb_len < 0) {
205                 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
206                 rc = lvb_len;
207                 goto out;
208         } else if (lvb_len > 0) {
209                 if (lock->l_lvb_len > 0) {
210                         /* for extent lock, lvb contains ost_lvb{}. */
211                         LASSERT(lock->l_lvb_data != NULL);
212
213                         if (unlikely(lock->l_lvb_len < lvb_len)) {
214                                 LDLM_ERROR(lock, "Replied LVB is larger than expectation, expected = %d, replied = %d",
215                                            lock->l_lvb_len, lvb_len);
216                                 rc = -EINVAL;
217                                 goto out;
218                         }
219                 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
220                                                      * variable length */
221                         void *lvb_data;
222
223                         OBD_ALLOC(lvb_data, lvb_len);
224                         if (lvb_data == NULL) {
225                                 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
226                                 rc = -ENOMEM;
227                                 goto out;
228                         }
229
230                         lock_res_and_lock(lock);
231                         LASSERT(lock->l_lvb_data == NULL);
232                         lock->l_lvb_type = LVB_T_LAYOUT;
233                         lock->l_lvb_data = lvb_data;
234                         lock->l_lvb_len = lvb_len;
235                         unlock_res_and_lock(lock);
236                 }
237         }
238
239         lock_res_and_lock(lock);
240         if ((lock->l_flags & LDLM_FL_DESTROYED) ||
241             lock->l_granted_mode == lock->l_req_mode) {
242                 /* bug 11300: the lock has already been granted */
243                 unlock_res_and_lock(lock);
244                 LDLM_DEBUG(lock, "Double grant race happened");
245                 rc = 0;
246                 goto out;
247         }
248
249         /* If we receive the completion AST before the actual enqueue returned,
250          * then we might need to switch lock modes, resources, or extents. */
251         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
252                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
253                 LDLM_DEBUG(lock, "completion AST, new lock mode");
254         }
255
256         if (lock->l_resource->lr_type != LDLM_PLAIN) {
257                 ldlm_convert_policy_to_local(req->rq_export,
258                                           dlm_req->lock_desc.l_resource.lr_type,
259                                           &dlm_req->lock_desc.l_policy_data,
260                                           &lock->l_policy_data);
261                 LDLM_DEBUG(lock, "completion AST, new policy data");
262         }
263
264         ldlm_resource_unlink_lock(lock);
265         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
266                    &lock->l_resource->lr_name,
267                    sizeof(lock->l_resource->lr_name)) != 0) {
268                 unlock_res_and_lock(lock);
269                 rc = ldlm_lock_change_resource(ns, lock,
270                                 &dlm_req->lock_desc.l_resource.lr_name);
271                 if (rc < 0) {
272                         LDLM_ERROR(lock, "Failed to allocate resource");
273                         goto out;
274                 }
275                 LDLM_DEBUG(lock, "completion AST, new resource");
276                 CERROR("change resource!\n");
277                 lock_res_and_lock(lock);
278         }
279
280         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
281                 /* BL_AST locks are not needed in LRU.
282                  * Let ldlm_cancel_lru() be fast. */
283                 ldlm_lock_remove_from_lru(lock);
284                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
285                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
286         }
287
288         if (lock->l_lvb_len > 0) {
289                 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
290                                    lock->l_lvb_data, lvb_len);
291                 if (rc < 0) {
292                         unlock_res_and_lock(lock);
293                         goto out;
294                 }
295         }
296
297         ldlm_grant_lock(lock, &ast_list);
298         unlock_res_and_lock(lock);
299
300         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
301
302         /* Let Enqueue to call osc_lock_upcall() and initialize
303          * l_ast_data */
304         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
305
306         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
307
308         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
309                           lock);
310         goto out;
311
312 out:
313         if (rc < 0) {
314                 lock_res_and_lock(lock);
315                 lock->l_flags |= LDLM_FL_FAILED;
316                 unlock_res_and_lock(lock);
317                 wake_up(&lock->l_waitq);
318         }
319         LDLM_LOCK_RELEASE(lock);
320 }
321
322 /**
323  * Callback handler for receiving incoming glimpse ASTs.
