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.sun.com/software/products/lustre/docs/GPLv2.pdf
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
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 /** \defgroup LDLM Lustre Distributed Lock Manager
39 * Lustre DLM is based on VAX DLM.
40 * Its two main roles are:
41 * - To provide locking assuring consistency of data on all Lustre nodes.
42 * - To allow clients to cache state protected by a lock by holding the
43 * lock until a conflicting lock is requested or it is expired by the LRU.
48 #ifndef _LUSTRE_DLM_H__
49 #define _LUSTRE_DLM_H__
51 #include "lustre_lib.h"
52 #include "lustre_net.h"
53 #include "lustre_import.h"
54 #include "lustre_handles.h"
55 #include "interval_tree.h" /* for interval_node{}, ldlm_extent */
58 #include "lustre_dlm_flags.h"
63 extern struct kset *ldlm_ns_kset;
64 extern struct kset *ldlm_svc_kset;
66 #define OBD_LDLM_DEVICENAME "ldlm"
68 #define LDLM_DEFAULT_LRU_SIZE (100 * num_online_cpus())
69 #define LDLM_DEFAULT_MAX_ALIVE (cfs_time_seconds(36000))
70 #define LDLM_DEFAULT_PARALLEL_AST_LIMIT 1024
73 * LDLM non-error return states
78 ELDLM_LOCK_CHANGED = 300,
79 ELDLM_LOCK_ABORTED = 301,
80 ELDLM_LOCK_REPLACED = 302,
81 ELDLM_NO_LOCK_DATA = 303,
82 ELDLM_LOCK_WOULDBLOCK = 304,
84 ELDLM_NAMESPACE_EXISTS = 400,
85 ELDLM_BAD_NAMESPACE = 401
89 * LDLM namespace type.
90 * The "client" type is actually an indication that this is a narrow local view
91 * into complete namespace on the server. Such namespaces cannot make any
92 * decisions about lack of conflicts or do any autonomous lock granting without
93 * first speaking to a server.
96 LDLM_NAMESPACE_SERVER = 1 << 0,
97 LDLM_NAMESPACE_CLIENT = 1 << 1
101 * The blocking callback is overloaded to perform two functions. These flags
102 * indicate which operation should be performed.
104 #define LDLM_CB_BLOCKING 1
105 #define LDLM_CB_CANCELING 2
108 * \name Lock Compatibility Matrix.
110 * A lock has both a type (extent, flock, inode bits, or plain) and a mode.
111 * Lock types are described in their respective implementation files:
112 * ldlm_{extent,flock,inodebits,plain}.c.
114 * There are six lock modes along with a compatibility matrix to indicate if
115 * two locks are compatible.
117 * - EX: Exclusive mode. Before a new file is created, MDS requests EX lock
119 * - PW: Protective Write (normal write) mode. When a client requests a write
120 * lock from an OST, a lock with PW mode will be issued.
121 * - PR: Protective Read (normal read) mode. When a client requests a read from
122 * an OST, a lock with PR mode will be issued. Also, if the client opens a
123 * file for execution, it is granted a lock with PR mode.
124 * - CW: Concurrent Write mode. The type of lock that the MDS grants if a client
125 * requests a write lock during a file open operation.
126 * - CR Concurrent Read mode. When a client performs a path lookup, MDS grants
127 * an inodebit lock with the CR mode on the intermediate path component.
141 #define LCK_COMPAT_EX LCK_NL
142 #define LCK_COMPAT_PW (LCK_COMPAT_EX | LCK_CR)
143 #define LCK_COMPAT_PR (LCK_COMPAT_PW | LCK_PR)
144 #define LCK_COMPAT_CW (LCK_COMPAT_PW | LCK_CW)
145 #define LCK_COMPAT_CR (LCK_COMPAT_CW | LCK_PR | LCK_PW)
146 #define LCK_COMPAT_NL (LCK_COMPAT_CR | LCK_EX | LCK_GROUP)
147 #define LCK_COMPAT_GROUP (LCK_GROUP | LCK_NL)
148 #define LCK_COMPAT_COS (LCK_COS)
149 /** @} Lock Compatibility Matrix */
151 extern ldlm_mode_t lck_compat_array[];
153 static inline void lockmode_verify(ldlm_mode_t mode)
155 LASSERT(mode > LCK_MINMODE && mode < LCK_MAXMODE);
158 static inline int lockmode_compat(ldlm_mode_t exist_mode, ldlm_mode_t new_mode)
160 return (lck_compat_array[exist_mode] & new_mode);
165 * cluster name spaces
169 #define DLM_OST_NAMESPACE 1
170 #define DLM_MDS_NAMESPACE 2
173 - do we just separate this by security domains and use a prefix for
174 multiple namespaces in the same domain?
179 * Locking rules for LDLM:
184 * waiting_locks_spinlock
199 struct ldlm_resource;
200 struct ldlm_namespace;
203 * Operations on LDLM pools.
204 * LDLM pool is a pool of locks in the namespace without any implicitly
206 * Locks in the pool are organized in LRU.
207 * Local memory pressure or server instructions (e.g. mempressure on server)
208 * can trigger freeing of locks from the pool
210 struct ldlm_pool_ops {
211 /** Recalculate pool \a pl usage */
212 int (*po_recalc)(struct ldlm_pool *pl);
213 /** Cancel at least \a nr locks from pool \a pl */
214 int (*po_shrink)(struct ldlm_pool *pl, int nr,
216 int (*po_setup)(struct ldlm_pool *pl, int limit);
219 /** One second for pools thread check interval. Each pool has own period. */
220 #define LDLM_POOLS_THREAD_PERIOD (1)
222 /** ~6% margin for modest pools. See ldlm_pool.c for details. */
223 #define LDLM_POOLS_MODEST_MARGIN_SHIFT (4)
225 /** Default recalc period for server side pools in sec. */
226 #define LDLM_POOL_SRV_DEF_RECALC_PERIOD (1)
228 /** Default recalc period for client side pools in sec. */
229 #define LDLM_POOL_CLI_DEF_RECALC_PERIOD (10)
232 * LDLM pool structure to track granted locks.
