4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ldlm/ldlm_lockd.c
34 * Author: Peter Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_LDLM
40 #include "../../include/linux/libcfs/libcfs.h"
41 #include "../include/lustre_dlm.h"
42 #include "../include/obd_class.h"
43 #include <linux/list.h>
44 #include "ldlm_internal.h"
46 static int ldlm_num_threads;
47 module_param(ldlm_num_threads, int, 0444);
48 MODULE_PARM_DESC(ldlm_num_threads, "number of DLM service threads to start");
50 static char *ldlm_cpts;
51 module_param(ldlm_cpts, charp, 0444);
52 MODULE_PARM_DESC(ldlm_cpts, "CPU partitions ldlm threads should run on");
54 static struct mutex ldlm_ref_mutex;
55 static int ldlm_refcount;
57 static struct kobject *ldlm_kobj;
58 struct kset *ldlm_ns_kset;
59 static struct kset *ldlm_svc_kset;
61 struct ldlm_cb_async_args {
62 struct ldlm_cb_set_arg *ca_set_arg;
63 struct ldlm_lock *ca_lock;
68 static struct ldlm_state *ldlm_state;
72 #define ELT_TERMINATE 2
78 * blp_prio_list is used for callbacks that should be handled
79 * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
82 struct list_head blp_prio_list;
85 * blp_list is used for all other callbacks which are likely
86 * to take longer to process.
88 struct list_head blp_list;
90 wait_queue_head_t blp_waitq;
91 struct completion blp_comp;
92 atomic_t blp_num_threads;
93 atomic_t blp_busy_threads;
98 struct ldlm_bl_work_item {
99 struct list_head blwi_entry;
100 struct ldlm_namespace *blwi_ns;
101 struct ldlm_lock_desc blwi_ld;
102 struct ldlm_lock *blwi_lock;
103 struct list_head blwi_head;
105 struct completion blwi_comp;
106 enum ldlm_cancel_flags blwi_flags;
107 int blwi_mem_pressure;
111 * Callback handler for receiving incoming blocking ASTs.
113 * This can only happen on client side.
115 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
116 struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
120 LDLM_DEBUG(lock, "client blocking AST callback handler");
122 lock_res_and_lock(lock);
123 ldlm_set_cbpending(lock);
125 if (ldlm_is_cancel_on_block(lock))
126 ldlm_set_cancel(lock);
128 do_ast = !lock->l_readers && !lock->l_writers;
129 unlock_res_and_lock(lock);
133 "Lock %p already unused, calling callback (%p)\n", lock,
134 lock->l_blocking_ast);
135 if (lock->l_blocking_ast)
136 lock->l_blocking_ast(lock, ld, lock->l_ast_data,
140 "Lock %p is referenced, will be cancelled later\n",
144 LDLM_DEBUG(lock, "client blocking callback handler END");
145 LDLM_LOCK_RELEASE(lock);
149 * Callback handler for receiving incoming completion ASTs.
151 * This only can happen on client side.
153 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
154 struct ldlm_namespace *ns,
155 struct ldlm_request *dlm_req,
156 struct ldlm_lock *lock)
162 LDLM_DEBUG(lock, "client completion callback handler START");
164 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
165 int to = cfs_time_seconds(1);
168 set_current_state(TASK_INTERRUPTIBLE);
169 schedule_timeout(to);
170 if (lock->l_granted_mode == lock->l_req_mode ||
171 ldlm_is_destroyed(lock))
176 lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
178 LDLM_ERROR(lock, "Fail to get lvb_len, rc = %d", lvb_len);
181 } else if (lvb_len > 0) {
182 if (lock->l_lvb_len > 0) {
183 /* for extent lock, lvb contains ost_lvb{}. */
184 LASSERT(lock->l_lvb_data);
186 if (unlikely(lock->l_lvb_len < lvb_len)) {
187 LDLM_ERROR(lock, "Replied LVB is larger than expectation, expected = %d, replied = %d",
188 lock->l_lvb_len, lvb_len);
192 } else if (ldlm_has_layout(lock)) { /* for layout lock, lvb has
197 lvb_data = kzalloc(lvb_len, GFP_NOFS);
199 LDLM_ERROR(lock, "No memory: %d.\n", lvb_len);
204 lock_res_and_lock(lock);
205 LASSERT(!lock->l_lvb_data);
206 lock->l_lvb_type = LVB_T_LAYOUT;
207 lock->l_lvb_data = lvb_data;
208 lock->l_lvb_len = lvb_len;
209 unlock_res_and_lock(lock);
213 lock_res_and_lock(lock);
214 if (ldlm_is_destroyed(lock) ||
215 lock->l_granted_mode == lock->l_req_mode) {
216 /* bug 11300: the lock has already been granted */
217 unlock_res_and_lock(lock);
218 LDLM_DEBUG(lock, "Double grant race happened");
223 /* If we receive the completion AST before the actual enqueue returned,
224 * then we might need to switch lock modes, resources, or extents.
