fs/dlm/dlm_internal.h

   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/semaphore.h>
  41 #include <asm/uaccess.h>
  42
  43 #include <linux/dlm.h>
  44 #include "config.h"
  45
  46 #define DLM_LOCKSPACE_LEN       64
  47
  48 /* Size of the temp buffer midcomms allocates on the stack.
  49    We try to make this large enough so most messages fit.
  50    FIXME: should sctp make this unnecessary? */
  51
  52 #define DLM_INBUF_LEN           148
  53
  54 struct dlm_ls;
  55 struct dlm_lkb;
  56 struct dlm_rsb;
  57 struct dlm_member;
  58 struct dlm_lkbtable;
  59 struct dlm_rsbtable;
  60 struct dlm_dirtable;
  61 struct dlm_direntry;
  62 struct dlm_recover;
  63 struct dlm_header;
  64 struct dlm_message;
  65 struct dlm_rcom;
  66 struct dlm_mhandle;
  67
  68 #define log_print(fmt, args...) \
  69         printk(KERN_ERR "dlm: "fmt"\n" , ##args)
  70 #define log_error(ls, fmt, args...) \
  71         printk(KERN_ERR "dlm: %s: " fmt "\n", (ls)->ls_name , ##args)
  72
  73 #define log_debug(ls, fmt, args...) \
  74 do { \
  75         if (dlm_config.ci_log_debug) \
  76                 printk(KERN_DEBUG "dlm: %s: " fmt "\n", \
  77                        (ls)->ls_name , ##args); \
  78 } while (0)
  79
  80 #define DLM_ASSERT(x, do) \
  81 { \
  82   if (!(x)) \
  83   { \
  84     printk(KERN_ERR "\nDLM:  Assertion failed on line %d of file %s\n" \
  85                "DLM:  assertion:  \"%s\"\n" \
  86                "DLM:  time = %lu\n", \
  87                __LINE__, __FILE__, #x, jiffies); \
  88     {do} \
  89     printk("\n"); \
  90     BUG(); \
  91     panic("DLM:  Record message above and reboot.\n"); \
  92   } \
  93 }
  94
  95 #define DLM_FAKE_USER_AST ERR_PTR(-EINVAL)
  96
  97
  98 struct dlm_direntry {
  99         struct list_head        list;
 100         uint32_t                master_nodeid;
 101         uint16_t                length;
 102         char                    name[1];
 103 };
 104
 105 struct dlm_dirtable {
 106         struct list_head        list;
 107         rwlock_t                lock;
 108 };
 109
 110 struct dlm_rsbtable {
 111         struct list_head        list;
 112         struct list_head        toss;
 113         rwlock_t                lock;
 114 };
 115
 116 struct dlm_lkbtable {
 117         struct list_head        list;
 118         rwlock_t                lock;
 119         uint16_t                counter;
 120 };
 121
 122 /*
 123  * Lockspace member (per node in a ls)
 124  */
 125
 126 struct dlm_member {
 127         struct list_head        list;
 128         int                     nodeid;
 129         int                     weight;
 130 };
 131
 132 /*
 133  * Save and manage recovery state for a lockspace.
 134  */
 135
 136 struct dlm_recover {
 137         struct list_head        list;
 138         int                     *nodeids;
 139         int                     node_count;
 140         uint64_t                seq;
 141 };
 142
 143 /*
 144  * Pass input args to second stage locking function.
 145  */
 146
 147 struct dlm_args {
 148         uint32_t                flags;
 149         void                    *astaddr;
 150         long                    astparam;
 151         void                    *bastaddr;
 152         int                     mode;
 153         struct dlm_lksb         *lksb;
 154         unsigned long           timeout;
 155 };
 156
 157
 158 /*
 159  * Lock block
 160  *
 161  * A lock can be one of three types:
 162  *
 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)
 169  *
 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.
 174  *
 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.
 181  *
 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.
 185  *
 186  * lkb_status: the lock status indicates which rsb queue the lock is
 187  * on, grant, convert, or wait.  DLM_LKSTS_ WAITING/GRANTED/CONVERT
 188  *
 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.
 192  *
 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.