233 * For purposes of determining when to release locks on e.g. memory pressure.
234 * This feature is commonly referred to as lru_resize.
237 /** Pool debugfs directory. */
238 struct dentry *pl_debugfs_entry;
239 /** Pool name, must be long enough to hold compound proc entry name. */
241 /** Lock for protecting SLV/CLV updates. */
243 /** Number of allowed locks in in pool, both, client and server side. */
245 /** Number of granted locks in */
247 /** Grant rate per T. */
248 atomic_t pl_grant_rate;
249 /** Cancel rate per T. */
250 atomic_t pl_cancel_rate;
251 /** Server lock volume (SLV). Protected by pl_lock. */
252 __u64 pl_server_lock_volume;
253 /** Current biggest client lock volume. Protected by pl_lock. */
254 __u64 pl_client_lock_volume;
255 /** Lock volume factor. SLV on client is calculated as following:
256 * server_slv * lock_volume_factor. */
257 atomic_t pl_lock_volume_factor;
258 /** Time when last SLV from server was obtained. */
259 time64_t pl_recalc_time;
260 /** Recalculation period for pool. */
261 time64_t pl_recalc_period;
262 /** Recalculation and shrink operations. */
263 const struct ldlm_pool_ops *pl_ops;
264 /** Number of planned locks for next period. */
266 /** Pool statistics. */
267 struct lprocfs_stats *pl_stats;
270 struct kobject pl_kobj;
271 struct completion pl_kobj_unregister;
274 typedef int (*ldlm_cancel_for_recovery)(struct ldlm_lock *lock);
278 * LVB is Lock Value Block. This is a special opaque (to LDLM) value that could
279 * be associated with an LDLM lock and transferred from client to server and
282 * Currently LVBs are used by:
283 * - OSC-OST code to maintain current object size/times
284 * - layout lock code to return the layout when the layout lock is granted
286 struct ldlm_valblock_ops {
287 int (*lvbo_init)(struct ldlm_resource *res);
288 int (*lvbo_update)(struct ldlm_resource *res,
289 struct ptlrpc_request *r,
291 int (*lvbo_free)(struct ldlm_resource *res);
292 /* Return size of lvb data appropriate RPC size can be reserved */
293 int (*lvbo_size)(struct ldlm_lock *lock);
294 /* Called to fill in lvb data to RPC buffer @buf */
295 int (*lvbo_fill)(struct ldlm_lock *lock, void *buf, int buflen);
299 * LDLM pools related, type of lock pool in the namespace.
300 * Greedy means release cached locks aggressively
303 LDLM_NAMESPACE_GREEDY = 1 << 0,
304 LDLM_NAMESPACE_MODEST = 1 << 1
307 struct ldlm_ns_bucket {
308 /** back pointer to namespace */
309 struct ldlm_namespace *nsb_namespace;
311 * Estimated lock callback time. Used by adaptive timeout code to
312 * avoid spurious client evictions due to unresponsiveness when in
313 * fact the network or overall system load is at fault
315 struct adaptive_timeout nsb_at_estimate;
319 /** LDLM namespace lock stats */
326 LDLM_NS_TYPE_UNKNOWN = 0,
344 * Namespace serves to contain locks related to a particular service.
345 * There are two kinds of namespaces:
346 * - Server namespace has knowledge of all locks and is therefore authoritative
347 * to make decisions like what locks could be granted and what conflicts
348 * exist during new lock enqueue.
349 * - Client namespace only has limited knowledge about locks in the namespace,
350 * only seeing locks held by the client.
352 * Every Lustre service has one server namespace present on the server serving
353 * that service. Every client connected to the service has a client namespace
355 * Every lock obtained by client in that namespace is actually represented by
356 * two in-memory locks. One on the server and one on the client. The locks are
357 * linked by a special cookie by which one node can tell to the other which lock
358 * it actually means during communications. Such locks are called remote locks.
359 * The locks held by server only without any reference to a client are called
362 struct ldlm_namespace {
363 /** Backward link to OBD, required for LDLM pool to store new SLV. */
364 struct obd_device *ns_obd;
366 /** Flag indicating if namespace is on client instead of server */
367 ldlm_side_t ns_client;
369 /** Resource hash table for namespace. */
370 struct cfs_hash *ns_rs_hash;
375 /** big refcount (by bucket) */
379 * Namespace connect flags supported by server (may be changed via
380 * /proc, LRU resize may be disabled/enabled).
382 __u64 ns_connect_flags;
384 /** Client side original connect flags supported by server. */
385 __u64 ns_orig_connect_flags;
387 /* namespace debugfs dir entry */
388 struct dentry *ns_debugfs_entry;
391 * Position in global namespace list linking all namespaces on
394 struct list_head ns_list_chain;
397 * List of unused locks for this namespace. This list is also called
399 * Unused locks are locks with zero reader/writer reference counts.
400 * This list is only used on clients for lock caching purposes.
401 * When we want to release some locks voluntarily or if server wants
402 * us to release some locks due to e.g. memory pressure, we take locks
403 * to release from the head of this list.
404 * Locks are linked via l_lru field in \see struct ldlm_lock.