226 if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
227 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
228 LDLM_DEBUG(lock, "completion AST, new lock mode");
231 if (lock->l_resource->lr_type != LDLM_PLAIN) {
232 ldlm_convert_policy_to_local(req->rq_export,
233 dlm_req->lock_desc.l_resource.lr_type,
234 &dlm_req->lock_desc.l_policy_data,
235 &lock->l_policy_data);
236 LDLM_DEBUG(lock, "completion AST, new policy data");
239 ldlm_resource_unlink_lock(lock);
240 if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
241 &lock->l_resource->lr_name,
242 sizeof(lock->l_resource->lr_name)) != 0) {
243 unlock_res_and_lock(lock);
244 rc = ldlm_lock_change_resource(ns, lock,
245 &dlm_req->lock_desc.l_resource.lr_name);
247 LDLM_ERROR(lock, "Failed to allocate resource");
250 LDLM_DEBUG(lock, "completion AST, new resource");
251 CERROR("change resource!\n");
252 lock_res_and_lock(lock);
255 if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
256 /* BL_AST locks are not needed in LRU.
257 * Let ldlm_cancel_lru() be fast.
259 ldlm_lock_remove_from_lru(lock);
260 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
261 LDLM_DEBUG(lock, "completion AST includes blocking AST");
264 if (lock->l_lvb_len > 0) {
265 rc = ldlm_fill_lvb(lock, &req->rq_pill, RCL_CLIENT,
266 lock->l_lvb_data, lvb_len);
268 unlock_res_and_lock(lock);
273 ldlm_grant_lock(lock, &ast_list);
274 unlock_res_and_lock(lock);
276 LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
278 /* Let Enqueue to call osc_lock_upcall() and initialize l_ast_data */
279 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
281 ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
283 LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
289 lock_res_and_lock(lock);
290 ldlm_set_failed(lock);
291 unlock_res_and_lock(lock);
292 wake_up(&lock->l_waitq);
294 LDLM_LOCK_RELEASE(lock);
298 * Callback handler for receiving incoming glimpse ASTs.
