xref: /openbmc/linux/fs/dlm/dlm_internal.h (revision 22246614)
1 /******************************************************************************
2 *******************************************************************************
3 **
4 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
5 **  Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
6 **
7 **  This copyrighted material is made available to anyone wishing to use,
8 **  modify, copy, or redistribute it subject to the terms and conditions
9 **  of the GNU General Public License v.2.
10 **
11 *******************************************************************************
12 ******************************************************************************/
13 
14 #ifndef __DLM_INTERNAL_DOT_H__
15 #define __DLM_INTERNAL_DOT_H__
16 
17 /*
18  * This is the main header file to be included in each DLM source file.
19  */
20 
21 #include <linux/module.h>
22 #include <linux/slab.h>
23 #include <linux/sched.h>
24 #include <linux/types.h>
25 #include <linux/ctype.h>
26 #include <linux/spinlock.h>
27 #include <linux/vmalloc.h>
28 #include <linux/list.h>
29 #include <linux/errno.h>
30 #include <linux/random.h>
31 #include <linux/delay.h>
32 #include <linux/socket.h>
33 #include <linux/kthread.h>
34 #include <linux/kobject.h>
35 #include <linux/kref.h>
36 #include <linux/kernel.h>
37 #include <linux/jhash.h>
38 #include <linux/miscdevice.h>
39 #include <linux/mutex.h>
40 #include <asm/uaccess.h>
41 
42 #include <linux/dlm.h>
43 #include "config.h"
44 
45 /* Size of the temp buffer midcomms allocates on the stack.
46    We try to make this large enough so most messages fit.
47    FIXME: should sctp make this unnecessary? */
48 
49 #define DLM_INBUF_LEN		148
50 
51 struct dlm_ls;
52 struct dlm_lkb;
53 struct dlm_rsb;
54 struct dlm_member;
55 struct dlm_lkbtable;
56 struct dlm_rsbtable;
57 struct dlm_dirtable;
58 struct dlm_direntry;
59 struct dlm_recover;
60 struct dlm_header;
61 struct dlm_message;
62 struct dlm_rcom;
63 struct dlm_mhandle;
64 
65 #define log_print(fmt, args...) \
66 	printk(KERN_ERR "dlm: "fmt"\n" , ##args)
67 #define log_error(ls, fmt, args...) \
68 	printk(KERN_ERR "dlm: %s: " fmt "\n", (ls)->ls_name , ##args)
69 
70 #define log_debug(ls, fmt, args...) \
71 do { \
72 	if (dlm_config.ci_log_debug) \
73 		printk(KERN_DEBUG "dlm: %s: " fmt "\n", \
74 		       (ls)->ls_name , ##args); \
75 } while (0)
76 
77 #define DLM_ASSERT(x, do) \
78 { \
79   if (!(x)) \
80   { \
81     printk(KERN_ERR "\nDLM:  Assertion failed on line %d of file %s\n" \
82                "DLM:  assertion:  \"%s\"\n" \
83                "DLM:  time = %lu\n", \
84                __LINE__, __FILE__, #x, jiffies); \
85     {do} \
86     printk("\n"); \
87     BUG(); \
88     panic("DLM:  Record message above and reboot.\n"); \
89   } \
90 }
91 
92 
93 struct dlm_direntry {
94 	struct list_head	list;
95 	uint32_t		master_nodeid;
96 	uint16_t		length;
97 	char			name[1];
98 };
99 
100 struct dlm_dirtable {
101 	struct list_head	list;
102 	rwlock_t		lock;
103 };
104 
105 struct dlm_rsbtable {
106 	struct list_head	list;
107 	struct list_head	toss;
108 	rwlock_t		lock;
109 };
110 
111 struct dlm_lkbtable {
112 	struct list_head	list;
113 	rwlock_t		lock;
114 	uint16_t		counter;
115 };
116 
117 /*
118  * Lockspace member (per node in a ls)
119  */
120 
121 struct dlm_member {
122 	struct list_head	list;
123 	int			nodeid;
124 	int			weight;
125 };
126 
127 /*
128  * Save and manage recovery state for a lockspace.