324  *
325  * This only can happen on client side.  After handling the glimpse AST
326  * we also consider dropping the lock here if it is unused locally for a
327  * long time.
328  */
329 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
330                                     struct ldlm_namespace *ns,
331                                     struct ldlm_request *dlm_req,
332                                     struct ldlm_lock *lock)
333 {
334         int rc = -ENOSYS;
335
336         LDLM_DEBUG(lock, "client glimpse AST callback handler");
337
338         if (lock->l_glimpse_ast != NULL)
339                 rc = lock->l_glimpse_ast(lock, req);
340
341         if (req->rq_repmsg != NULL) {
342                 ptlrpc_reply(req);
343         } else {
344                 req->rq_status = rc;
345                 ptlrpc_error(req);
346         }
347
348         lock_res_and_lock(lock);
349         if (lock->l_granted_mode == LCK_PW &&
350             !lock->l_readers && !lock->l_writers &&
351             cfs_time_after(cfs_time_current(),
352                            cfs_time_add(lock->l_last_used,
353                                         cfs_time_seconds(10)))) {
354                 unlock_res_and_lock(lock);
355                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
356                         ldlm_handle_bl_callback(ns, NULL, lock);
357
358                 return;
359         }
360         unlock_res_and_lock(lock);
361         LDLM_LOCK_RELEASE(lock);
362 }
363
364 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
365 {
366         if (req->rq_no_reply)
367                 return 0;
368
369         req->rq_status = rc;
370         if (!req->rq_packed_final) {
371                 rc = lustre_pack_reply(req, 1, NULL, NULL);
372                 if (rc)
373                         return rc;
374         }
375         return ptlrpc_reply(req);
376 }
377
378 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
379                                ldlm_cancel_flags_t cancel_flags)
380 {
381         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
382
383         spin_lock(&blp->blp_lock);
384         if (blwi->blwi_lock &&
385             blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
386                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
387                 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
388         } else {
389                 /* other blocking callbacks are added to the regular list */
390                 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
391         }
392         spin_unlock(&blp->blp_lock);
393
394         wake_up(&blp->blp_waitq);
395
396         /* can not check blwi->blwi_flags as blwi could be already freed in
397            LCF_ASYNC mode */
398         if (!(cancel_flags & LCF_ASYNC))
399                 wait_for_completion(&blwi->blwi_comp);
400
401         return 0;
402 }
403
404 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
405                              struct ldlm_namespace *ns,
406                              struct ldlm_lock_desc *ld,
407                              struct list_head *cancels, int count,
408                              struct ldlm_lock *lock,
409                              ldlm_cancel_flags_t cancel_flags)
410 {
411         init_completion(&blwi->blwi_comp);
412         INIT_LIST_HEAD(&blwi->blwi_head);
413
414         if (memory_pressure_get())
415                 blwi->blwi_mem_pressure = 1;
416
417         blwi->blwi_ns = ns;
418         blwi->blwi_flags = cancel_flags;
419         if (ld != NULL)
420                 blwi->blwi_ld = *ld;
421         if (count) {
422                 list_add(&blwi->blwi_head, cancels);
423                 list_del_init(cancels);
424                 blwi->blwi_count = count;
425         } else {
426                 blwi->blwi_lock = lock;
427         }
428 }
429
430 /**
431  * Queues a list of locks \a cancels containing \a count locks
432  * for later processing by a blocking thread.  If \a count is zero,
433  * then the lock referenced as \a lock is queued instead.
434  *
435  * The blocking thread would then call ->l_blocking_ast callback in the lock.
436  * If list addition fails an error is returned and caller is supposed to
437  * call ->l_blocking_ast itself.