406 struct list_head ns_unused_list;
407 /** Number of locks in the LRU list above */
411 * Maximum number of locks permitted in the LRU. If 0, means locks
412 * are managed by pools and there is no preset limit, rather it is all
413 * controlled by available memory on this client and on server.
415 unsigned int ns_max_unused;
416 /** Maximum allowed age (last used time) for locks in the LRU */
417 unsigned int ns_max_age;
420 * Used to rate-limit ldlm_namespace_dump calls.
421 * \see ldlm_namespace_dump. Increased by 10 seconds every time
424 unsigned long ns_next_dump;
427 * LVB operations for this namespace.
428 * \see struct ldlm_valblock_ops
430 struct ldlm_valblock_ops *ns_lvbo;
433 * Used by filter code to store pointer to OBD of the service.
434 * Should be dropped in favor of \a ns_obd
439 * Wait queue used by __ldlm_namespace_free. Gets woken up every time
440 * a resource is removed.
442 wait_queue_head_t ns_waitq;
443 /** LDLM pool structure for this namespace */
444 struct ldlm_pool ns_pool;
445 /** Definition of how eagerly unused locks will be released from LRU */
446 ldlm_appetite_t ns_appetite;
448 /** Limit of parallel AST RPC count. */
449 unsigned ns_max_parallel_ast;
451 /** Callback to cancel locks before replaying it during recovery. */
452 ldlm_cancel_for_recovery ns_cancel_for_recovery;
454 /** LDLM lock stats */
455 struct lprocfs_stats *ns_stats;
458 * Flag to indicate namespace is being freed. Used to determine if
459 * recalculation of LDLM pool statistics should be skipped.
461 unsigned ns_stopping:1;
463 struct kobject ns_kobj; /* sysfs object */
464 struct completion ns_kobj_unregister;
468 * Returns 1 if namespace \a ns is a client namespace.
470 static inline int ns_is_client(struct ldlm_namespace *ns)
473 LASSERT(!(ns->ns_client & ~(LDLM_NAMESPACE_CLIENT |
474 LDLM_NAMESPACE_SERVER)));
475 LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT ||
476 ns->ns_client == LDLM_NAMESPACE_SERVER);
477 return ns->ns_client == LDLM_NAMESPACE_CLIENT;
481 * Returns 1 if namespace \a ns is a server namespace.
483 static inline int ns_is_server(struct ldlm_namespace *ns)
486 LASSERT(!(ns->ns_client & ~(LDLM_NAMESPACE_CLIENT |
487 LDLM_NAMESPACE_SERVER)));
488 LASSERT(ns->ns_client == LDLM_NAMESPACE_CLIENT ||
489 ns->ns_client == LDLM_NAMESPACE_SERVER);
490 return ns->ns_client == LDLM_NAMESPACE_SERVER;
494 * Returns 1 if namespace \a ns supports early lock cancel (ELC).
496 static inline int ns_connect_cancelset(struct ldlm_namespace *ns)
499 return !!(ns->ns_connect_flags & OBD_CONNECT_CANCELSET);
503 * Returns 1 if this namespace supports lru_resize.
505 static inline int ns_connect_lru_resize(struct ldlm_namespace *ns)
508 return !!(ns->ns_connect_flags & OBD_CONNECT_LRU_RESIZE);
511 static inline void ns_register_cancel(struct ldlm_namespace *ns,
512 ldlm_cancel_for_recovery arg)
515 ns->ns_cancel_for_recovery = arg;
520 /** Type for blocking callback function of a lock. */
521 typedef int (*ldlm_blocking_callback)(struct ldlm_lock *lock,
522 struct ldlm_lock_desc *new, void *data,
524 /** Type for completion callback function of a lock. */
525 typedef int (*ldlm_completion_callback)(struct ldlm_lock *lock, __u64 flags,
527 /** Type for glimpse callback function of a lock. */
528 typedef int (*ldlm_glimpse_callback)(struct ldlm_lock *lock, void *data);
530 /** Work list for sending GL ASTs to multiple locks. */
531 struct ldlm_glimpse_work {
532 struct ldlm_lock *gl_lock; /* lock to glimpse */
533 struct list_head gl_list; /* linkage to other gl work structs */
534 __u32 gl_flags;/* see LDLM_GL_WORK_* below */
535 union ldlm_gl_desc *gl_desc; /* glimpse descriptor to be packed in
536 * glimpse callback request */
539 /** The ldlm_glimpse_work is allocated on the stack and should not be freed. */
540 #define LDLM_GL_WORK_NOFREE 0x1
542 /** Interval node data for each LDLM_EXTENT lock. */
543 struct ldlm_interval {
544 struct interval_node li_node; /* node for tree management */
545 struct list_head li_group; /* the locks which have the same
546 * policy - group of the policy */
548 #define to_ldlm_interval(n) container_of(n, struct ldlm_interval, li_node)
551 * Interval tree for extent locks.
552 * The interval tree must be accessed under the resource lock.
553 * Interval trees are used for granted extent locks to speed up conflicts
554 * lookup. See ldlm/interval_tree.c for more details.
