1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /******************************************************************************
3 *******************************************************************************
5 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
6 ** Copyright (C) 2004-2011 Red Hat, Inc. All rights reserved.
9 *******************************************************************************
10 ******************************************************************************/
12 #ifndef __DLM_INTERNAL_DOT_H__
13 #define __DLM_INTERNAL_DOT_H__
16 * This is the main header file to be included in each DLM source file.
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/types.h>
22 #include <linux/ctype.h>
23 #include <linux/spinlock.h>
24 #include <linux/vmalloc.h>
25 #include <linux/list.h>
26 #include <linux/errno.h>
27 #include <linux/random.h>
28 #include <linux/delay.h>
29 #include <linux/socket.h>
30 #include <linux/kthread.h>
31 #include <linux/kobject.h>
32 #include <linux/kref.h>
33 #include <linux/kernel.h>
34 #include <linux/jhash.h>
35 #include <linux/miscdevice.h>
36 #include <linux/mutex.h>
37 #include <linux/idr.h>
38 #include <linux/ratelimit.h>
39 #include <linux/uaccess.h>
41 #include <linux/dlm.h>
44 /* Size of the temp buffer midcomms allocates on the stack.
45 We try to make this large enough so most messages fit.
46 FIXME: should sctp make this unnecessary? */
48 #define DLM_INBUF_LEN 148
62 #define log_print(fmt, args...) \
63 printk(KERN_ERR "dlm: "fmt"\n" , ##args)
64 #define log_print_ratelimited(fmt, args...) \
65 printk_ratelimited(KERN_ERR "dlm: "fmt"\n", ##args)
66 #define log_error(ls, fmt, args...) \
67 printk(KERN_ERR "dlm: %s: " fmt "\n", (ls)->ls_name , ##args)
69 #define log_rinfo(ls, fmt, args...) \
71 if (dlm_config.ci_log_info) \
72 printk(KERN_INFO "dlm: %s: " fmt "\n", \
73 (ls)->ls_name, ##args); \
74 else if (dlm_config.ci_log_debug) \
75 printk(KERN_DEBUG "dlm: %s: " fmt "\n", \
76 (ls)->ls_name , ##args); \
79 #define log_debug(ls, fmt, args...) \
81 if (dlm_config.ci_log_debug) \
82 printk(KERN_DEBUG "dlm: %s: " fmt "\n", \
83 (ls)->ls_name , ##args); \
86 #define log_limit(ls, fmt, args...) \
88 if (dlm_config.ci_log_debug) \
89 printk_ratelimited(KERN_DEBUG "dlm: %s: " fmt "\n", \
90 (ls)->ls_name , ##args); \
93 #define DLM_ASSERT(x, do) \
97 printk(KERN_ERR "\nDLM: Assertion failed on line %d of file %s\n" \
98 "DLM: assertion: \"%s\"\n" \
99 "DLM: time = %lu\n", \
100 __LINE__, __FILE__, #x, jiffies); \
103 panic("DLM: Record message above and reboot.\n"); \
108 #define DLM_RTF_SHRINK 0x00000001
110 struct dlm_rsbtable {
119 * Lockspace member (per node in a ls)
123 struct list_head list;
133 * Save and manage recovery state for a lockspace.
137 struct list_head list;
138 struct dlm_config_node *nodes;
144 * Pass input args to second stage locking function.
149 void (*astfn) (void *astparam);
151 void (*bastfn) (void *astparam, int mode);
153 struct dlm_lksb *lksb;
154 unsigned long timeout;
161 * A lock can be one of three types:
163 * local copy lock is mastered locally
164 * (lkb_nodeid is zero and DLM_LKF_MSTCPY is not set)
165 * process copy lock is mastered on a remote node
166 * (lkb_nodeid is non-zero and DLM_LKF_MSTCPY is not set)
167 * master copy master node's copy of a lock owned by remote node
168 * (lkb_nodeid is non-zero and DLM_LKF_MSTCPY is set)
170 * lkb_exflags: a copy of the most recent flags arg provided to dlm_lock or
171 * dlm_unlock. The dlm does not modify these or use any private flags in
172 * this field; it only contains DLM_LKF_ flags from dlm.h. These flags
173 * are sent as-is to the remote master when the lock is remote.