300 * This only can happen on client side. After handling the glimpse AST
301 * we also consider dropping the lock here if it is unused locally for a
304 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
305 struct ldlm_namespace *ns,
306 struct ldlm_request *dlm_req,
307 struct ldlm_lock *lock)
311 LDLM_DEBUG(lock, "client glimpse AST callback handler");
313 if (lock->l_glimpse_ast)
314 rc = lock->l_glimpse_ast(lock, req);
316 if (req->rq_repmsg) {
323 lock_res_and_lock(lock);
324 if (lock->l_granted_mode == LCK_PW &&
325 !lock->l_readers && !lock->l_writers &&
326 cfs_time_after(cfs_time_current(),
327 cfs_time_add(lock->l_last_used,
328 cfs_time_seconds(10)))) {
329 unlock_res_and_lock(lock);
330 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
331 ldlm_handle_bl_callback(ns, NULL, lock);
335 unlock_res_and_lock(lock);
336 LDLM_LOCK_RELEASE(lock);
339 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
341 if (req->rq_no_reply)
345 if (!req->rq_packed_final) {
346 rc = lustre_pack_reply(req, 1, NULL, NULL);
350 return ptlrpc_reply(req);
353 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi,
354 enum ldlm_cancel_flags cancel_flags)
356 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
358 spin_lock(&blp->blp_lock);
359 if (blwi->blwi_lock && ldlm_is_discard_data(blwi->blwi_lock)) {
360 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
361 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
363 /* other blocking callbacks are added to the regular list */
364 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
366 spin_unlock(&blp->blp_lock);
368 wake_up(&blp->blp_waitq);
370 /* can not check blwi->blwi_flags as blwi could be already freed in
373 if (!(cancel_flags & LCF_ASYNC))
374 wait_for_completion(&blwi->blwi_comp);
379 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
380 struct ldlm_namespace *ns,
381 struct ldlm_lock_desc *ld,
382 struct list_head *cancels, int count,
383 struct ldlm_lock *lock,
384 enum ldlm_cancel_flags cancel_flags)
386 init_completion(&blwi->blwi_comp);
387 INIT_LIST_HEAD(&blwi->blwi_head);
389 if (memory_pressure_get())
390 blwi->blwi_mem_pressure = 1;
393 blwi->blwi_flags = cancel_flags;
397 list_add(&blwi->blwi_head, cancels);
398 list_del_init(cancels);
399 blwi->blwi_count = count;
401 blwi->blwi_lock = lock;
406 * Queues a list of locks \a cancels containing \a count locks
407 * for later processing by a blocking thread. If \a count is zero,
408 * then the lock referenced as \a lock is queued instead.
410 * The blocking thread would then call ->l_blocking_ast callback in the lock.
411 * If list addition fails an error is returned and caller is supposed to
412 * call ->l_blocking_ast itself.
414 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
415 struct ldlm_lock_desc *ld,
416 struct ldlm_lock *lock,
417 struct list_head *cancels, int count,
418 enum ldlm_cancel_flags cancel_flags)
420 if (cancels && count == 0)
423 if (cancel_flags & LCF_ASYNC) {
424 struct ldlm_bl_work_item *blwi;
426 blwi = kzalloc(sizeof(*blwi), GFP_NOFS);
429 init_blwi(blwi, ns, ld, cancels, count, lock, cancel_flags);
431 return __ldlm_bl_to_thread(blwi, cancel_flags);
433 /* if it is synchronous call do minimum mem alloc, as it could
434 * be triggered from kernel shrinker
436 struct ldlm_bl_work_item blwi;
438 memset(&blwi, 0, sizeof(blwi));
439 init_blwi(&blwi, ns, ld, cancels, count, lock, cancel_flags);
440 return __ldlm_bl_to_thread(&blwi, cancel_flags);
444 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
445 struct ldlm_lock *lock)
447 return ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LCF_ASYNC);
450 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
451 struct list_head *cancels, int count,
452 enum ldlm_cancel_flags cancel_flags)
454 return ldlm_bl_to_thread(ns, ld, NULL, cancels, count, cancel_flags);
457 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
458 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
460 struct obd_device *obd = req->rq_export->exp_obd;
466 DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
468 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
470 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
472 DEBUG_REQ(D_IOCTL, req, "no set_info key");
475 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
477 val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
479 DEBUG_REQ(D_IOCTL, req, "no set_info val");
482 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
485 /* We are responsible for swabbing contents of val */
487 if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
488 /* Pass it on to mdc (the "export" in this case) */
489 rc = obd_set_info_async(req->rq_svc_thread->t_env,
491 sizeof(KEY_HSM_COPYTOOL_SEND),
492 KEY_HSM_COPYTOOL_SEND,
495 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
500 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
501 const char *msg, int rc,
502 const struct lustre_handle *handle)
504 DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
505 "%s: [nid %s] [rc %d] [lock %#llx]",
506 msg, libcfs_id2str(req->rq_peer), rc,
507 handle ? handle->cookie : 0);
508 if (req->rq_no_reply)
509 CWARN("No reply was sent, maybe cause bug 21636.\n");
511 CWARN("Send reply failed, maybe cause bug 21636.\n");
514 static int ldlm_handle_qc_callback(struct ptlrpc_request *req)
516 struct obd_quotactl *oqctl;
517 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
519 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
521 CERROR("Can't unpack obd_quotactl\n");
525 oqctl->qc_stat = ptlrpc_status_ntoh(oqctl->qc_stat);
527 cli->cl_qchk_stat = oqctl->qc_stat;
531 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
532 static int ldlm_callback_handler(struct ptlrpc_request *req)
534 struct ldlm_namespace *ns;
535 struct ldlm_request *dlm_req;
536 struct ldlm_lock *lock;
539 /* Requests arrive in sender's byte order. The ptlrpc service
540 * handler has already checked and, if necessary, byte-swapped the
541 * incoming request message body, but I am responsible for the
545 /* do nothing for sec context finalize */
546 if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
549 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
551 if (!req->rq_export) {
552 rc = ldlm_callback_reply(req, -ENOTCONN);
553 ldlm_callback_errmsg(req, "Operate on unconnected server",
558 LASSERT(req->rq_export->exp_obd);
560 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
561 case LDLM_BL_CALLBACK:
562 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET)) {
564 ldlm_callback_reply(req, -(int)cfs_fail_err);
568 case LDLM_CP_CALLBACK:
569 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
572 case LDLM_GL_CALLBACK:
573 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
577 rc = ldlm_handle_setinfo(req);
578 ldlm_callback_reply(req, rc);
580 case OBD_QC_CALLBACK:
581 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
582 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
584 rc = ldlm_handle_qc_callback(req);
585 ldlm_callback_reply(req, rc);
588 CERROR("unknown opcode %u\n",
589 lustre_msg_get_opc(req->rq_reqmsg));
590 ldlm_callback_reply(req, -EPROTO);
594 ns = req->rq_export->exp_obd->obd_namespace;
597 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
599 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
601 rc = ldlm_callback_reply(req, -EPROTO);
602 ldlm_callback_errmsg(req, "Operate without parameter", rc,
607 /* Force a known safe race, send a cancel to the server for a lock
608 * which the server has already started a blocking callback on.
610 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
611 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
612 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0], 0);
614 CERROR("ldlm_cli_cancel: %d\n", rc);
617 lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
619 CDEBUG(D_DLMTRACE, "callback on lock %#llx - lock disappeared\n",
620 dlm_req->lock_handle[0].cookie);
621 rc = ldlm_callback_reply(req, -EINVAL);
622 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
623 &dlm_req->lock_handle[0]);
627 if (ldlm_is_fail_loc(lock) &&
628 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
629 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
631 /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
632 lock_res_and_lock(lock);
633 lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
635 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
636 /* If somebody cancels lock and cache is already dropped,
637 * or lock is failed before cp_ast received on client,
638 * we can tell the server we have no lock. Otherwise, we
639 * should send cancel after dropping the cache.
641 if ((ldlm_is_canceling(lock) && ldlm_is_bl_done(lock)) ||
642 ldlm_is_failed(lock)) {
644 "callback on lock %#llx - lock disappeared",
645 dlm_req->lock_handle[0].cookie);
646 unlock_res_and_lock(lock);
647 LDLM_LOCK_RELEASE(lock);
648 rc = ldlm_callback_reply(req, -EINVAL);
649 ldlm_callback_errmsg(req, "Operate on stale lock", rc,
650 &dlm_req->lock_handle[0]);
653 /* BL_AST locks are not needed in LRU.
654 * Let ldlm_cancel_lru() be fast.
656 ldlm_lock_remove_from_lru(lock);
657 ldlm_set_bl_ast(lock);
659 unlock_res_and_lock(lock);
661 /* We want the ost thread to get this reply so that it can respond
662 * to ost requests (write cache writeback) that might be triggered
665 * But we'd also like to be able to indicate in the reply that we're
666 * cancelling right now, because it's unused, or have an intent result
667 * in the reply, so we might have to push the responsibility for sending
668 * the reply down into the AST handlers, alas.
671 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
672 case LDLM_BL_CALLBACK:
673 CDEBUG(D_INODE, "blocking ast\n");
674 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
675 if (!ldlm_is_cancel_on_block(lock)) {
676 rc = ldlm_callback_reply(req, 0);
677 if (req->rq_no_reply || rc)
678 ldlm_callback_errmsg(req, "Normal process", rc,
679 &dlm_req->lock_handle[0]);
681 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
682 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
684 case LDLM_CP_CALLBACK:
685 CDEBUG(D_INODE, "completion ast\n");
686 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
687 ldlm_callback_reply(req, 0);
688 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
690 case LDLM_GL_CALLBACK:
691 CDEBUG(D_INODE, "glimpse ast\n");
692 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
693 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
696 LBUG(); /* checked above */
702 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
704 struct ldlm_bl_work_item *blwi = NULL;
705 static unsigned int num_bl;
707 spin_lock(&blp->blp_lock);
708 /* process a request from the blp_list at least every blp_num_threads */
709 if (!list_empty(&blp->blp_list) &&
710 (list_empty(&blp->blp_prio_list) || num_bl == 0))
711 blwi = list_entry(blp->blp_list.next,
712 struct ldlm_bl_work_item, blwi_entry);
714 if (!list_empty(&blp->blp_prio_list))
715 blwi = list_entry(blp->blp_prio_list.next,
716 struct ldlm_bl_work_item,
720 if (++num_bl >= atomic_read(&blp->blp_num_threads))
722 list_del(&blwi->blwi_entry);
724 spin_unlock(&blp->blp_lock);
729 /* This only contains temporary data until the thread starts */
730 struct ldlm_bl_thread_data {
731 char bltd_name[CFS_CURPROC_COMM_MAX];
732 struct ldlm_bl_pool *bltd_blp;
733 struct completion bltd_comp;
737 static int ldlm_bl_thread_main(void *arg);
739 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
741 struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
742 struct task_struct *task;
744 init_completion(&bltd.bltd_comp);
745 bltd.bltd_num = atomic_read(&blp->blp_num_threads);
746 snprintf(bltd.bltd_name, sizeof(bltd.bltd_name),
747 "ldlm_bl_%02d", bltd.bltd_num);
748 task = kthread_run(ldlm_bl_thread_main, &bltd, "%s", bltd.bltd_name);
750 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %ld\n",
751 atomic_read(&blp->blp_num_threads), PTR_ERR(task));
752 return PTR_ERR(task);
754 wait_for_completion(&bltd.bltd_comp);
760 * Main blocking requests processing thread.
762 * Callers put locks into its queue by calling ldlm_bl_to_thread.
763 * This thread in the end ends up doing actual call to ->l_blocking_ast
766 static int ldlm_bl_thread_main(void *arg)
768 struct ldlm_bl_pool *blp;
771 struct ldlm_bl_thread_data *bltd = arg;
773 blp = bltd->bltd_blp;
775 atomic_inc(&blp->blp_num_threads);
776 atomic_inc(&blp->blp_busy_threads);
778 complete(&bltd->bltd_comp);
779 /* cannot use bltd after this, it is only on caller's stack */
783 struct l_wait_info lwi = { 0 };
784 struct ldlm_bl_work_item *blwi = NULL;
787 blwi = ldlm_bl_get_work(blp);
790 atomic_dec(&blp->blp_busy_threads);
791 l_wait_event_exclusive(blp->blp_waitq,
792 (blwi = ldlm_bl_get_work(blp)),
794 busy = atomic_inc_return(&blp->blp_busy_threads);
796 busy = atomic_read(&blp->blp_busy_threads);
800 /* added by ldlm_cleanup() */
803 /* Not fatal if racy and have a few too many threads */
804 if (unlikely(busy < blp->blp_max_threads &&
805 busy >= atomic_read(&blp->blp_num_threads) &&
806 !blwi->blwi_mem_pressure))
807 /* discard the return value, we tried */
808 ldlm_bl_thread_start(blp);
810 if (blwi->blwi_mem_pressure)
811 memory_pressure_set();
813 if (blwi->blwi_count) {
815 /* The special case when we cancel locks in LRU
816 * asynchronously, we pass the list of locks here.
817 * Thus locks are marked LDLM_FL_CANCELING, but NOT
818 * canceled locally yet.
820 count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
823 ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL,
826 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
829 if (blwi->blwi_mem_pressure)
830 memory_pressure_clr();
832 if (blwi->blwi_flags & LCF_ASYNC)
835 complete(&blwi->blwi_comp);
838 atomic_dec(&blp->blp_busy_threads);
839 atomic_dec(&blp->blp_num_threads);
840 complete(&blp->blp_comp);
844 static int ldlm_setup(void);
845 static int ldlm_cleanup(void);
847 int ldlm_get_ref(void)
851 mutex_lock(&ldlm_ref_mutex);
852 if (++ldlm_refcount == 1) {
857 mutex_unlock(&ldlm_ref_mutex);
861 EXPORT_SYMBOL(ldlm_get_ref);
863 void ldlm_put_ref(void)
865 mutex_lock(&ldlm_ref_mutex);
866 if (ldlm_refcount == 1) {
867 int rc = ldlm_cleanup();
870 CERROR("ldlm_cleanup failed: %d\n", rc);
876 mutex_unlock(&ldlm_ref_mutex);
878 EXPORT_SYMBOL(ldlm_put_ref);
880 static ssize_t cancel_unused_locks_before_replay_show(struct kobject *kobj,
881 struct attribute *attr,
884 return sprintf(buf, "%d\n", ldlm_cancel_unused_locks_before_replay);
887 static ssize_t cancel_unused_locks_before_replay_store(struct kobject *kobj,
888 struct attribute *attr,
895 rc = kstrtoul(buffer, 10, &val);
899 ldlm_cancel_unused_locks_before_replay = val;
903 LUSTRE_RW_ATTR(cancel_unused_locks_before_replay);
905 /* These are for root of /sys/fs/lustre/ldlm */
906 static struct attribute *ldlm_attrs[] = {
907 &lustre_attr_cancel_unused_locks_before_replay.attr,
911 static struct attribute_group ldlm_attr_group = {
915 static int ldlm_setup(void)
917 static struct ptlrpc_service_conf conf;
918 struct ldlm_bl_pool *blp = NULL;
925 ldlm_state = kzalloc(sizeof(*ldlm_state), GFP_NOFS);
929 ldlm_kobj = kobject_create_and_add("ldlm", lustre_kobj);
935 rc = sysfs_create_group(ldlm_kobj, &ldlm_attr_group);
939 ldlm_ns_kset = kset_create_and_add("namespaces", NULL, ldlm_kobj);
945 ldlm_svc_kset = kset_create_and_add("services", NULL, ldlm_kobj);
946 if (!ldlm_svc_kset) {
951 rc = ldlm_debugfs_setup();
955 memset(&conf, 0, sizeof(conf));
956 conf = (typeof(conf)) {
957 .psc_name = "ldlm_cbd",
958 .psc_watchdog_factor = 2,
960 .bc_nbufs = LDLM_CLIENT_NBUFS,
961 .bc_buf_size = LDLM_BUFSIZE,
962 .bc_req_max_size = LDLM_MAXREQSIZE,
963 .bc_rep_max_size = LDLM_MAXREPSIZE,
964 .bc_req_portal = LDLM_CB_REQUEST_PORTAL,
965 .bc_rep_portal = LDLM_CB_REPLY_PORTAL,
968 .tc_thr_name = "ldlm_cb",
969 .tc_thr_factor = LDLM_THR_FACTOR,
970 .tc_nthrs_init = LDLM_NTHRS_INIT,
971 .tc_nthrs_base = LDLM_NTHRS_BASE,
972 .tc_nthrs_max = LDLM_NTHRS_MAX,
973 .tc_nthrs_user = ldlm_num_threads,
974 .tc_cpu_affinity = 1,
975 .tc_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
978 .cc_pattern = ldlm_cpts,
981 .so_req_handler = ldlm_callback_handler,
984 ldlm_state->ldlm_cb_service =
985 ptlrpc_register_service(&conf, ldlm_svc_kset,
986 ldlm_svc_debugfs_dir);
987 if (IS_ERR(ldlm_state->ldlm_cb_service)) {
988 CERROR("failed to start service\n");
989 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
990 ldlm_state->ldlm_cb_service = NULL;
994 blp = kzalloc(sizeof(*blp), GFP_NOFS);
999 ldlm_state->ldlm_bl_pool = blp;
1001 spin_lock_init(&blp->blp_lock);
1002 INIT_LIST_HEAD(&blp->blp_list);
1003 INIT_LIST_HEAD(&blp->blp_prio_list);
1004 init_waitqueue_head(&blp->blp_waitq);
1005 atomic_set(&blp->blp_num_threads, 0);
1006 atomic_set(&blp->blp_busy_threads, 0);
1008 if (ldlm_num_threads == 0) {
1009 blp->blp_min_threads = LDLM_NTHRS_INIT;
1010 blp->blp_max_threads = LDLM_NTHRS_MAX;
1012 blp->blp_min_threads = min_t(int, LDLM_NTHRS_MAX,
1013 max_t(int, LDLM_NTHRS_INIT,
1016 blp->blp_max_threads = blp->blp_min_threads;
1019 for (i = 0; i < blp->blp_min_threads; i++) {
1020 rc = ldlm_bl_thread_start(blp);
1025 rc = ldlm_pools_init();
1027 CERROR("Failed to initialize LDLM pools: %d\n", rc);
1037 static int ldlm_cleanup(void)
1039 if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
1040 !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
1041 CERROR("ldlm still has namespaces; clean these up first.\n");
1042 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
1043 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
1049 if (ldlm_state->ldlm_bl_pool) {
1050 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1052 while (atomic_read(&blp->blp_num_threads) > 0) {
1053 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1055 init_completion(&blp->blp_comp);
1057 spin_lock(&blp->blp_lock);
1058 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1059 wake_up(&blp->blp_waitq);
1060 spin_unlock(&blp->blp_lock);
1062 wait_for_completion(&blp->blp_comp);
1068 if (ldlm_state->ldlm_cb_service)
1069 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1072 kset_unregister(ldlm_ns_kset);
1074 kset_unregister(ldlm_svc_kset);
1076 kobject_put(ldlm_kobj);
1078 ldlm_debugfs_cleanup();
1088 mutex_init(&ldlm_ref_mutex);
1089 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
1090 mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1091 ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1092 sizeof(struct ldlm_resource), 0,
1093 SLAB_HWCACHE_ALIGN, NULL);
1094 if (!ldlm_resource_slab)
1097 ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1098 sizeof(struct ldlm_lock), 0,
1099 SLAB_HWCACHE_ALIGN |
1100 SLAB_DESTROY_BY_RCU, NULL);
1101 if (!ldlm_lock_slab) {
1102 kmem_cache_destroy(ldlm_resource_slab);
1106 ldlm_interval_slab = kmem_cache_create("interval_node",
1107 sizeof(struct ldlm_interval),
1108 0, SLAB_HWCACHE_ALIGN, NULL);
1109 if (!ldlm_interval_slab) {
1110 kmem_cache_destroy(ldlm_resource_slab);
1111 kmem_cache_destroy(ldlm_lock_slab);
1114 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1115 class_export_dump_hook = ldlm_dump_export_locks;
1120 void ldlm_exit(void)
1123 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1124 kmem_cache_destroy(ldlm_resource_slab);
1125 /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
1126 * synchronize_rcu() to wait a grace period elapsed, so that
1127 * ldlm_lock_free() get a chance to be called.
1130 kmem_cache_destroy(ldlm_lock_slab);
1131 kmem_cache_destroy(ldlm_interval_slab);