556 struct ldlm_interval_tree {
559 ldlm_mode_t lit_mode; /* lock mode */
560 struct interval_node *lit_root; /* actual ldlm_interval */
563 /** Whether to track references to exports by LDLM locks. */
564 #define LUSTRE_TRACKS_LOCK_EXP_REFS (0)
568 LCF_ASYNC = 0x1, /* Cancel locks asynchronously. */
569 LCF_LOCAL = 0x2, /* Cancel locks locally, not notifing server */
570 LCF_BL_AST = 0x4, /* Cancel locks marked as LDLM_FL_BL_AST
572 } ldlm_cancel_flags_t;
578 __u64 blocking_owner;
579 struct obd_export *blocking_export;
580 /* Protected by the hash lock */
586 struct ldlm_extent l_extent;
587 struct ldlm_flock l_flock;
588 struct ldlm_inodebits l_inodebits;
589 } ldlm_policy_data_t;
591 void ldlm_convert_policy_to_wire(ldlm_type_t type,
592 const ldlm_policy_data_t *lpolicy,
593 ldlm_wire_policy_data_t *wpolicy);
594 void ldlm_convert_policy_to_local(struct obd_export *exp, ldlm_type_t type,
595 const ldlm_wire_policy_data_t *wpolicy,
596 ldlm_policy_data_t *lpolicy);
606 * LDLM lock structure
608 * Represents a single LDLM lock and its state in memory. Each lock is
609 * associated with a single ldlm_resource, the object which is being
610 * locked. There may be multiple ldlm_locks on a single resource,
611 * depending on the lock type and whether the locks are conflicting or
617 * When remote side wants to tell us about a lock, they address
618 * it by this opaque handle. The handle does not hold a
619 * reference on the ldlm_lock, so it can be safely passed to
620 * other threads or nodes. When the lock needs to be accessed
621 * from the handle, it is looked up again in the lock table, and
622 * may no longer exist.
624 * Must be first in the structure.
626 struct portals_handle l_handle;
628 * Lock reference count.
629 * This is how many users have pointers to actual structure, so that
630 * we do not accidentally free lock structure that is in use.
634 * Internal spinlock protects l_resource. We should hold this lock
635 * first before taking res_lock.
639 * Pointer to actual resource this lock is in.
640 * ldlm_lock_change_resource() can change this.
642 struct ldlm_resource *l_resource;
644 * List item for client side LRU list.
645 * Protected by ns_lock in struct ldlm_namespace.
647 struct list_head l_lru;
649 * Linkage to resource's lock queues according to current lock state.
650 * (could be granted, waiting or converting)
651 * Protected by lr_lock in struct ldlm_resource.
653 struct list_head l_res_link;
655 * Tree node for ldlm_extent.
657 struct ldlm_interval *l_tree_node;
659 * Per export hash of locks.
660 * Protected by per-bucket exp->exp_lock_hash locks.
662 struct hlist_node l_exp_hash;
664 * Per export hash of flock locks.
665 * Protected by per-bucket exp->exp_flock_hash locks.
667 struct hlist_node l_exp_flock_hash;
670 * Protected by lr_lock.
672 ldlm_mode_t l_req_mode;
674 * Granted mode, also protected by lr_lock.
676 ldlm_mode_t l_granted_mode;
677 /** Lock completion handler pointer. Called when lock is granted. */
678 ldlm_completion_callback l_completion_ast;
680 * Lock blocking AST handler pointer.
681 * It plays two roles:
682 * - as a notification of an attempt to queue a conflicting lock (once)
683 * - as a notification when the lock is being cancelled.
685 * As such it's typically called twice: once for the initial conflict
686 * and then once more when the last user went away and the lock is
687 * cancelled (could happen recursively).
689 ldlm_blocking_callback l_blocking_ast;
691 * Lock glimpse handler.
692 * Glimpse handler is used to obtain LVB updates from a client by
695 ldlm_glimpse_callback l_glimpse_ast;
699 * This is a pointer to actual client export for locks that were granted
700 * to clients. Used server-side.
702 struct obd_export *l_export;
704 * Lock connection export.
705 * Pointer to server export on a client.
707 struct obd_export *l_conn_export;
710 * Remote lock handle.
711 * If the lock is remote, this is the handle of the other side lock
714 struct lustre_handle l_remote_handle;
717 * Representation of private data specific for a lock type.
718 * Examples are: extent range for extent lock or bitmask for ibits locks
720 ldlm_policy_data_t l_policy_data;
723 * Lock state flags. Protected by lr_lock.
724 * \see lustre_dlm_flags.h where the bits are defined.
729 * Lock r/w usage counters.
730 * Protected by lr_lock.
735 * If the lock is granted, a process sleeps on this waitq to learn when
736 * it's no longer in use. If the lock is not granted, a process sleeps
737 * on this waitq to learn when it becomes granted.
739 wait_queue_head_t l_waitq;
742 * Seconds. It will be updated if there is any activity related to
743 * the lock, e.g. enqueue the lock or send blocking AST.
745 time64_t l_last_activity;
748 * Time last used by e.g. being matched by lock match.
749 * Jiffies. Should be converted to time if needed.
751 unsigned long l_last_used;
753 /** Originally requested extent for the extent lock. */
754 struct ldlm_extent l_req_extent;
757 * Client-side-only members.
760 enum lvb_type l_lvb_type;
763 * Temporary storage for a LVB received during an enqueue operation.
768 /** Private storage for lock user. Opaque to LDLM. */
772 * Server-side-only members.
776 * Connection cookie for the client originating the operation.
777 * Used by Commit on Share (COS) code. Currently only used for
778 * inodebits locks on MDS.
780 __u64 l_client_cookie;
783 * List item for locks waiting for cancellation from clients.
784 * The lists this could be linked into are:
785 * waiting_locks_list (protected by waiting_locks_spinlock),
786 * then if the lock timed out, it is moved to
787 * expired_lock_thread.elt_expired_locks for further processing.
788 * Protected by elt_lock.
790 struct list_head l_pending_chain;
793 * Set when lock is sent a blocking AST. Time in seconds when timeout
794 * is reached and client holding this lock could be evicted.