175 * lkb_flags: internal dlm flags (DLM_IFL_ prefix) from dlm_internal.h.
176 * Some internal flags are shared between the master and process nodes;
177 * these shared flags are kept in the lower two bytes. One of these
178 * flags set on the master copy will be propagated to the process copy
179 * and v.v. Other internal flags are private to the master or process
180 * node (e.g. DLM_IFL_MSTCPY). These are kept in the high two bytes.
182 * lkb_sbflags: status block flags. These flags are copied directly into
183 * the caller's lksb.sb_flags prior to the dlm_lock/dlm_unlock completion
184 * ast. All defined in dlm.h with DLM_SBF_ prefix.
186 * lkb_status: the lock status indicates which rsb queue the lock is
187 * on, grant, convert, or wait. DLM_LKSTS_ WAITING/GRANTED/CONVERT
189 * lkb_wait_type: the dlm message type (DLM_MSG_ prefix) for which a
190 * reply is needed. Only set when the lkb is on the lockspace waiters
191 * list awaiting a reply from a remote node.
193 * lkb_nodeid: when the lkb is a local copy, nodeid is 0; when the lkb
194 * is a master copy, nodeid specifies the remote lock holder, when the
195 * lkb is a process copy, the nodeid specifies the lock master.
200 #define DLM_LKSTS_WAITING 1
201 #define DLM_LKSTS_GRANTED 2
202 #define DLM_LKSTS_CONVERT 3
206 #define DLM_IFL_MSTCPY 0x00010000
207 #define DLM_IFL_RESEND 0x00020000
208 #define DLM_IFL_DEAD 0x00040000
209 #define DLM_IFL_OVERLAP_UNLOCK 0x00080000
210 #define DLM_IFL_OVERLAP_CANCEL 0x00100000
211 #define DLM_IFL_ENDOFLIFE 0x00200000
212 #define DLM_IFL_WATCH_TIMEWARN 0x00400000
213 #define DLM_IFL_TIMEOUT_CANCEL 0x00800000
214 #define DLM_IFL_DEADLOCK_CANCEL 0x01000000
215 #define DLM_IFL_STUB_MS 0x02000000 /* magic number for m_flags */
216 #define DLM_IFL_USER 0x00000001
217 #define DLM_IFL_ORPHAN 0x00000002
219 #define DLM_CALLBACKS_SIZE 6
221 #define DLM_CB_CAST 0x00000001
222 #define DLM_CB_BAST 0x00000002
223 #define DLM_CB_SKIP 0x00000004
225 struct dlm_callback {
227 uint32_t flags; /* DLM_CBF_ */
228 int sb_status; /* copy to lksb status */
229 uint8_t sb_flags; /* copy to lksb flags */
230 int8_t mode; /* rq mode of bast, gr mode of cast */
234 struct dlm_rsb *lkb_resource; /* the rsb */
236 int lkb_nodeid; /* copied from rsb */
237 int lkb_ownpid; /* pid of lock owner */
238 uint32_t lkb_id; /* our lock ID */
239 uint32_t lkb_remid; /* lock ID on remote partner */
240 uint32_t lkb_exflags; /* external flags from caller */
241 uint32_t lkb_sbflags; /* lksb flags */
242 uint32_t lkb_flags; /* internal flags */
243 uint32_t lkb_lvbseq; /* lvb sequence number */
245 int8_t lkb_status; /* granted, waiting, convert */
246 int8_t lkb_rqmode; /* requested lock mode */
247 int8_t lkb_grmode; /* granted lock mode */
248 int8_t lkb_highbast; /* highest mode bast sent for */