129  */
130 
131 struct dlm_recover {
132 	struct list_head	list;
133 	int			*nodeids;   /* nodeids of all members */
134 	int			node_count;
135 	int			*new;       /* nodeids of new members */
136 	int			new_count;
137 	uint64_t		seq;
138 };
139 
140 /*
141  * Pass input args to second stage locking function.
142  */
143 
144 struct dlm_args {
145 	uint32_t		flags;
146 	void			(*astfn) (void *astparam);
147 	void			*astparam;
148 	void			(*bastfn) (void *astparam, int mode);
149 	int			mode;
150 	struct dlm_lksb		*lksb;
151 	unsigned long		timeout;
152 };
153 
154 
155 /*
156  * Lock block
157  *
158  * A lock can be one of three types:
159  *
160  * local copy      lock is mastered locally
161  *                 (lkb_nodeid is zero and DLM_LKF_MSTCPY is not set)
162  * process copy    lock is mastered on a remote node
163  *                 (lkb_nodeid is non-zero and DLM_LKF_MSTCPY is not set)
164  * master copy     master node's copy of a lock owned by remote node
165  *                 (lkb_nodeid is non-zero and DLM_LKF_MSTCPY is set)
166  *
167  * lkb_exflags: a copy of the most recent flags arg provided to dlm_lock or
168  * dlm_unlock.  The dlm does not modify these or use any private flags in
169  * this field; it only contains DLM_LKF_ flags from dlm.h.  These flags
170  * are sent as-is to the remote master when the lock is remote.
171  *
172  * lkb_flags: internal dlm flags (DLM_IFL_ prefix) from dlm_internal.h.
173  * Some internal flags are shared between the master and process nodes;
174  * these shared flags are kept in the lower two bytes.  One of these
175  * flags set on the master copy will be propagated to the process copy
176  * and v.v.  Other internal flags are private to the master or process
177  * node (e.g. DLM_IFL_MSTCPY).  These are kept in the high two bytes.
178  *
179  * lkb_sbflags: status block flags.  These flags are copied directly into
180  * the caller's lksb.sb_flags prior to the dlm_lock/dlm_unlock completion
181  * ast.  All defined in dlm.h with DLM_SBF_ prefix.
182  *
183  * lkb_status: the lock status indicates which rsb queue the lock is
184  * on, grant, convert, or wait.  DLM_LKSTS_ WAITING/GRANTED/CONVERT
185  *
186  * lkb_wait_type: the dlm message type (DLM_MSG_ prefix) for which a
187  * reply is needed.  Only set when the lkb is on the lockspace waiters
188  * list awaiting a reply from a remote node.
189  *
190  * lkb_nodeid: when the lkb is a local copy, nodeid is 0; when the lkb
191  * is a master copy, nodeid specifies the remote lock holder, when the
192  * lkb is a process copy, the nodeid specifies the lock master.
193  */
194 
195 /* lkb_ast_type */
196 
197 #define AST_COMP		1
198 #define AST_BAST		2
199 
200 /* lkb_status */
201 
202 #define DLM_LKSTS_WAITING	1
203 #define DLM_LKSTS_GRANTED	2
204 #define DLM_LKSTS_CONVERT	3
205 
206 /* lkb_flags */
207 
208 #define DLM_IFL_MSTCPY		0x00010000
209 #define DLM_IFL_RESEND		0x00020000
210 #define DLM_IFL_DEAD		0x00040000
211 #define DLM_IFL_OVERLAP_UNLOCK  0x00080000
212 #define DLM_IFL_OVERLAP_CANCEL  0x00100000
213 #define DLM_IFL_ENDOFLIFE	0x00200000
214 #define DLM_IFL_WATCH_TIMEWARN	0x00400000
215 #define DLM_IFL_TIMEOUT_CANCEL	0x00800000
216 #define DLM_IFL_DEADLOCK_CANCEL	0x01000000
217 #define DLM_IFL_USER		0x00000001
218 #define DLM_IFL_ORPHAN		0x00000002
219 
220 struct dlm_lkb {
221 	struct dlm_rsb		*lkb_resource;	/* the rsb */
222 	struct kref		lkb_ref;
223 	int			lkb_nodeid;	/* copied from rsb */
224 	int			lkb_ownpid;	/* pid of lock owner */
225 	uint32_t		lkb_id;		/* our lock ID */
226 	uint32_t		lkb_remid;	/* lock ID on remote partner */
227 	uint32_t		lkb_exflags;	/* external flags from caller */
228 	uint32_t		lkb_sbflags;	/* lksb flags */
229 	uint32_t		lkb_flags;	/* internal flags */
230 	uint32_t		lkb_lvbseq;	/* lvb sequence number */
231 
232 	int8_t			lkb_status;     /* granted, waiting, convert */
233 	int8_t			lkb_rqmode;	/* requested lock mode */
234 	int8_t			lkb_grmode;	/* granted lock mode */
235 	int8_t			lkb_bastmode;	/* requested mode */
236 	int8_t			lkb_highbast;	/* highest mode bast sent for */
237 	int8_t			lkb_wait_type;	/* type of reply waiting for */
238 	int8_t			lkb_wait_count;
239 	int8_t			lkb_ast_type;	/* type of ast queued for */
240 
241 	struct list_head	lkb_idtbl_list;	/* lockspace lkbtbl */
242 	struct list_head	lkb_statequeue;	/* rsb g/c/w list */
243 	struct list_head	lkb_rsb_lookup;	/* waiting for rsb lookup */
244 	struct list_head	lkb_wait_reply;	/* waiting for remote reply */
245 	struct list_head	lkb_astqueue;	/* need ast to be sent */
246 	struct list_head	lkb_ownqueue;	/* list of locks for a process */
247 	struct list_head	lkb_time_list;
248 	unsigned long		lkb_timestamp;
249 	unsigned long		lkb_timeout_cs;
250 
251 	char			*lkb_lvbptr;
252 	struct dlm_lksb		*lkb_lksb;      /* caller's status block */
253 	void			(*lkb_astfn) (void *astparam);
254 	void			(*lkb_bastfn) (void *astparam, int mode);
255 	union {
256 		void			*lkb_astparam;	/* caller's ast arg */
257 		struct dlm_user_args	*lkb_ua;
258 	};
259 };
260 
261 
262 struct dlm_rsb {
263 	struct dlm_ls		*res_ls;	/* the lockspace */
264 	struct kref		res_ref;
265 	struct mutex		res_mutex;
266 	unsigned long		res_flags;
267 	int			res_length;	/* length of rsb name */
268 	int			res_nodeid;
269 	uint32_t                res_lvbseq;
270 	uint32_t		res_hash;
271 	uint32_t		res_bucket;	/* rsbtbl */
272 	unsigned long		res_toss_time;
273 	uint32_t		res_first_lkid;
274 	struct list_head	res_lookup;	/* lkbs waiting on first */
275 	struct list_head	res_hashchain;	/* rsbtbl */
276 	struct list_head	res_grantqueue;
277 	struct list_head	res_convertqueue;
278 	struct list_head	res_waitqueue;
279 
280 	struct list_head	res_root_list;	    /* used for recovery */
281 	struct list_head	res_recover_list;   /* used for recovery */
282 	int			res_recover_locks_count;
283 
284 	char			*res_lvbptr;
285 	char			res_name[1];
286 };
287 
288 /* find_rsb() flags */
289 
290 #define R_MASTER		1	/* only return rsb if it's a master */
291 #define R_CREATE		2	/* create/add rsb if not found */
292 
293 /* rsb_flags */
294 
295 enum rsb_flags {
296 	RSB_MASTER_UNCERTAIN,
297 	RSB_VALNOTVALID,
298 	RSB_VALNOTVALID_PREV,
299 	RSB_NEW_MASTER,
300 	RSB_NEW_MASTER2,
301 	RSB_RECOVER_CONVERT,
302 	RSB_LOCKS_PURGED,
303 };
304 
305 static inline void rsb_set_flag(struct dlm_rsb *r, enum rsb_flags flag)
306 {
307 	__set_bit(flag, &r->res_flags);
308 }
309 
310 static inline void rsb_clear_flag(struct dlm_rsb *r, enum rsb_flags flag)
311 {
312 	__clear_bit(flag, &r->res_flags);
313 }
314 
315 static inline int rsb_flag(struct dlm_rsb *r, enum rsb_flags flag)
316 {
317 	return test_bit(flag, &r->res_flags);
318 }
319 
320 
321 /* dlm_header is first element of all structs sent between nodes */
322 
323 #define DLM_HEADER_MAJOR	0x00030000
324 #define DLM_HEADER_MINOR	0x00000000
325 
326 #define DLM_MSG			1
327 #define DLM_RCOM		2
328 
329 struct dlm_header {
330 	uint32_t		h_version;
331 	uint32_t		h_lockspace;
332 	uint32_t		h_nodeid;	/* nodeid of sender */
333 	uint16_t		h_length;
334 	uint8_t			h_cmd;		/* DLM_MSG, DLM_RCOM */
335 	uint8_t			h_pad;
336 };
337 
338 
339 #define DLM_MSG_REQUEST		1
340 #define DLM_MSG_CONVERT		2
341 #define DLM_MSG_UNLOCK		3
342 #define DLM_MSG_CANCEL		4
343 #define DLM_MSG_REQUEST_REPLY	5
344 #define DLM_MSG_CONVERT_REPLY	6
345 #define DLM_MSG_UNLOCK_REPLY	7
346 #define DLM_MSG_CANCEL_REPLY	8
347 #define DLM_MSG_GRANT		9
348 #define DLM_MSG_BAST		10
349 #define DLM_MSG_LOOKUP		11
350 #define DLM_MSG_REMOVE		12
351 #define DLM_MSG_LOOKUP_REPLY	13
352 #define DLM_MSG_PURGE		14
353 
354 struct dlm_message {
355 	struct dlm_header	m_header;
356 	uint32_t		m_type;		/* DLM_MSG_ */
357 	uint32_t		m_nodeid;
358 	uint32_t		m_pid;
359 	uint32_t		m_lkid;		/* lkid on sender */
360 	uint32_t		m_remid;	/* lkid on receiver */
361 	uint32_t		m_parent_lkid;
362 	uint32_t		m_parent_remid;
363 	uint32_t		m_exflags;
364 	uint32_t		m_sbflags;
365 	uint32_t		m_flags;
366 	uint32_t		m_lvbseq;
367 	uint32_t		m_hash;
368 	int			m_status;
369 	int			m_grmode;
370 	int			m_rqmode;
371 	int			m_bastmode;
372 	int			m_asts;
373 	int			m_result;	/* 0 or -EXXX */
374 	char			m_extra[0];	/* name or lvb */
375 };
376 
377 
378 #define DLM_RS_NODES		0x00000001
379 #define DLM_RS_NODES_ALL	0x00000002
380 #define DLM_RS_DIR		0x00000004
381 #define DLM_RS_DIR_ALL		0x00000008
382 #define DLM_RS_LOCKS		0x00000010
383 #define DLM_RS_LOCKS_ALL	0x00000020
384 #define DLM_RS_DONE		0x00000040
385 #define DLM_RS_DONE_ALL		0x00000080
386 
387 #define DLM_RCOM_STATUS		1
388 #define DLM_RCOM_NAMES		2
389 #define DLM_RCOM_LOOKUP		3
390 #define DLM_RCOM_LOCK		4
391 #define DLM_RCOM_STATUS_REPLY	5
392 #define DLM_RCOM_NAMES_REPLY	6
393 #define DLM_RCOM_LOOKUP_REPLY	7
394 #define DLM_RCOM_LOCK_REPLY	8
395 
396 struct dlm_rcom {
397 	struct dlm_header	rc_header;
398 	uint32_t		rc_type;	/* DLM_RCOM_ */
399 	int			rc_result;	/* multi-purpose */
400 	uint64_t		rc_id;		/* match reply with request */
401 	uint64_t		rc_seq;		/* sender's ls_recover_seq */
402 	uint64_t		rc_seq_reply;	/* remote ls_recover_seq */
403 	char			rc_buf[0];
404 };
405 
406 union dlm_packet {
407 	struct dlm_header	header;		/* common to other two */
408 	struct dlm_message	message;
409 	struct dlm_rcom		rcom;
410 };
411 
412 struct rcom_config {
413 	__le32			rf_lvblen;
414 	__le32			rf_lsflags;
415 	__le64			rf_unused;
416 };
417 
418 struct rcom_lock {
419 	__le32			rl_ownpid;
420 	__le32			rl_lkid;
421 	__le32			rl_remid;
422 	__le32			rl_parent_lkid;
423 	__le32			rl_parent_remid;
424 	__le32			rl_exflags;
425 	__le32			rl_flags;
426 	__le32			rl_lvbseq;
427 	__le32			rl_result;
428 	int8_t			rl_rqmode;
429 	int8_t			rl_grmode;
430 	int8_t			rl_status;
431 	int8_t			rl_asts;
432 	__le16			rl_wait_type;
433 	__le16			rl_namelen;
434 	char			rl_name[DLM_RESNAME_MAXLEN];
435 	char			rl_lvb[0];
436 };
437 
438 struct dlm_ls {
439 	struct list_head	ls_list;	/* list of lockspaces */
440 	dlm_lockspace_t		*ls_local_handle;
441 	uint32_t		ls_global_id;	/* global unique lockspace ID */
442 	uint32_t		ls_exflags;
443 	int			ls_lvblen;
444 	int			ls_count;	/* reference count */
445 	unsigned long		ls_flags;	/* LSFL_ */
446 	struct kobject		ls_kobj;
447 
448 	struct dlm_rsbtable	*ls_rsbtbl;
449 	uint32_t		ls_rsbtbl_size;
450 
451 	struct dlm_lkbtable	*ls_lkbtbl;
452 	uint32_t		ls_lkbtbl_size;
453 
454 	struct dlm_dirtable	*ls_dirtbl;
455 	uint32_t		ls_dirtbl_size;
456 
457 	struct mutex		ls_waiters_mutex;
458 	struct list_head	ls_waiters;	/* lkbs needing a reply */
459 
460 	struct mutex		ls_orphans_mutex;
461 	struct list_head	ls_orphans;
462 
463 	struct mutex		ls_timeout_mutex;
464 	struct list_head	ls_timeout;
465 
466 	struct list_head	ls_nodes;	/* current nodes in ls */
467 	struct list_head	ls_nodes_gone;	/* dead node list, recovery */
468 	int			ls_num_nodes;	/* number of nodes in ls */
469 	int			ls_low_nodeid;
470 	int			ls_total_weight;
471 	int			*ls_node_array;
472 	gfp_t			ls_allocation;
473 
474 	struct dlm_rsb		ls_stub_rsb;	/* for returning errors */
475 	struct dlm_lkb		ls_stub_lkb;	/* for returning errors */
476 	struct dlm_message	ls_stub_ms;	/* for faking a reply */
477 
478 	struct dentry		*ls_debug_rsb_dentry; /* debugfs */
479 	struct dentry		*ls_debug_waiters_dentry; /* debugfs */
480 	struct dentry		*ls_debug_locks_dentry; /* debugfs */
481 
482 	wait_queue_head_t	ls_uevent_wait;	/* user part of join/leave */
483 	int			ls_uevent_result;
484 	struct completion	ls_members_done;
485 	int			ls_members_result;
486 
487 	struct miscdevice       ls_device;
488 
489 	/* recovery related */
490 
491 	struct timer_list	ls_timer;
492 	struct task_struct	*ls_recoverd_task;
493 	struct mutex		ls_recoverd_active;
494 	spinlock_t		ls_recover_lock;
495 	unsigned long		ls_recover_begin; /* jiffies timestamp */
496 	uint32_t		ls_recover_status; /* DLM_RS_ */
497 	uint64_t		ls_recover_seq;
498 	struct dlm_recover	*ls_recover_args;
499 	struct rw_semaphore	ls_in_recovery;	/* block local requests */
500 	struct rw_semaphore	ls_recv_active;	/* block dlm_recv */
501 	struct list_head	ls_requestqueue;/* queue remote requests */