795 * This timeout could be further extended by e.g. certain IO activity
797 * \see ost_rw_prolong_locks
799 unsigned long l_callback_timeout;
801 /** Local PID of process which created this lock. */
805 * Number of times blocking AST was sent for this lock.
806 * This is for debugging. Valid values are 0 and 1, if there is an
807 * attempt to send blocking AST more than once, an assertion would be
808 * hit. \see ldlm_work_bl_ast_lock
811 /** List item ldlm_add_ast_work_item() for case of blocking ASTs. */
812 struct list_head l_bl_ast;
813 /** List item ldlm_add_ast_work_item() for case of completion ASTs. */
814 struct list_head l_cp_ast;
815 /** For ldlm_add_ast_work_item() for "revoke" AST used in COS. */
816 struct list_head l_rk_ast;
819 * Pointer to a conflicting lock that caused blocking AST to be sent
822 struct ldlm_lock *l_blocking_lock;
825 * Protected by lr_lock, linkages to "skip lists".
826 * For more explanations of skip lists see ldlm/ldlm_inodebits.c
828 struct list_head l_sl_mode;
829 struct list_head l_sl_policy;
831 /** Reference tracking structure to debug leaked locks. */
832 struct lu_ref l_reference;
833 #if LUSTRE_TRACKS_LOCK_EXP_REFS
834 /* Debugging stuff for bug 20498, for tracking export references. */
835 /** number of export references taken */
837 /** link all locks referencing one export */
838 struct list_head l_exp_refs_link;
839 /** referenced export object */
840 struct obd_export *l_exp_refs_target;
843 * export blocking dlm lock list, protected by
844 * l_export->exp_bl_list_lock.
845 * Lock order of waiting_lists_spinlock, exp_bl_list_lock and res lock
846 * is: res lock -> exp_bl_list_lock -> wanting_lists_spinlock.
848 struct list_head l_exp_list;
852 * LDLM resource description.
853 * Basically, resource is a representation for a single object.
854 * Object has a name which is currently 4 64-bit integers. LDLM user is
855 * responsible for creation of a mapping between objects it wants to be
856 * protected and resource names.
858 * A resource can only hold locks of a single lock type, though there may be
859 * multiple ldlm_locks on a single resource, depending on the lock type and
860 * whether the locks are conflicting or not.
862 struct ldlm_resource {
863 struct ldlm_ns_bucket *lr_ns_bucket;
866 * List item for list in namespace hash.
867 * protected by ns_lock
869 struct hlist_node lr_hash;
871 /** Spinlock to protect locks under this resource. */
875 * protected by lr_lock
877 /** List of locks in granted state */
878 struct list_head lr_granted;
880 * List of locks that could not be granted due to conflicts and
881 * that are waiting for conflicts to go away */
882 struct list_head lr_waiting;
885 /* XXX No longer needed? Remove ASAP */
886 ldlm_mode_t lr_most_restr;
888 /** Type of locks this resource can hold. Only one type per resource. */
889 ldlm_type_t lr_type; /* LDLM_{PLAIN,EXTENT,FLOCK,IBITS} */
892 struct ldlm_res_id lr_name;
893 /** Reference count for this resource */
894 atomic_t lr_refcount;
897 * Interval trees (only for extent locks) for all modes of this resource
899 struct ldlm_interval_tree lr_itree[LCK_MODE_NUM];
902 * Server-side-only lock value block elements.
903 * To serialize lvbo_init.
905 struct mutex lr_lvb_mutex;
907 /** protected by lr_lock */
910 /** When the resource was considered as contended. */
911 unsigned long lr_contention_time;
912 /** List of references to this resource. For debugging. */
913 struct lu_ref lr_reference;
915 struct inode *lr_lvb_inode;
918 static inline bool ldlm_has_layout(struct ldlm_lock *lock)
920 return lock->l_resource->lr_type == LDLM_IBITS &&
921 lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_LAYOUT;
925 ldlm_ns_name(struct ldlm_namespace *ns)
927 return ns->ns_rs_hash->hs_name;
930 static inline struct ldlm_namespace *
931 ldlm_res_to_ns(struct ldlm_resource *res)
933 return res->lr_ns_bucket->nsb_namespace;
936 static inline struct ldlm_namespace *
937 ldlm_lock_to_ns(struct ldlm_lock *lock)
939 return ldlm_res_to_ns(lock->l_resource);
943 ldlm_lock_to_ns_name(struct ldlm_lock *lock)
945 return ldlm_ns_name(ldlm_lock_to_ns(lock));
948 static inline struct adaptive_timeout *
949 ldlm_lock_to_ns_at(struct ldlm_lock *lock)
951 return &lock->l_resource->lr_ns_bucket->nsb_at_estimate;
954 static inline int ldlm_lvbo_init(struct ldlm_resource *res)
956 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
958 if (ns->ns_lvbo != NULL && ns->ns_lvbo->lvbo_init != NULL)
959 return ns->ns_lvbo->lvbo_init(res);
964 static inline int ldlm_lvbo_size(struct ldlm_lock *lock)
966 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
968 if (ns->ns_lvbo != NULL && ns->ns_lvbo->lvbo_size != NULL)
969 return ns->ns_lvbo->lvbo_size(lock);
974 static inline int ldlm_lvbo_fill(struct ldlm_lock *lock, void *buf, int len)
976 struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
978 if (ns->ns_lvbo != NULL) {
979 LASSERT(ns->ns_lvbo->lvbo_fill != NULL);
980 return ns->ns_lvbo->lvbo_fill(lock, buf, len);
985 struct ldlm_ast_work {
986 struct ldlm_lock *w_lock;
988 struct ldlm_lock_desc w_desc;
989 struct list_head w_list;
996 * Common ldlm_enqueue parameters
998 struct ldlm_enqueue_info {
999 __u32 ei_type; /** Type of the lock being enqueued. */
1000 __u32 ei_mode; /** Mode of the lock being enqueued. */
1001 void *ei_cb_bl; /** blocking lock callback */
1002 void *ei_cb_cp; /** lock completion callback */
1003 void *ei_cb_gl; /** lock glimpse callback */
1004 void *ei_cbdata; /** Data to be passed into callbacks. */
1007 extern struct obd_ops ldlm_obd_ops;
1009 extern char *ldlm_lockname[];
1010 extern char *ldlm_typename[];
1011 char *ldlm_it2str(int it);
1014 * Just a fancy CDEBUG call with log level preset to LDLM_DEBUG.