250 int8_t lkb_wait_type; /* type of reply waiting for */
251 int8_t lkb_wait_count;
252 int lkb_wait_nodeid; /* for debugging */
254 struct list_head lkb_statequeue; /* rsb g/c/w list */
255 struct list_head lkb_rsb_lookup; /* waiting for rsb lookup */
256 struct list_head lkb_wait_reply; /* waiting for remote reply */
257 struct list_head lkb_ownqueue; /* list of locks for a process */
258 struct list_head lkb_time_list;
259 ktime_t lkb_timestamp;
260 ktime_t lkb_wait_time;
261 unsigned long lkb_timeout_cs;
263 struct mutex lkb_cb_mutex;
264 struct work_struct lkb_cb_work;
265 struct list_head lkb_cb_list; /* for ls_cb_delay or proc->asts */
266 struct dlm_callback lkb_callbacks[DLM_CALLBACKS_SIZE];
267 struct dlm_callback lkb_last_cast;
268 struct dlm_callback lkb_last_bast;
269 ktime_t lkb_last_cast_time; /* for debugging */
270 ktime_t lkb_last_bast_time; /* for debugging */
272 uint64_t lkb_recover_seq; /* from ls_recover_seq */
275 struct dlm_lksb *lkb_lksb; /* caller's status block */
276 void (*lkb_astfn) (void *astparam);
277 void (*lkb_bastfn) (void *astparam, int mode);
279 void *lkb_astparam; /* caller's ast arg */
280 struct dlm_user_args *lkb_ua;
285 * res_master_nodeid is "normal": 0 is unset/invalid, non-zero is the real
286 * nodeid, even when nodeid is our_nodeid.
288 * res_nodeid is "odd": -1 is unset/invalid, zero means our_nodeid,
289 * greater than zero when another nodeid.
291 * (TODO: remove res_nodeid and only use res_master_nodeid)
295 struct dlm_ls *res_ls; /* the lockspace */
297 struct mutex res_mutex;
298 unsigned long res_flags;
299 int res_length; /* length of rsb name */
301 int res_master_nodeid;
303 int res_id; /* for ls_recover_idr */
306 uint32_t res_bucket; /* rsbtbl */
307 unsigned long res_toss_time;
308 uint32_t res_first_lkid;
309 struct list_head res_lookup; /* lkbs waiting on first */
311 struct list_head res_hashchain;
312 struct rb_node res_hashnode; /* rsbtbl */
314 struct list_head res_grantqueue;
315 struct list_head res_convertqueue;
316 struct list_head res_waitqueue;
318 struct list_head res_root_list; /* used for recovery */
319 struct list_head res_recover_list; /* used for recovery */
320 int res_recover_locks_count;
323 char res_name[DLM_RESNAME_MAXLEN+1];
326 /* dlm_master_lookup() flags */
328 #define DLM_LU_RECOVER_DIR 1
329 #define DLM_LU_RECOVER_MASTER 2
331 /* dlm_master_lookup() results */
333 #define DLM_LU_MATCH 1
336 /* find_rsb() flags */
338 #define R_REQUEST 0x00000001
339 #define R_RECEIVE_REQUEST 0x00000002
340 #define R_RECEIVE_RECOVER 0x00000004
345 RSB_MASTER_UNCERTAIN,
347 RSB_VALNOTVALID_PREV,
352 RSB_RECOVER_LVB_INVAL,
355 static inline void rsb_set_flag(struct dlm_rsb *r, enum rsb_flags flag)
357 __set_bit(flag, &r->res_flags);
360 static inline void rsb_clear_flag(struct dlm_rsb *r, enum rsb_flags flag)