438  */
439 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
440                              struct ldlm_lock_desc *ld,
441                              struct ldlm_lock *lock,
442                              struct list_head *cancels, int count,
443                              ldlm_cancel_flags_t cancel_flags)
444 {
445         if (cancels && count == 0)
446                 return 0;
447
448         if (cancel_flags & LCF_ASYNC) {
449                 struct ldlm_bl_work_item *blwi;
450
451                 OBD_ALLOC(blwi, sizeof(*blwi));
452                 if (blwi == NULL)
453                         return -ENOMEM;
454                 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
455
456                 return __ldlm_bl_to_thread(blwi, cancel_flags);
457         } else {
458                 /* if it is synchronous call do minimum mem alloc, as it could
459                  * be triggered from kernel shrinker
460                  */
461                 struct ldlm_bl_work_item blwi;
462
463                 memset(&blwi, 0, sizeof(blwi));
464                 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
465                 return __ldlm_bl_to_thread(&blwi, cancel_flags);
466         }
467 }
468
469
470 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
471                            struct ldlm_lock *lock)
472 {
473         return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
474 }
475
476 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
477                            struct list_head *cancels, int count,
478                            ldlm_cancel_flags_t cancel_flags)
479 {
480         return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
481 }
482
483 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
484 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
485 {
486         struct obd_device *obd = req->rq_export->exp_obd;
487         char *key;
488         void *val;
489         int keylen, vallen;
490         int rc = -ENOSYS;
491
492         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
493
494         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
495
496         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
497         if (key == NULL) {
498                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
499                 return -EFAULT;
500         }
501         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
502                                       RCL_CLIENT);
503         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
504         if (val == NULL) {
505                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
506                 return -EFAULT;
507         }
508         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
509                                       RCL_CLIENT);
510
511         /* We are responsible for swabbing contents of val */
512
513         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
514                 /* Pass it on to mdc (the "export" in this case) */
515                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
516                                         req->rq_export,
517                                         sizeof(KEY_HSM_COPYTOOL_SEND),
518                                         KEY_HSM_COPYTOOL_SEND,
519                                         vallen, val, NULL);
520         else
521                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
522
523         return rc;
524 }
525
526 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
527                                         const char *msg, int rc,
528                                         struct lustre_handle *handle)
529 {
530         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
531                   "%s: [nid %s] [rc %d] [lock %#llx]",
532                   msg, libcfs_id2str(req->rq_peer), rc,
533                   handle ? handle->cookie : 0);
534         if (req->rq_no_reply)
535                 CWARN("No reply was sent, maybe cause bug 21636.\n");
536         else if (rc)
537                 CWARN("Send reply failed, maybe cause bug 21636.\n");
538 }
539
540 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
541 {
542         struct obd_quotactl *oqctl;
543         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
544
545         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
546         if (oqctl == NULL) {
547                 CERROR("Can't unpack obd_quotactl\n");
548                 return -EPROTO;
549         }
550
551         oqctl->qc_stat = ptlrpc_status_ntoh(oqctl->qc_stat);
552
553         cli->cl_qchk_stat = oqctl->qc_stat;
554         return 0;
555 }
556
557 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
558 static int ldlm_callback_handler(struct ptlrpc_request *req)
559 {
560         struct ldlm_namespace *ns;
561         struct ldlm_request *dlm_req;
562         struct ldlm_lock *lock;
563         int rc;
564
565         /* Requests arrive in sender's byte order.  The ptlrpc service
566          * handler has already checked and, if necessary, byte-swapped the
567          * incoming request message body, but I am responsible for the
568          * message buffers. */
569
570         /* do nothing for sec context finalize */
571         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