 196  */
 197
 198 /* lkb_ast_type */
 199
 200 #define AST_COMP                1
 201 #define AST_BAST                2
 202
 203 /* lkb_status */
 204
 205 #define DLM_LKSTS_WAITING       1
 206 #define DLM_LKSTS_GRANTED       2
 207 #define DLM_LKSTS_CONVERT       3
 208
 209 /* lkb_flags */
 210
 211 #define DLM_IFL_MSTCPY          0x00010000
 212 #define DLM_IFL_RESEND          0x00020000
 213 #define DLM_IFL_DEAD            0x00040000
 214 #define DLM_IFL_OVERLAP_UNLOCK  0x00080000
 215 #define DLM_IFL_OVERLAP_CANCEL  0x00100000
 216 #define DLM_IFL_ENDOFLIFE       0x00200000
 217 #define DLM_IFL_WATCH_TIMEWARN  0x00400000
 218 #define DLM_IFL_TIMEOUT_CANCEL  0x00800000
 219 #define DLM_IFL_DEADLOCK_CANCEL 0x01000000
 220 #define DLM_IFL_USER            0x00000001
 221 #define DLM_IFL_ORPHAN          0x00000002
 222
 223 struct dlm_lkb {
 224         struct dlm_rsb          *lkb_resource;  /* the rsb */
 225         struct kref             lkb_ref;
 226         int                     lkb_nodeid;     /* copied from rsb */
 227         int                     lkb_ownpid;     /* pid of lock owner */
 228         uint32_t                lkb_id;         /* our lock ID */
 229         uint32_t                lkb_remid;      /* lock ID on remote partner */
 230         uint32_t                lkb_exflags;    /* external flags from caller */
 231         uint32_t                lkb_sbflags;    /* lksb flags */
 232         uint32_t                lkb_flags;      /* internal flags */
 233         uint32_t                lkb_lvbseq;     /* lvb sequence number */
 234
 235         int8_t                  lkb_status;     /* granted, waiting, convert */
 236         int8_t                  lkb_rqmode;     /* requested lock mode */
 237         int8_t                  lkb_grmode;     /* granted lock mode */
 238         int8_t                  lkb_bastmode;   /* requested mode */
 239         int8_t                  lkb_highbast;   /* highest mode bast sent for */
 240         int8_t                  lkb_wait_type;  /* type of reply waiting for */
 241         int8_t                  lkb_wait_count;
 242         int8_t                  lkb_ast_type;   /* type of ast queued for */
 243
 244         struct list_head        lkb_idtbl_list; /* lockspace lkbtbl */
 245         struct list_head        lkb_statequeue; /* rsb g/c/w list */
 246         struct list_head        lkb_rsb_lookup; /* waiting for rsb lookup */
 247         struct list_head        lkb_wait_reply; /* waiting for remote reply */
 248         struct list_head        lkb_astqueue;   /* need ast to be sent */
 249         struct list_head        lkb_ownqueue;   /* list of locks for a process */
 250         struct list_head        lkb_time_list;
 251         unsigned long           lkb_timestamp;
 252         unsigned long           lkb_timeout_cs;
 253
 254         char                    *lkb_lvbptr;
 255         struct dlm_lksb         *lkb_lksb;      /* caller's status block */
 256         void                    *lkb_astaddr;   /* caller's ast function */
 257         void                    *lkb_bastaddr;  /* caller's bast function */
 258         long                    lkb_astparam;   /* caller's ast arg */
 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 struct rcom_config {
 407         uint32_t                rf_lvblen;
 408         uint32_t                rf_lsflags;
 409         uint64_t                rf_unused;
 410 };
 411
 412 struct rcom_lock {
 413         uint32_t                rl_ownpid;
 414         uint32_t                rl_lkid;
 415         uint32_t                rl_remid;
 416         uint32_t                rl_parent_lkid;
 417         uint32_t                rl_parent_remid;
 418         uint32_t                rl_exflags;
 419         uint32_t                rl_flags;
 420         uint32_t                rl_lvbseq;
 421         int                     rl_result;
 422         int8_t                  rl_rqmode;
 423         int8_t                  rl_grmode;
 424         int8_t                  rl_status;
 425         int8_t                  rl_asts;
 426         uint16_t                rl_wait_type;
 427         uint16_t                rl_namelen;
 428         char                    rl_name[DLM_RESNAME_MAXLEN];
 429         char                    rl_lvb[0];
 430 };
 431
 432 struct dlm_ls {
 433         struct list_head        ls_list;        /* list of lockspaces */
 434         dlm_lockspace_t         *ls_local_handle;
 435         uint32_t                ls_global_id;   /* global unique lockspace ID */
 436         uint32_t                ls_exflags;
 437         int                     ls_lvblen;
 438         int                     ls_count;       /* reference count */
 439         unsigned long           ls_flags;       /* LSFL_ */
 440         struct kobject          ls_kobj;
 441
 442         struct dlm_rsbtable     *ls_rsbtbl;
 443         uint32_t                ls_rsbtbl_size;
 444
 445         struct dlm_lkbtable     *ls_lkbtbl;
 446         uint32_t                ls_lkbtbl_size;
 447
 448         struct dlm_dirtable     *ls_dirtbl;
 449         uint32_t                ls_dirtbl_size;
 450
 451         struct mutex            ls_waiters_mutex;
 452         struct list_head        ls_waiters;     /* lkbs needing a reply */
 453