362 __clear_bit(flag, &r->res_flags);
365 static inline int rsb_flag(struct dlm_rsb *r, enum rsb_flags flag)
367 return test_bit(flag, &r->res_flags);
371 /* dlm_header is first element of all structs sent between nodes */
373 #define DLM_HEADER_MAJOR 0x00030000
374 #define DLM_HEADER_MINOR 0x00000002
376 #define DLM_VERSION_3_1 0x00030001
377 #define DLM_VERSION_3_2 0x00030002
379 #define DLM_HEADER_SLOTS 0x00000001
390 /* for DLM_MSG and DLM_RCOM */
391 uint32_t h_lockspace;
392 /* for DLM_ACK and DLM_OPTS */
395 uint32_t h_nodeid; /* nodeid of sender */
397 uint8_t h_cmd; /* DLM_MSG, DLM_RCOM */
401 #define DLM_MSG_REQUEST 1
402 #define DLM_MSG_CONVERT 2
403 #define DLM_MSG_UNLOCK 3
404 #define DLM_MSG_CANCEL 4
405 #define DLM_MSG_REQUEST_REPLY 5
406 #define DLM_MSG_CONVERT_REPLY 6
407 #define DLM_MSG_UNLOCK_REPLY 7
408 #define DLM_MSG_CANCEL_REPLY 8
409 #define DLM_MSG_GRANT 9
410 #define DLM_MSG_BAST 10
411 #define DLM_MSG_LOOKUP 11
412 #define DLM_MSG_REMOVE 12
413 #define DLM_MSG_LOOKUP_REPLY 13
414 #define DLM_MSG_PURGE 14
417 struct dlm_header m_header;
418 uint32_t m_type; /* DLM_MSG_ */
421 uint32_t m_lkid; /* lkid on sender */
422 uint32_t m_remid; /* lkid on receiver */
423 uint32_t m_parent_lkid;
424 uint32_t m_parent_remid;
435 int m_result; /* 0 or -EXXX */
436 char m_extra[]; /* name or lvb */
440 #define DLM_RS_NODES 0x00000001
441 #define DLM_RS_NODES_ALL 0x00000002
442 #define DLM_RS_DIR 0x00000004
443 #define DLM_RS_DIR_ALL 0x00000008
444 #define DLM_RS_LOCKS 0x00000010
445 #define DLM_RS_LOCKS_ALL 0x00000020
446 #define DLM_RS_DONE 0x00000040
447 #define DLM_RS_DONE_ALL 0x00000080
449 #define DLM_RCOM_STATUS 1
450 #define DLM_RCOM_NAMES 2
451 #define DLM_RCOM_LOOKUP 3
452 #define DLM_RCOM_LOCK 4
453 #define DLM_RCOM_STATUS_REPLY 5
454 #define DLM_RCOM_NAMES_REPLY 6
455 #define DLM_RCOM_LOOKUP_REPLY 7
456 #define DLM_RCOM_LOCK_REPLY 8
459 struct dlm_header rc_header;
460 uint32_t rc_type; /* DLM_RCOM_ */
461 int rc_result; /* multi-purpose */
462 uint64_t rc_id; /* match reply with request */
463 uint64_t rc_seq; /* sender's ls_recover_seq */
464 uint64_t rc_seq_reply; /* remote ls_recover_seq */
468 struct dlm_opt_header {
472 /* need to be 8 byte aligned */
476 /* encapsulation header */
478 struct dlm_header o_header;
487 struct dlm_header header; /* common to other two */
488 struct dlm_message message;
489 struct dlm_rcom rcom;
490 struct dlm_opts opts;
493 #define DLM_RSF_NEED_SLOTS 0x00000001
495 /* RCOM_STATUS data */
502 /* RCOM_STATUS_REPLY data */
507 /* DLM_HEADER_SLOTS adds: */
511 __le32 rf_generation;
527 __le32 rl_parent_lkid;
528 __le32 rl_parent_remid;
539 char rl_name[DLM_RESNAME_MAXLEN];
544 * The max number of resources per rsbtbl bucket that shrink will attempt
545 * to remove in each iteration.