1015 * For the cases where we do not have actual lock to print along
1016 * with a debugging message that is ldlm-related
1018 #define LDLM_DEBUG_NOLOCK(format, a...) \
1019 CDEBUG(D_DLMTRACE, "### " format "\n", ##a)
1022 * Support function for lock information printing into debug logs.
1025 #define ldlm_lock_debug(msgdata, mask, cdls, lock, fmt, a...) do { \
1026 CFS_CHECK_STACK(msgdata, mask, cdls); \
1028 if (((mask) & D_CANTMASK) != 0 || \
1029 ((libcfs_debug & (mask)) != 0 && \
1030 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1031 _ldlm_lock_debug(lock, msgdata, fmt, ##a); \
1034 void _ldlm_lock_debug(struct ldlm_lock *lock,
1035 struct libcfs_debug_msg_data *data,
1036 const char *fmt, ...)
1040 * Rate-limited version of lock printing function.
1042 #define LDLM_DEBUG_LIMIT(mask, lock, fmt, a...) do { \
1043 static struct cfs_debug_limit_state _ldlm_cdls; \
1044 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, mask, &_ldlm_cdls); \
1045 ldlm_lock_debug(&msgdata, mask, &_ldlm_cdls, lock, "### " fmt, ##a);\
1048 #define LDLM_ERROR(lock, fmt, a...) LDLM_DEBUG_LIMIT(D_ERROR, lock, fmt, ## a)
1049 #define LDLM_WARN(lock, fmt, a...) LDLM_DEBUG_LIMIT(D_WARNING, lock, fmt, ## a)
1051 /** Non-rate-limited lock printing function for debugging purposes. */
1052 #define LDLM_DEBUG(lock, fmt, a...) do { \
1053 if (likely(lock != NULL)) { \
1054 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_DLMTRACE, NULL); \
1055 ldlm_lock_debug(&msgdata, D_DLMTRACE, NULL, lock, \
1058 LDLM_DEBUG_NOLOCK("no dlm lock: " fmt, ##a); \
1062 typedef int (*ldlm_processing_policy)(struct ldlm_lock *lock, __u64 *flags,
1063 int first_enq, ldlm_error_t *err,
1064 struct list_head *work_list);
1067 * Return values for lock iterators.
1068 * Also used during deciding of lock grants and cancellations.
1070 #define LDLM_ITER_CONTINUE 1 /* keep iterating */
1071 #define LDLM_ITER_STOP 2 /* stop iterating */
1073 typedef int (*ldlm_iterator_t)(struct ldlm_lock *, void *);
1074 typedef int (*ldlm_res_iterator_t)(struct ldlm_resource *, void *);
1076 /** \defgroup ldlm_iterator Lock iterators
1078 * LDLM provides for a way to iterate through every lock on a resource or
1079 * namespace or every resource in a namespace.
1081 int ldlm_resource_foreach(struct ldlm_resource *res, ldlm_iterator_t iter,
1083 void ldlm_namespace_foreach(struct ldlm_namespace *ns, ldlm_iterator_t iter,
1085 int ldlm_resource_iterate(struct ldlm_namespace *, const struct ldlm_res_id *,
1086 ldlm_iterator_t iter, void *data);
1087 /** @} ldlm_iterator */
1089 int ldlm_replay_locks(struct obd_import *imp);
1092 int ldlm_flock_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data);
1095 __u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms);
1097 struct ldlm_callback_suite {
1098 ldlm_completion_callback lcs_completion;
1099 ldlm_blocking_callback lcs_blocking;
1100 ldlm_glimpse_callback lcs_glimpse;
1104 int ldlm_get_ref(void);
1105 void ldlm_put_ref(void);
1106 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req);
1109 void ldlm_lock2handle(const struct ldlm_lock *lock,
1110 struct lustre_handle *lockh);
1111 struct ldlm_lock *__ldlm_handle2lock(const struct lustre_handle *, __u64 flags);
1112 void ldlm_cancel_callback(struct ldlm_lock *);
1113 int ldlm_lock_remove_from_lru(struct ldlm_lock *);
1114 int ldlm_lock_set_data(struct lustre_handle *, void *);
1117 * Obtain a lock reference by its handle.