572                 return 0;
573
574         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
575
576         if (req->rq_export == NULL) {
577                 rc = ldlm_callback_reply(req, -ENOTCONN);
578                 ldlm_callback_errmsg(req, "Operate on unconnected server",
579                                      rc, NULL);
580                 return 0;
581         }
582
583         LASSERT(req->rq_export != NULL);
584         LASSERT(req->rq_export->exp_obd != NULL);
585
586         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
587         case LDLM_BL_CALLBACK:
588                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
589                         return 0;
590                 break;
591         case LDLM_CP_CALLBACK:
592                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
593                         return 0;
594                 break;
595         case LDLM_GL_CALLBACK:
596                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
597                         return 0;
598                 break;
599         case LDLM_SET_INFO:
600                 rc = ldlm_handle_setinfo(req);
601                 ldlm_callback_reply(req, rc);
602                 return 0;
603         case OBD_QC_CALLBACK:
604                 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
605                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
606                         return 0;
607                 rc = ldlm_handle_qc_callback(req);
608                 ldlm_callback_reply(req, rc);
609                 return 0;
610         default:
611                 CERROR("unknown opcode %u\n",
612                        lustre_msg_get_opc(req->rq_reqmsg));
613                 ldlm_callback_reply(req, -EPROTO);
614                 return 0;
615         }
616
617         ns = req->rq_export->exp_obd->obd_namespace;
618         LASSERT(ns != NULL);
619
620         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
621
622         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
623         if (dlm_req == NULL) {
624                 rc = ldlm_callback_reply(req, -EPROTO);
625                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
626                                      NULL);
627                 return 0;
628         }
629
630         /* Force a known safe race, send a cancel to the server for a lock
631          * which the server has already started a blocking callback on. */
632         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
633             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
634                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
635                 if (rc < 0)
636                         CERROR("ldlm_cli_cancel: %d\n", rc);
637         }
638
639         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
640         if (!lock) {
641                 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock disappeared\n",
642                        dlm_req->lock_handle[0].cookie);
643                 rc = ldlm_callback_reply(req, -EINVAL);
644                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
645                                      &dlm_req->lock_handle[0]);
646                 return 0;
647         }
648
649         if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
650             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
651                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
652
653         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
654         lock_res_and_lock(lock);
655         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
656                                               LDLM_AST_FLAGS);
657         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
658                 /* If somebody cancels lock and cache is already dropped,
659                  * or lock is failed before cp_ast received on client,
660                  * we can tell the server we have no lock. Otherwise, we
661                  * should send cancel after dropping the cache. */
662                 if (((lock->l_flags & LDLM_FL_CANCELING) &&
663                     (lock->l_flags & LDLM_FL_BL_DONE)) ||
664                     (lock->l_flags & LDLM_FL_FAILED)) {
665                         LDLM_DEBUG(lock, "callback on lock %#llx - lock disappeared\n",
666                                    dlm_req->lock_handle[0].cookie);
667                         unlock_res_and_lock(lock);
668                         LDLM_LOCK_RELEASE(lock);
669                         rc = ldlm_callback_reply(req, -EINVAL);
670                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
671                                              &dlm_req->lock_handle[0]);
672                         return 0;
673                 }
674                 /* BL_AST locks are not needed in LRU.
675                  * Let ldlm_cancel_lru() be fast. */
676                 ldlm_lock_remove_from_lru(lock);
677                 lock->l_flags |= LDLM_FL_BL_AST;
678         }
679         unlock_res_and_lock(lock);
680
681         /* We want the ost thread to get this reply so that it can respond
682          * to ost requests (write cache writeback) that might be triggered
683          * in the callback.