548 #define DLM_REMOVE_NAMES_MAX 8
551 struct list_head ls_list; /* list of lockspaces */
552 dlm_lockspace_t *ls_local_handle;
553 uint32_t ls_global_id; /* global unique lockspace ID */
554 uint32_t ls_generation;
557 int ls_count; /* refcount of processes in
558 the dlm using this ls */
559 int ls_create_count; /* create/release refcount */
560 unsigned long ls_flags; /* LSFL_ */
561 unsigned long ls_scan_time;
562 struct kobject ls_kobj;
564 struct idr ls_lkbidr;
565 spinlock_t ls_lkbidr_spin;
567 struct dlm_rsbtable *ls_rsbtbl;
568 uint32_t ls_rsbtbl_size;
570 struct mutex ls_waiters_mutex;
571 struct list_head ls_waiters; /* lkbs needing a reply */
573 struct mutex ls_orphans_mutex;
574 struct list_head ls_orphans;
576 struct mutex ls_timeout_mutex;
577 struct list_head ls_timeout;
579 spinlock_t ls_new_rsb_spin;
580 int ls_new_rsb_count;
581 struct list_head ls_new_rsb; /* new rsb structs */
583 spinlock_t ls_remove_spin;
584 char ls_remove_name[DLM_RESNAME_MAXLEN+1];
585 char *ls_remove_names[DLM_REMOVE_NAMES_MAX];
587 int ls_remove_lens[DLM_REMOVE_NAMES_MAX];
589 struct list_head ls_nodes; /* current nodes in ls */
590 struct list_head ls_nodes_gone; /* dead node list, recovery */
591 int ls_num_nodes; /* number of nodes in ls */
599 struct dlm_slot *ls_slots;
601 struct dlm_rsb ls_stub_rsb; /* for returning errors */
602 struct dlm_lkb ls_stub_lkb; /* for returning errors */
603 struct dlm_message ls_stub_ms; /* for faking a reply */
605 struct dentry *ls_debug_rsb_dentry; /* debugfs */
606 struct dentry *ls_debug_waiters_dentry; /* debugfs */
607 struct dentry *ls_debug_locks_dentry; /* debugfs */
608 struct dentry *ls_debug_all_dentry; /* debugfs */
609 struct dentry *ls_debug_toss_dentry; /* debugfs */
611 wait_queue_head_t ls_uevent_wait; /* user part of join/leave */
612 int ls_uevent_result;
613 struct completion ls_members_done;
614 int ls_members_result;
616 struct miscdevice ls_device;
618 struct workqueue_struct *ls_callback_wq;
620 /* recovery related */
622 struct mutex ls_cb_mutex;
623 struct list_head ls_cb_delay; /* save for queue_work later */
624 struct timer_list ls_timer;
625 struct task_struct *ls_recoverd_task;
626 struct mutex ls_recoverd_active;
627 spinlock_t ls_recover_lock;
628 unsigned long ls_recover_begin; /* jiffies timestamp */
629 uint32_t ls_recover_status; /* DLM_RS_ */
630 uint64_t ls_recover_seq;
631 struct dlm_recover *ls_recover_args;
632 struct rw_semaphore ls_in_recovery; /* block local requests */
633 struct rw_semaphore ls_recv_active; /* block dlm_recv */
634 struct list_head ls_requestqueue;/* queue remote requests */
635 struct mutex ls_requestqueue_mutex;
636 struct dlm_rcom *ls_recover_buf;
637 int ls_recover_nodeid; /* for debugging */
638 unsigned int ls_recover_dir_sent_res; /* for log info */
639 unsigned int ls_recover_dir_sent_msg; /* for log info */
640 unsigned int ls_recover_locks_in; /* for log info */
641 uint64_t ls_rcom_seq;
642 spinlock_t ls_rcom_spin;
643 struct list_head ls_recover_list;
644 spinlock_t ls_recover_list_lock;
645 int ls_recover_list_count;
646 struct idr ls_recover_idr;
647 spinlock_t ls_recover_idr_lock;
648 wait_queue_head_t ls_wait_general;
649 wait_queue_head_t ls_recover_lock_wait;
650 struct mutex ls_clear_proc_locks;
652 struct list_head ls_root_list; /* root resources */
653 struct rw_semaphore ls_root_sem; /* protect root_list */
655 const struct dlm_lockspace_ops *ls_ops;
663 * LSFL_RECOVER_STOP - dlm_ls_stop() sets this to tell dlm recovery routines
664 * that they should abort what they're doing so new recovery can be started.