502 	struct mutex		ls_requestqueue_mutex;
503 	struct dlm_rcom		*ls_recover_buf;
504 	int			ls_recover_nodeid; /* for debugging */
505 	uint64_t		ls_rcom_seq;
506 	spinlock_t		ls_rcom_spin;
507 	struct list_head	ls_recover_list;
508 	spinlock_t		ls_recover_list_lock;
509 	int			ls_recover_list_count;
510 	wait_queue_head_t	ls_wait_general;
511 	struct mutex		ls_clear_proc_locks;
512 
513 	struct list_head	ls_root_list;	/* root resources */
514 	struct rw_semaphore	ls_root_sem;	/* protect root_list */
515 
516 	int			ls_namelen;
517 	char			ls_name[1];
518 };
519 
520 #define LSFL_WORK		0
521 #define LSFL_RUNNING		1
522 #define LSFL_RECOVERY_STOP	2
523 #define LSFL_RCOM_READY		3
524 #define LSFL_RCOM_WAIT		4
525 #define LSFL_UEVENT_WAIT	5
526 #define LSFL_TIMEWARN		6
527 
528 /* much of this is just saving user space pointers associated with the
529    lock that we pass back to the user lib with an ast */
530 
531 struct dlm_user_args {
532 	struct dlm_user_proc	*proc; /* each process that opens the lockspace
533 					  device has private data
534 					  (dlm_user_proc) on the struct file,
535 					  the process's locks point back to it*/
536 	struct dlm_lksb		lksb;
537 	int			old_mode;
538 	int			update_user_lvb;
539 	struct dlm_lksb __user	*user_lksb;
540 	void __user		*castparam;
541 	void __user		*castaddr;
542 	void __user		*bastparam;
543 	void __user		*bastaddr;
544 	uint64_t		xid;
545 };
546 
547 #define DLM_PROC_FLAGS_CLOSING 1
548 #define DLM_PROC_FLAGS_COMPAT  2
549 
550 /* locks list is kept so we can remove all a process's locks when it
551    exits (or orphan those that are persistent) */
552 
553 struct dlm_user_proc {
554 	dlm_lockspace_t		*lockspace;
555 	unsigned long		flags; /* DLM_PROC_FLAGS */
556 	struct list_head	asts;
557 	spinlock_t		asts_spin;
558 	struct list_head	locks;
559 	spinlock_t		locks_spin;
560 	struct list_head	unlocking;
561 	wait_queue_head_t	wait;
562 };
563 
564 static inline int dlm_locking_stopped(struct dlm_ls *ls)
565 {
566 	return !test_bit(LSFL_RUNNING, &ls->ls_flags);
567 }
568 
569 static inline int dlm_recovery_stopped(struct dlm_ls *ls)
570 {
571 	return test_bit(LSFL_RECOVERY_STOP, &ls->ls_flags);
572 }
573 
574 static inline int dlm_no_directory(struct dlm_ls *ls)
575 {
576 	return (ls->ls_exflags & DLM_LSFL_NODIR) ? 1 : 0;
577 }
578 
579 int dlm_netlink_init(void);
580 void dlm_netlink_exit(void);
581 void dlm_timeout_warn(struct dlm_lkb *lkb);
582 int dlm_plock_init(void);
583 void dlm_plock_exit(void);
584 
585 #ifdef CONFIG_DLM_DEBUG
586 int dlm_register_debugfs(void);
587 void dlm_unregister_debugfs(void);
588 int dlm_create_debug_file(struct dlm_ls *ls);
589 void dlm_delete_debug_file(struct dlm_ls *ls);
590 #else
591 static inline int dlm_register_debugfs(void) { return 0; }
592 static inline void dlm_unregister_debugfs(void) { }
593 static inline int dlm_create_debug_file(struct dlm_ls *ls) { return 0; }
594 static inline void dlm_delete_debug_file(struct dlm_ls *ls) { }
595 #endif
596 
597 #endif				/* __DLM_INTERNAL_DOT_H__ */
598 
599