684          *
685          * But we'd also like to be able to indicate in the reply that we're
686          * cancelling right now, because it's unused, or have an intent result
687          * in the reply, so we might have to push the responsibility for sending
688          * the reply down into the AST handlers, alas. */
689
690         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
691         case LDLM_BL_CALLBACK:
692                 CDEBUG(D_INODE, "blocking ast\n");
693                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
694                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
695                         rc = ldlm_callback_reply(req, 0);
696                         if (req->rq_no_reply || rc)
697                                 ldlm_callback_errmsg(req, "Normal process", rc,
698                                                      &dlm_req->lock_handle[0]);
699                 }
700                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
701                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
702                 break;
703         case LDLM_CP_CALLBACK:
704                 CDEBUG(D_INODE, "completion ast\n");
705                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
706                 ldlm_callback_reply(req, 0);
707                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
708                 break;
709         case LDLM_GL_CALLBACK:
710                 CDEBUG(D_INODE, "glimpse ast\n");
711                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
712                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
713                 break;
714         default:
715                 LBUG();                  /* checked above */
716         }
717
718         return 0;
719 }
720
721
722 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
723 {
724         struct ldlm_bl_work_item *blwi = NULL;
725         static unsigned int num_bl;
726
727         spin_lock(&blp->blp_lock);
728         /* process a request from the blp_list at least every blp_num_threads */
729         if (!list_empty(&blp->blp_list) &&
730             (list_empty(&blp->blp_prio_list) || num_bl == 0))
731                 blwi = list_entry(blp->blp_list.next,
732                                       struct ldlm_bl_work_item, blwi_entry);
733         else
734                 if (!list_empty(&blp->blp_prio_list))
735                         blwi = list_entry(blp->blp_prio_list.next,
736                                               struct ldlm_bl_work_item,
737                                               blwi_entry);
738
739         if (blwi) {
740                 if (++num_bl >= atomic_read(&blp->blp_num_threads))
741                         num_bl = 0;
742                 list_del(&blwi->blwi_entry);
743         }
744         spin_unlock(&blp->blp_lock);
745
746         return blwi;
747 }
748
749 /* This only contains temporary data until the thread starts */
750 struct ldlm_bl_thread_data {
751         char                    bltd_name[CFS_CURPROC_COMM_MAX];
752         struct ldlm_bl_pool     *bltd_blp;
753         struct completion       bltd_comp;
754         int                     bltd_num;
755 };
756
757 static int ldlm_bl_thread_main(void *arg);
758
759 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
760 {
761         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
762         struct task_struct *task;
763
764         init_completion(&bltd.bltd_comp);
765         bltd.bltd_num = atomic_read(&blp->blp_num_threads);
766         snprintf(bltd.bltd_name, sizeof(bltd.bltd_name),
767                 "ldlm_bl_%02d", bltd.bltd_num);
768         task = kthread_run(ldlm_bl_thread_main, &bltd, "%s", bltd.bltd_name);
769         if (IS_ERR(task)) {
770                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
771                        atomic_read(&blp->blp_num_threads), PTR_ERR(task));
772                 return PTR_ERR(task);
773         }
774         wait_for_completion(&bltd.bltd_comp);
775
776         return 0;
777 }
778
779 /**
780  * Main blocking requests processing thread.
781  *
782  * Callers put locks into its queue by calling ldlm_bl_to_thread.
783  * This thread in the end ends up doing actual call to ->l_blocking_ast
784  * for queued locks.
785  */
786 static int ldlm_bl_thread_main(void *arg)
787 {
788         struct ldlm_bl_pool *blp;
789
790         {
791                 struct ldlm_bl_thread_data *bltd = arg;
792
793                 blp = bltd->bltd_blp;
794
795                 atomic_inc(&blp->blp_num_threads);
796                 atomic_inc(&blp->blp_busy_threads);
797
798                 complete(&bltd->bltd_comp);
799                 /* cannot use bltd after this, it is only on caller's stack */
800         }
801
802         while (1) {
803                 struct l_wait_info lwi = { 0 };
804                 struct ldlm_bl_work_item *blwi = NULL;
805                 int busy;
806
807                 blwi = ldlm_bl_get_work(blp);
808
809                 if (blwi == NULL) {
810                         atomic_dec(&blp->blp_busy_threads);
811                         l_wait_event_exclusive(blp->blp_waitq,
812                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
813                                          &lwi);
814                         busy = atomic_inc_return(&blp->blp_busy_threads);
815                 } else {
816                         busy = atomic_read(&blp->blp_busy_threads);
817                 }
818
819                 if (blwi->blwi_ns == NULL)
820                         /* added by ldlm_cleanup() */
821                         break;
822
823                 /* Not fatal if racy and have a few too many threads */
824                 if (unlikely(busy < blp->blp_max_threads &&
825                              busy >= atomic_read(&blp->blp_num_threads) &&
826                              !blwi->blwi_mem_pressure))
827                         /* discard the return value, we tried */
828                         ldlm_bl_thread_start(blp);
829
830                 if (blwi->blwi_mem_pressure)
831                         memory_pressure_set();
832
833                 if (blwi->blwi_count) {
834                         int count;
835                         /* The special case when we cancel locks in LRU
836                          * asynchronously, we pass the list of locks here.