666 * LSFL_RECOVER_DOWN - dlm_ls_stop() sets this to tell dlm_recoverd that it
667 * should do down_write() on the in_recovery rw_semaphore. (doing down_write
668 * within dlm_ls_stop causes complaints about the lock acquired/released
669 * in different contexts.)
671 * LSFL_RECOVER_LOCK - dlm_recoverd holds the in_recovery rw_semaphore.
672 * It sets this after it is done with down_write() on the in_recovery
673 * rw_semaphore and clears it after it has released the rw_semaphore.
675 * LSFL_RECOVER_WORK - dlm_ls_start() sets this to tell dlm_recoverd that it
676 * should begin recovery of the lockspace.
678 * LSFL_RUNNING - set when normal locking activity is enabled.
679 * dlm_ls_stop() clears this to tell dlm locking routines that they should
680 * quit what they are doing so recovery can run. dlm_recoverd sets
681 * this after recovery is finished.
684 #define LSFL_RECOVER_STOP 0
685 #define LSFL_RECOVER_DOWN 1
686 #define LSFL_RECOVER_LOCK 2
687 #define LSFL_RECOVER_WORK 3
688 #define LSFL_RUNNING 4
690 #define LSFL_RCOM_READY 5
691 #define LSFL_RCOM_WAIT 6
692 #define LSFL_UEVENT_WAIT 7
693 #define LSFL_TIMEWARN 8
694 #define LSFL_CB_DELAY 9
695 #define LSFL_NODIR 10
697 /* much of this is just saving user space pointers associated with the
698 lock that we pass back to the user lib with an ast */
700 struct dlm_user_args {
701 struct dlm_user_proc *proc; /* each process that opens the lockspace
702 device has private data
703 (dlm_user_proc) on the struct file,
704 the process's locks point back to it*/
705 struct dlm_lksb lksb;
706 struct dlm_lksb __user *user_lksb;
707 void __user *castparam;
708 void __user *castaddr;
709 void __user *bastparam;
710 void __user *bastaddr;
714 #define DLM_PROC_FLAGS_CLOSING 1
715 #define DLM_PROC_FLAGS_COMPAT 2
717 /* locks list is kept so we can remove all a process's locks when it
718 exits (or orphan those that are persistent) */
720 struct dlm_user_proc {
721 dlm_lockspace_t *lockspace;
722 unsigned long flags; /* DLM_PROC_FLAGS */
723 struct list_head asts;
724 spinlock_t asts_spin;
725 struct list_head locks;
726 spinlock_t locks_spin;
727 struct list_head unlocking;
728 wait_queue_head_t wait;
731 static inline int dlm_locking_stopped(struct dlm_ls *ls)
733 return !test_bit(LSFL_RUNNING, &ls->ls_flags);
736 static inline int dlm_recovery_stopped(struct dlm_ls *ls)
738 return test_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
741 static inline int dlm_no_directory(struct dlm_ls *ls)
743 return test_bit(LSFL_NODIR, &ls->ls_flags);
746 int dlm_netlink_init(void);
747 void dlm_netlink_exit(void);
748 void dlm_timeout_warn(struct dlm_lkb *lkb);
749 int dlm_plock_init(void);
750 void dlm_plock_exit(void);
752 #ifdef CONFIG_DLM_DEBUG
753 void dlm_register_debugfs(void);
754 void dlm_unregister_debugfs(void);
755 void dlm_create_debug_file(struct dlm_ls *ls);
756 void dlm_delete_debug_file(struct dlm_ls *ls);
757 void *dlm_create_debug_comms_file(int nodeid, void *data);
758 void dlm_delete_debug_comms_file(void *ctx);
760 static inline void dlm_register_debugfs(void) { }
761 static inline void dlm_unregister_debugfs(void) { }
762 static inline void dlm_create_debug_file(struct dlm_ls *ls) { }
763 static inline void dlm_delete_debug_file(struct dlm_ls *ls) { }
764 static inline void *dlm_create_debug_comms_file(int nodeid, void *data) { return NULL; }
765 static inline void dlm_delete_debug_comms_file(void *ctx) { }
768 #endif /* __DLM_INTERNAL_DOT_H__ */