1119 static inline struct ldlm_lock *ldlm_handle2lock(const struct lustre_handle *h)
1121 return __ldlm_handle2lock(h, 0);
1124 #define LDLM_LOCK_REF_DEL(lock) \
1125 lu_ref_del(&lock->l_reference, "handle", current)
1127 static inline struct ldlm_lock *
1128 ldlm_handle2lock_long(const struct lustre_handle *h, __u64 flags)
1130 struct ldlm_lock *lock;
1132 lock = __ldlm_handle2lock(h, flags);
1134 LDLM_LOCK_REF_DEL(lock);
1139 * Update Lock Value Block Operations (LVBO) on a resource taking into account
1140 * data from request \a r
1142 static inline int ldlm_res_lvbo_update(struct ldlm_resource *res,
1143 struct ptlrpc_request *r, int increase)
1145 if (ldlm_res_to_ns(res)->ns_lvbo &&
1146 ldlm_res_to_ns(res)->ns_lvbo->lvbo_update) {
1147 return ldlm_res_to_ns(res)->ns_lvbo->lvbo_update(res, r,
1153 int ldlm_error2errno(ldlm_error_t error);
1155 #if LUSTRE_TRACKS_LOCK_EXP_REFS
1156 void ldlm_dump_export_locks(struct obd_export *exp);
1160 * Release a temporary lock reference obtained by ldlm_handle2lock() or
1161 * __ldlm_handle2lock().
1163 #define LDLM_LOCK_PUT(lock) \
1165 LDLM_LOCK_REF_DEL(lock); \
1166 /*LDLM_DEBUG((lock), "put");*/ \
1167 ldlm_lock_put(lock); \
1171 * Release a lock reference obtained by some other means (see
1174 #define LDLM_LOCK_RELEASE(lock) \
1176 /*LDLM_DEBUG((lock), "put");*/ \
1177 ldlm_lock_put(lock); \
1180 #define LDLM_LOCK_GET(lock) \
1182 ldlm_lock_get(lock); \
1183 /*LDLM_DEBUG((lock), "get");*/ \
1187 #define ldlm_lock_list_put(head, member, count) \
1189 struct ldlm_lock *_lock, *_next; \
1191 list_for_each_entry_safe(_lock, _next, head, member) { \
1194 list_del_init(&_lock->member); \
1195 LDLM_LOCK_RELEASE(_lock); \
1200 struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
1201 void ldlm_lock_put(struct ldlm_lock *lock);
1202 void ldlm_lock_destroy(struct ldlm_lock *lock);
1203 void ldlm_lock2desc(struct ldlm_lock *lock, struct ldlm_lock_desc *desc);
1204 void ldlm_lock_addref(struct lustre_handle *lockh, __u32 mode);
1205 int ldlm_lock_addref_try(struct lustre_handle *lockh, __u32 mode);
1206 void ldlm_lock_decref(struct lustre_handle *lockh, __u32 mode);
1207 void ldlm_lock_decref_and_cancel(struct lustre_handle *lockh, __u32 mode);
1208 void ldlm_lock_fail_match_locked(struct ldlm_lock *lock);
1209 void ldlm_lock_allow_match(struct ldlm_lock *lock);
1210 void ldlm_lock_allow_match_locked(struct ldlm_lock *lock);
1211 ldlm_mode_t ldlm_lock_match(struct ldlm_namespace *ns, __u64 flags,
1212 const struct ldlm_res_id *, ldlm_type_t type,
1213 ldlm_policy_data_t *, ldlm_mode_t mode,
1214 struct lustre_handle *, int unref);
1215 ldlm_mode_t ldlm_revalidate_lock_handle(struct lustre_handle *lockh,
1217 void ldlm_lock_cancel(struct ldlm_lock *lock);
1218 void ldlm_lock_dump_handle(int level, struct lustre_handle *);
1219 void ldlm_unlink_lock_skiplist(struct ldlm_lock *req);
1222 struct ldlm_namespace *
1223 ldlm_namespace_new(struct obd_device *obd, char *name,
1224 ldlm_side_t client, ldlm_appetite_t apt,
1225 ldlm_ns_type_t ns_type);
1226 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags);
1227 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client);
1228 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client);
1229 void ldlm_namespace_get(struct ldlm_namespace *ns);
1230 void ldlm_namespace_put(struct ldlm_namespace *ns);
1231 int ldlm_debugfs_setup(void);
1232 void ldlm_debugfs_cleanup(void);
1234 /* resource.c - internal */
1235 struct ldlm_resource *ldlm_resource_get(struct ldlm_namespace *ns,
1236 struct ldlm_resource *parent,
1237 const struct ldlm_res_id *,
1238 ldlm_type_t type, int create);
1239 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res);
1240 int ldlm_resource_putref(struct ldlm_resource *res);
1241 void ldlm_resource_add_lock(struct ldlm_resource *res,
1242 struct list_head *head,
1243 struct ldlm_lock *lock);
1244 void ldlm_resource_unlink_lock(struct ldlm_lock *lock);
1245 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc);
1246 void ldlm_dump_all_namespaces(ldlm_side_t client, int level);
1247 void ldlm_namespace_dump(int level, struct ldlm_namespace *);
1248 void ldlm_resource_dump(int level, struct ldlm_resource *);
1249 int ldlm_lock_change_resource(struct ldlm_namespace *, struct ldlm_lock *,
1250 const struct ldlm_res_id *);
1252 #define LDLM_RESOURCE_ADDREF(res) do { \
1253 lu_ref_add_atomic(&(res)->lr_reference, __func__, current); \
1256 #define LDLM_RESOURCE_DELREF(res) do { \
1257 lu_ref_del(&(res)->lr_reference, __func__, current); \
1260 /* ldlm_request.c */
1261 int ldlm_expired_completion_wait(void *data);
1262 /** \defgroup ldlm_local_ast Default AST handlers for local locks
1263 * These AST handlers are typically used for server-side local locks and are
1264 * also used by client-side lock handlers to perform minimum level base
1267 int ldlm_blocking_ast_nocheck(struct ldlm_lock *lock);
1268 int ldlm_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1269 void *data, int flag);
1270 int ldlm_completion_ast_async(struct ldlm_lock *lock, __u64 flags, void *data);
1271 int ldlm_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data);
1272 /** @} ldlm_local_ast */
1274 /** \defgroup ldlm_cli_api API to operate on locks from actual LDLM users.