837                          * Thus locks are marked LDLM_FL_CANCELING, but NOT
838                          * canceled locally yet. */
839                         count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
840                                                            blwi->blwi_count,
841                                                            LCF_BL_AST);
842                         ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
843                                              blwi->blwi_flags);
844                 } else {
845                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
846                                                 blwi->blwi_lock);
847                 }
848                 if (blwi->blwi_mem_pressure)
849                         memory_pressure_clr();
850
851                 if (blwi->blwi_flags & LCF_ASYNC)
852                         OBD_FREE(blwi, sizeof(*blwi));
853                 else
854                         complete(&blwi->blwi_comp);
855         }
856
857         atomic_dec(&blp->blp_busy_threads);
858         atomic_dec(&blp->blp_num_threads);
859         complete(&blp->blp_comp);
860         return 0;
861 }
862
863
864 static int ldlm_setup(void);
865 static int ldlm_cleanup(void);
866
867 int ldlm_get_ref(void)
868 {
869         int rc = 0;
870
871         mutex_lock(&ldlm_ref_mutex);
872         if (++ldlm_refcount == 1) {
873                 rc = ldlm_setup();
874                 if (rc)
875                         ldlm_refcount--;
876         }
877         mutex_unlock(&ldlm_ref_mutex);
878
879         return rc;
880 }
881 EXPORT_SYMBOL(ldlm_get_ref);
882
883 void ldlm_put_ref(void)
884 {
885         mutex_lock(&ldlm_ref_mutex);
886         if (ldlm_refcount == 1) {
887                 int rc = ldlm_cleanup();
888
889                 if (rc)
890                         CERROR("ldlm_cleanup failed: %d\n", rc);
891                 else
892                         ldlm_refcount--;
893         } else {
894                 ldlm_refcount--;
895         }
896         mutex_unlock(&ldlm_ref_mutex);
897 }
898 EXPORT_SYMBOL(ldlm_put_ref);
899
900 /*
901  * Export handle<->lock hash operations.
902  */
903 static unsigned
904 ldlm_export_lock_hash(struct cfs_hash *hs, const void *key, unsigned mask)
905 {
906         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
907 }
908
909 static void *
910 ldlm_export_lock_key(struct hlist_node *hnode)
911 {
912         struct ldlm_lock *lock;
913
914         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
915         return &lock->l_remote_handle;
916 }
917
918 static void
919 ldlm_export_lock_keycpy(struct hlist_node *hnode, void *key)
920 {
921         struct ldlm_lock     *lock;
922
923         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
924         lock->l_remote_handle = *(struct lustre_handle *)key;
925 }
926
927 static int
928 ldlm_export_lock_keycmp(const void *key, struct hlist_node *hnode)
929 {
930         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
931 }
932
933 static void *
934 ldlm_export_lock_object(struct hlist_node *hnode)
935 {
936         return hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
937 }
938
939 static void
940 ldlm_export_lock_get(struct cfs_hash *hs, struct hlist_node *hnode)
941 {
942         struct ldlm_lock *lock;
943
944         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
945         LDLM_LOCK_GET(lock);
946 }
947
948 static void
949 ldlm_export_lock_put(struct cfs_hash *hs, struct hlist_node *hnode)
950 {
951         struct ldlm_lock *lock;
952
953         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
954         LDLM_LOCK_RELEASE(lock);
955 }
956
957 static cfs_hash_ops_t ldlm_export_lock_ops = {
958         .hs_hash        = ldlm_export_lock_hash,
959         .hs_key  = ldlm_export_lock_key,
960         .hs_keycmp      = ldlm_export_lock_keycmp,
961         .hs_keycpy      = ldlm_export_lock_keycpy,
962         .hs_object      = ldlm_export_lock_object,
963         .hs_get  = ldlm_export_lock_get,
964         .hs_put  = ldlm_export_lock_put,
965         .hs_put_locked  = ldlm_export_lock_put,
966 };
967
968 int ldlm_init_export(struct obd_export *exp)
969 {
970         int rc;
971
972         exp->exp_lock_hash =
973                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
974                                 HASH_EXP_LOCK_CUR_BITS,
975                                 HASH_EXP_LOCK_MAX_BITS,
976                                 HASH_EXP_LOCK_BKT_BITS, 0,
977                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
978                                 &ldlm_export_lock_ops,
979                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
980                                 CFS_HASH_NBLK_CHANGE);
981
982         if (!exp->exp_lock_hash)
983                 return -ENOMEM;
984
985         rc = ldlm_init_flock_export(exp);
986         if (rc)
987                 goto err;
988
989         return 0;
990 err:
991         ldlm_destroy_export(exp);
992         return rc;
993 }
994 EXPORT_SYMBOL(ldlm_init_export);
995
996 void ldlm_destroy_export(struct obd_export *exp)
997 {
998         cfs_hash_putref(exp->exp_lock_hash);
999         exp->exp_lock_hash = NULL;
1000
1001         ldlm_destroy_flock_export(exp);
1002 }
1003 EXPORT_SYMBOL(ldlm_destroy_export);
1004
1005 static int ldlm_setup(void)
1006 {
1007         static struct ptlrpc_service_conf       conf;
1008         struct ldlm_bl_pool                     *blp = NULL;
1009         int rc = 0;
1010         int i;
1011
1012         if (ldlm_state != NULL)
1013                 return -EALREADY;
1014
1015         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1016         if (ldlm_state == NULL)
1017                 return -ENOMEM;
1018
1019         rc = ldlm_proc_setup();
1020         if (rc != 0)
1021                 goto out;
1022
1023         memset(&conf, 0, sizeof(conf));
1024         conf = (typeof(conf)) {
1025                 .psc_name               = "ldlm_cbd",
1026                 .psc_watchdog_factor    = 2,
1027                 .psc_buf                = {
1028                         .bc_nbufs               = LDLM_CLIENT_NBUFS,
1029                         .bc_buf_size            = LDLM_BUFSIZE,
1030                         .bc_req_max_size        = LDLM_MAXREQSIZE,
1031                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
1032                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
1033                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
1034                 },
1035                 .psc_thr                = {
1036                         .tc_thr_name            = "ldlm_cb",
1037                         .tc_thr_factor          = LDLM_THR_FACTOR,
1038                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
1039                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
1040                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
1041                         .tc_nthrs_user          = ldlm_num_threads,
1042                         .tc_cpu_affinity        = 1,
1043                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
1044                 },
1045                 .psc_cpt                = {
1046                         .cc_pattern             = ldlm_cpts,
1047                 },
1048                 .psc_ops                = {
1049                         .so_req_handler         = ldlm_callback_handler,
1050                 },
1051         };
1052         ldlm_state->ldlm_cb_service =
1053                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
1054         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
1055                 CERROR("failed to start service\n");
1056                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
1057                 ldlm_state->ldlm_cb_service = NULL;
1058                 goto out;
1059         }
1060
1061
1062         OBD_ALLOC(blp, sizeof(*blp));
1063         if (blp == NULL) {
1064                 rc = -ENOMEM;
1065                 goto out;
1066         }
1067         ldlm_state->ldlm_bl_pool = blp;
1068
1069         spin_lock_init(&blp->blp_lock);
1070         INIT_LIST_HEAD(&blp->blp_list);
1071         INIT_LIST_HEAD(&blp->blp_prio_list);
1072         init_waitqueue_head(&blp->blp_waitq);
1073         atomic_set(&blp->blp_num_threads, 0);
1074         atomic_set(&blp->blp_busy_threads, 0);
1075
1076         if (ldlm_num_threads == 0) {
1077                 blp->blp_min_threads = LDLM_NTHRS_INIT;
1078                 blp->blp_max_threads = LDLM_NTHRS_MAX;
1079         } else {
1080                 blp->blp_min_threads = blp->blp_max_threads =
1081                         min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
1082                                                          ldlm_num_threads));
1083         }
1084
1085         for (i = 0; i < blp->blp_min_threads; i++) {
1086                 rc = ldlm_bl_thread_start(blp);
1087                 if (rc < 0)
1088                         goto out;
1089         }
1090
1091
1092         rc = ldlm_pools_init();
1093         if (rc) {
1094                 CERROR("Failed to initialize LDLM pools: %d\n", rc);
1095                 goto out;
1096         }
1097         return 0;
1098
1099  out:
1100         ldlm_cleanup();
1101         return rc;
1102 }
1103
1104 static int ldlm_cleanup(void)
1105 {
1106         if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
1107             !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
1108                 CERROR("ldlm still has namespaces; clean these up first.\n");
1109                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
1110                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
1111                 return -EBUSY;
1112         }
1113
1114         ldlm_pools_fini();
1115
1116         if (ldlm_state->ldlm_bl_pool != NULL) {
1117                 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1118
1119                 while (atomic_read(&blp->blp_num_threads) > 0) {
1120                         struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1121
1122                         init_completion(&blp->blp_comp);
1123
1124                         spin_lock(&blp->blp_lock);
1125                         list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1126                         wake_up(&blp->blp_waitq);
1127                         spin_unlock(&blp->blp_lock);
1128
1129                         wait_for_completion(&blp->blp_comp);
1130                 }
1131
1132                 OBD_FREE(blp, sizeof(*blp));
1133         }
1134
1135         if (ldlm_state->ldlm_cb_service != NULL)
1136                 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1137
1138         ldlm_proc_cleanup();
1139
1140
1141         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1142         ldlm_state = NULL;
1143
1144         return 0;
1145 }
1146
1147 int ldlm_init(void)
1148 {
1149         mutex_init(&ldlm_ref_mutex);
1150         mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
1151         mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1152         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1153                                                sizeof(struct ldlm_resource), 0,
1154                                                SLAB_HWCACHE_ALIGN, NULL);
1155         if (ldlm_resource_slab == NULL)
1156                 return -ENOMEM;
1157
1158         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1159                               sizeof(struct ldlm_lock), 0,
1160                               SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU, NULL);
1161         if (ldlm_lock_slab == NULL) {
1162                 kmem_cache_destroy(ldlm_resource_slab);
1163                 return -ENOMEM;
1164         }
1165
1166         ldlm_interval_slab = kmem_cache_create("interval_node",
1167                                         sizeof(struct ldlm_interval),
1168                                         0, SLAB_HWCACHE_ALIGN, NULL);
1169         if (ldlm_interval_slab == NULL) {
1170                 kmem_cache_destroy(ldlm_resource_slab);
1171                 kmem_cache_destroy(ldlm_lock_slab);
1172                 return -ENOMEM;
1173         }
1174 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1175         class_export_dump_hook = ldlm_dump_export_locks;
1176 #endif
1177         return 0;
1178 }
1179
1180 void ldlm_exit(void)
1181 {
1182         if (ldlm_refcount)
1183                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1184         kmem_cache_destroy(ldlm_resource_slab);
1185         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
1186          * synchronize_rcu() to wait a grace period elapsed, so that
1187          * ldlm_lock_free() get a chance to be called. */
1188         synchronize_rcu();
1189         kmem_cache_destroy(ldlm_lock_slab);
1190         kmem_cache_destroy(ldlm_interval_slab);
1191 }