 454         struct mutex            ls_orphans_mutex;
 455         struct list_head        ls_orphans;
 456
 457         struct mutex            ls_timeout_mutex;
 458         struct list_head        ls_timeout;
 459
 460         struct list_head        ls_nodes;       /* current nodes in ls */
 461         struct list_head        ls_nodes_gone;  /* dead node list, recovery */
 462         int                     ls_num_nodes;   /* number of nodes in ls */
 463         int                     ls_low_nodeid;
 464         int                     ls_total_weight;
 465         int                     *ls_node_array;
 466         gfp_t                   ls_allocation;
 467
 468         struct dlm_rsb          ls_stub_rsb;    /* for returning errors */
 469         struct dlm_lkb          ls_stub_lkb;    /* for returning errors */
 470         struct dlm_message      ls_stub_ms;     /* for faking a reply */
 471
 472         struct dentry           *ls_debug_rsb_dentry; /* debugfs */
 473         struct dentry           *ls_debug_waiters_dentry; /* debugfs */
 474         struct dentry           *ls_debug_locks_dentry; /* debugfs */
 475
 476         wait_queue_head_t       ls_uevent_wait; /* user part of join/leave */
 477         int                     ls_uevent_result;
 478         struct completion       ls_members_done;
 479         int                     ls_members_result;
 480
 481         struct miscdevice       ls_device;
 482
 483         /* recovery related */
 484
 485         struct timer_list       ls_timer;
 486         struct task_struct      *ls_recoverd_task;
 487         struct mutex            ls_recoverd_active;
 488         spinlock_t              ls_recover_lock;
 489         unsigned long           ls_recover_begin; /* jiffies timestamp */
 490         uint32_t                ls_recover_status; /* DLM_RS_ */
 491         uint64_t                ls_recover_seq;
 492         struct dlm_recover      *ls_recover_args;
 493         struct rw_semaphore     ls_in_recovery; /* block local requests */
 494         struct list_head        ls_requestqueue;/* queue remote requests */
 495         struct mutex            ls_requestqueue_mutex;
 496         char                    *ls_recover_buf;
 497         int                     ls_recover_nodeid; /* for debugging */
 498         uint64_t                ls_rcom_seq;
 499         spinlock_t              ls_rcom_spin;
 500         struct list_head        ls_recover_list;
 501         spinlock_t              ls_recover_list_lock;
 502         int                     ls_recover_list_count;
 503         wait_queue_head_t       ls_wait_general;
 504         struct mutex            ls_clear_proc_locks;
 505
 506         struct list_head        ls_root_list;   /* root resources */
 507         struct rw_semaphore     ls_root_sem;    /* protect root_list */
 508
 509         int                     ls_namelen;
 510         char                    ls_name[1];
 511 };
 512
 513 #define LSFL_WORK               0
 514 #define LSFL_RUNNING            1
 515 #define LSFL_RECOVERY_STOP      2
 516 #define LSFL_RCOM_READY         3
 517 #define LSFL_RCOM_WAIT          4
 518 #define LSFL_UEVENT_WAIT        5
 519 #define LSFL_TIMEWARN           6
 520
 521 /* much of this is just saving user space pointers associated with the
 522    lock that we pass back to the user lib with an ast */
 523
 524 struct dlm_user_args {
 525         struct dlm_user_proc    *proc; /* each process that opens the lockspace
 526                                           device has private data
 527                                           (dlm_user_proc) on the struct file,
 528                                           the process's locks point back to it*/
 529         struct dlm_lksb         lksb;
 530         int                     old_mode;
 531         int                     update_user_lvb;
 532         struct dlm_lksb __user  *user_lksb;
 533         void __user             *castparam;
 534         void __user             *castaddr;
 535         void __user             *bastparam;
 536         void __user             *bastaddr;
 537         uint64_t                xid;
 538 };
 539
 540 #define DLM_PROC_FLAGS_CLOSING 1
 541 #define DLM_PROC_FLAGS_COMPAT  2
 542
 543 /* locks list is kept so we can remove all a process's locks when it
 544    exits (or orphan those that are persistent) */
 545
 546 struct dlm_user_proc {
 547         dlm_lockspace_t         *lockspace;
 548         unsigned long           flags; /* DLM_PROC_FLAGS */
 549         struct list_head        asts;
 550         spinlock_t              asts_spin;
 551         struct list_head        locks;
 552         spinlock_t              locks_spin;
 553         struct list_head        unlocking;
 554         wait_queue_head_t       wait;
 555 };
 556
 557 static inline int dlm_locking_stopped(struct dlm_ls *ls)
 558 {
 559         return !test_bit(LSFL_RUNNING, &ls->ls_flags);
 560 }
 561
 562 static inline int dlm_recovery_stopped(struct dlm_ls *ls)
 563 {
 564         return test_bit(LSFL_RECOVERY_STOP, &ls->ls_flags);
 565 }
 566
 567 static inline int dlm_no_directory(struct dlm_ls *ls)
 568 {
 569         return (ls->ls_exflags & DLM_LSFL_NODIR) ? 1 : 0;
 570 }
 571
 572 #endif                          /* __DLM_INTERNAL_DOT_H__ */
 573