1275 * These are typically used by client and server (*_local versions)
1276 * to obtain and release locks.
1278 int ldlm_cli_enqueue(struct obd_export *exp, struct ptlrpc_request **reqp,
1279 struct ldlm_enqueue_info *einfo,
1280 const struct ldlm_res_id *res_id,
1281 ldlm_policy_data_t const *policy, __u64 *flags,
1282 void *lvb, __u32 lvb_len, enum lvb_type lvb_type,
1283 struct lustre_handle *lockh, int async);
1284 int ldlm_prep_enqueue_req(struct obd_export *exp,
1285 struct ptlrpc_request *req,
1286 struct list_head *cancels,
1288 int ldlm_prep_elc_req(struct obd_export *exp,
1289 struct ptlrpc_request *req,
1290 int version, int opc, int canceloff,
1291 struct list_head *cancels, int count);
1293 int ldlm_cli_enqueue_fini(struct obd_export *exp, struct ptlrpc_request *req,
1294 ldlm_type_t type, __u8 with_policy, ldlm_mode_t mode,
1295 __u64 *flags, void *lvb, __u32 lvb_len,
1296 struct lustre_handle *lockh, int rc);
1297 int ldlm_cli_update_pool(struct ptlrpc_request *req);
1298 int ldlm_cli_cancel(struct lustre_handle *lockh,
1299 ldlm_cancel_flags_t cancel_flags);
1300 int ldlm_cli_cancel_unused(struct ldlm_namespace *, const struct ldlm_res_id *,
1301 ldlm_cancel_flags_t flags, void *opaque);
1302 int ldlm_cli_cancel_unused_resource(struct ldlm_namespace *ns,
1303 const struct ldlm_res_id *res_id,
1304 ldlm_policy_data_t *policy,
1306 ldlm_cancel_flags_t flags,
1308 int ldlm_cli_cancel_req(struct obd_export *exp, struct list_head *head,
1309 int count, ldlm_cancel_flags_t flags);
1310 int ldlm_cancel_resource_local(struct ldlm_resource *res,
1311 struct list_head *cancels,
1312 ldlm_policy_data_t *policy,
1313 ldlm_mode_t mode, __u64 lock_flags,
1314 ldlm_cancel_flags_t cancel_flags, void *opaque);
1315 int ldlm_cli_cancel_list_local(struct list_head *cancels, int count,
1316 ldlm_cancel_flags_t flags);
1317 int ldlm_cli_cancel_list(struct list_head *head, int count,
1318 struct ptlrpc_request *req, ldlm_cancel_flags_t flags);
1319 /** @} ldlm_cli_api */
1322 /* This has to be here because recursive inclusion sucks. */
1323 int intent_disposition(struct ldlm_reply *rep, int flag);
1324 void intent_set_disposition(struct ldlm_reply *rep, int flag);
1327 /* ioctls for trying requests */
1328 #define IOC_LDLM_TYPE 'f'
1329 #define IOC_LDLM_MIN_NR 40
1331 #define IOC_LDLM_TEST _IOWR('f', 40, long)
1332 #define IOC_LDLM_DUMP _IOWR('f', 41, long)
1333 #define IOC_LDLM_REGRESS_START _IOWR('f', 42, long)
1334 #define IOC_LDLM_REGRESS_STOP _IOWR('f', 43, long)
1335 #define IOC_LDLM_MAX_NR 43
1338 * "Modes" of acquiring lock_res, necessary to tell lockdep that taking more
1339 * than one lock_res is dead-lock safe.
1341 enum lock_res_type {
1346 /** Lock resource. */
1347 static inline void lock_res(struct ldlm_resource *res)
1349 spin_lock(&res->lr_lock);
1352 /** Lock resource with a way to instruct lockdep code about nestedness-safe. */
1353 static inline void lock_res_nested(struct ldlm_resource *res,
1354 enum lock_res_type mode)
1356 spin_lock_nested(&res->lr_lock, mode);
1359 /** Unlock resource. */
1360 static inline void unlock_res(struct ldlm_resource *res)
1362 spin_unlock(&res->lr_lock);
1365 /** Check if resource is already locked, assert if not. */
1366 static inline void check_res_locked(struct ldlm_resource *res)
1368 assert_spin_locked(&res->lr_lock);
1371 struct ldlm_resource *lock_res_and_lock(struct ldlm_lock *lock);
1372 void unlock_res_and_lock(struct ldlm_lock *lock);
1375 /** \defgroup ldlm_pools Various LDLM pool related functions
1376 * There are not used outside of ldlm.
1379 int ldlm_pools_recalc(ldlm_side_t client);
1380 int ldlm_pools_init(void);
1381 void ldlm_pools_fini(void);
1383 int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
1384 int idx, ldlm_side_t client);
1385 int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
1387 void ldlm_pool_fini(struct ldlm_pool *pl);
1388 int ldlm_pool_setup(struct ldlm_pool *pl, int limit);
1389 int ldlm_pool_recalc(struct ldlm_pool *pl);
1390 __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl);
1391 __u64 ldlm_pool_get_slv(struct ldlm_pool *pl);
1392 __u32 ldlm_pool_get_limit(struct ldlm_pool *pl);
1393 void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv);
1394 void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit);
1395 void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock);
1396 void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock);