fs/gfs2/glock.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2006 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License v.2.
   8  */
   9
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/completion.h>
  14 #include <linux/buffer_head.h>
  15 #include <linux/delay.h>
  16 #include <linux/sort.h>
  17 #include <linux/jhash.h>
  18 #include <linux/kref.h>
  19 #include <linux/kallsyms.h>
  20 #include <linux/gfs2_ondisk.h>
  21 #include <asm/uaccess.h>
  22
  23 #include "gfs2.h"
  24 #include "lm_interface.h"
  25 #include "incore.h"
  26 #include "glock.h"
  27 #include "glops.h"
  28 #include "inode.h"
  29 #include "lm.h"
  30 #include "lops.h"
  31 #include "meta_io.h"
  32 #include "quota.h"
  33 #include "super.h"
  34 #include "util.h"
  35
  36 /*  Must be kept in sync with the beginning of struct gfs2_glock  */
  37 struct glock_plug {
  38         struct list_head gl_list;
  39         unsigned long gl_flags;
  40 };
  41
  42 struct greedy {
  43         struct gfs2_holder gr_gh;
  44         struct work_struct gr_work;
  45 };
  46
  47 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  48
  49 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp);
  50 static int dump_glock(struct gfs2_glock *gl);
  51
  52 /**
  53  * relaxed_state_ok - is a requested lock compatible with the current lock mode?
  54  * @actual: the current state of the lock
  55  * @requested: the lock state that was requested by the caller
  56  * @flags: the modifier flags passed in by the caller
  57  *
  58  * Returns: 1 if the locks are compatible, 0 otherwise
  59  */
  60
  61 static inline int relaxed_state_ok(unsigned int actual, unsigned requested,
  62                                    int flags)
  63 {
  64         if (actual == requested)
  65                 return 1;
  66
  67         if (flags & GL_EXACT)
  68                 return 0;
  69
  70         if (actual == LM_ST_EXCLUSIVE && requested == LM_ST_SHARED)
  71                 return 1;
  72
  73         if (actual != LM_ST_UNLOCKED && (flags & LM_FLAG_ANY))
  74                 return 1;
  75
  76         return 0;
  77 }
  78
  79 /**
  80  * gl_hash() - Turn glock number into hash bucket number
  81  * @lock: The glock number
  82  *
  83  * Returns: The number of the corresponding hash bucket
  84  */
  85
  86 static unsigned int gl_hash(struct lm_lockname *name)
  87 {
  88         unsigned int h;
  89
  90         h = jhash(&name->ln_number, sizeof(uint64_t), 0);
  91         h = jhash(&name->ln_type, sizeof(unsigned int), h);
  92         h &= GFS2_GL_HASH_MASK;
  93
  94         return h;
  95 }
  96
  97 /**
  98  * glock_free() - Perform a few checks and then release struct gfs2_glock
  99  * @gl: The glock to release
 100  *
 101  * Also calls lock module to release its internal structure for this glock.
 102  *
 103  */
 104
 105 static void glock_free(struct gfs2_glock *gl)
 106 {
 107         struct gfs2_sbd *sdp = gl->gl_sbd;
 108         struct inode *aspace = gl->gl_aspace;
 109
 110         gfs2_lm_put_lock(sdp, gl->gl_lock);
 111
 112         if (aspace)
 113                 gfs2_aspace_put(aspace);
 114
 115         kmem_cache_free(gfs2_glock_cachep, gl);
 116 }
 117
 118 /**
 119  * gfs2_glock_hold() - increment reference count on glock
 120  * @gl: The glock to hold
 121  *
 122  */
 123
 124 void gfs2_glock_hold(struct gfs2_glock *gl)
 125 {
 126         kref_get(&gl->gl_ref);
 127 }
 128
 129 /* All work is done after the return from kref_put() so we
 130    can release the write_lock before the free. */
 131
 132 static void kill_glock(struct kref *kref)
 133 {
 134         struct gfs2_glock *gl = container_of(kref, struct gfs2_glock, gl_ref);
 135         struct gfs2_sbd *sdp = gl->gl_sbd;
 136
 137         gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
 138         gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
 139         gfs2_assert(sdp, list_empty(&gl->gl_holders));
 140         gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
 141         gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
 142         gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
 143 }
 144
 145 /**
 146  * gfs2_glock_put() - Decrement reference count on glock
 147  * @gl: The glock to put
 148  *
 149  */
 150
 151 int gfs2_glock_put(struct gfs2_glock *gl)
 152 {
 153         struct gfs2_sbd *sdp = gl->gl_sbd;
 154         struct gfs2_gl_hash_bucket *bucket = gl->gl_bucket;
 155         int rv = 0;
 156
 157         mutex_lock(&sdp->sd_invalidate_inodes_mutex);
 158
 159         write_lock(&bucket->hb_lock);
 160         if (kref_put(&gl->gl_ref, kill_glock)) {
 161                 list_del_init(&gl->gl_list);
 162                 write_unlock(&bucket->hb_lock);
 163                 BUG_ON(spin_is_locked(&gl->gl_spin));
 164                 glock_free(gl);
 165                 rv = 1;
 166                 goto out;
 167         }
 168         write_unlock(&bucket->hb_lock);
 169  out:
 170         mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
 171         return rv;
 172 }
 173
 174 /**
 175  * queue_empty - check to see if a glock's queue is empty
 176  * @gl: the glock
 177  * @head: the head of the queue to check
 178  *
 179  * This function protects the list in the event that a process already
 180  * has a holder on the list and is adding a second holder for itself.
 181  * The glmutex lock is what generally prevents processes from working
 182  * on the same glock at once, but the special case of adding a second
 183  * holder for yourself ("recursive" locking) doesn't involve locking
 184  * glmutex, making the spin lock necessary.
 185  *
 186  * Returns: 1 if the queue is empty
 187  */
 188
 189 static inline int queue_empty(struct gfs2_glock *gl, struct list_head *head)
 190 {
 191         int empty;
 192         spin_lock(&gl->gl_spin);
 193         empty = list_empty(head);
 194         spin_unlock(&gl->gl_spin);
 195         return empty;
 196 }
 197
 198 /**
 199  * search_bucket() - Find struct gfs2_glock by lock number
 200  * @bucket: the bucket to search
 201  * @name: The lock name
 202  *
 203  * Returns: NULL, or the struct gfs2_glock with the requested number
 204  */
 205
 206 static struct gfs2_glock *search_bucket(struct gfs2_gl_hash_bucket *bucket,
 207                                         struct lm_lockname *name)
 208 {
 209         struct gfs2_glock *gl;
 210
 211         list_for_each_entry(gl, &bucket->hb_list, gl_list) {
 212                 if (test_bit(GLF_PLUG, &gl->gl_flags))
 213                         continue;
 214                 if (!lm_name_equal(&gl->gl_name, name))
 215                         continue;
 216
 217                 kref_get(&gl->gl_ref);
 218
 219                 return gl;
 220         }
 221
 222         return NULL;
 223 }
 224
 225 /**
 226  * gfs2_glock_find() - Find glock by lock number
 227  * @sdp: The GFS2 superblock
 228  * @name: The lock name
 229  *
 230  * Returns: NULL, or the struct gfs2_glock with the requested number
 231  */
 232
 233 static struct gfs2_glock *gfs2_glock_find(struct gfs2_sbd *sdp,
 234                                           struct lm_lockname *name)
 235 {
 236         struct gfs2_gl_hash_bucket *bucket = &sdp->sd_gl_hash[gl_hash(name)];
 237         struct gfs2_glock *gl;
 238
 239         read_lock(&bucket->hb_lock);
 240         gl = search_bucket(bucket, name);
 241         read_unlock(&bucket->hb_lock);
 242
 243         return gl;
 244 }
 245
 246 /**
 247  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 248  * @sdp: The GFS2 superblock
 249  * @number: the lock number
 250  * @glops: The glock_operations to use
 251  * @create: If 0, don't create the glock if it doesn't exist
 252  * @glp: the glock is returned here
 253  *
 254  * This does not lock a glock, just finds/creates structures for one.
 255  *
 256  * Returns: errno
 257  */
 258
 259 int gfs2_glock_get(struct gfs2_sbd *sdp, uint64_t number,
 260                    struct gfs2_glock_operations *glops, int create,
 261                    struct gfs2_glock **glp)
 262 {
 263         struct lm_lockname name;
 264         struct gfs2_glock *gl, *tmp;
 265         struct gfs2_gl_hash_bucket *bucket;
 266         int error;
 267
 268         name.ln_number = number;
 269         name.ln_type = glops->go_type;
 270         bucket = &sdp->sd_gl_hash[gl_hash(&name)];
 271
 272         read_lock(&bucket->hb_lock);
 273         gl = search_bucket(bucket, &name);
 274         read_unlock(&bucket->hb_lock);
 275
 276         if (gl || !create) {
 277                 *glp = gl;
 278                 return 0;
 279         }
 280
 281         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
 282         if (!gl)
 283                 return -ENOMEM;
 284
 285         memset(gl, 0, sizeof(struct gfs2_glock));
 286
 287         INIT_LIST_HEAD(&gl->gl_list);
 288         gl->gl_name = name;
 289         kref_init(&gl->gl_ref);
 290
 291         spin_lock_init(&gl->gl_spin);
 292
 293         gl->gl_state = LM_ST_UNLOCKED;
 294         gl->gl_owner = NULL;
 295         gl->gl_ip = 0;
 296         INIT_LIST_HEAD(&gl->gl_holders);
 297         INIT_LIST_HEAD(&gl->gl_waiters1);
 298         INIT_LIST_HEAD(&gl->gl_waiters2);
 299         INIT_LIST_HEAD(&gl->gl_waiters3);
 300
 301         gl->gl_ops = glops;
 302
 303         gl->gl_bucket = bucket;
 304         INIT_LIST_HEAD(&gl->gl_reclaim);
 305
 306         gl->gl_sbd = sdp;
 307
 308         lops_init_le(&gl->gl_le, &gfs2_glock_lops);
 309         INIT_LIST_HEAD(&gl->gl_ail_list);
 310
 311         /* If this glock protects actual on-disk data or metadata blocks,
 312            create a VFS inode to manage the pages/buffers holding them. */
 313         if (glops == &gfs2_inode_glops ||
 314             glops == &gfs2_rgrp_glops ||
 315             glops == &gfs2_meta_glops) {
 316                 gl->gl_aspace = gfs2_aspace_get(sdp);
 317                 if (!gl->gl_aspace) {
 318                         error = -ENOMEM;
 319                         goto fail;
 320                 }
 321         }
 322
 323         error = gfs2_lm_get_lock(sdp, &name, &gl->gl_lock);
 324         if (error)
 325                 goto fail_aspace;
 326
 327         write_lock(&bucket->hb_lock);
 328         tmp = search_bucket(bucket, &name);
 329         if (tmp) {
 330                 write_unlock(&bucket->hb_lock);
 331                 glock_free(gl);
 332                 gl = tmp;
 333         } else {
 334                 list_add_tail(&gl->gl_list, &bucket->hb_list);
 335                 write_unlock(&bucket->hb_lock);
 336         }
 337
 338         *glp = gl;
 339
 340         return 0;
 341
 342  fail_aspace:
 343         if (gl->gl_aspace)
 344                 gfs2_aspace_put(gl->gl_aspace);
 345
 346  fail:
 347         kmem_cache_free(gfs2_glock_cachep, gl);
 348
 349         return error;
 350 }
 351
 352 /**
 353  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 354  * @gl: the glock
 355  * @state: the state we're requesting
 356  * @flags: the modifier flags
 357  * @gh: the holder structure
 358  *
 359  */
 360
 361 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
 362                       struct gfs2_holder *gh)
 363 {
 364         INIT_LIST_HEAD(&gh->gh_list);
 365         gh->gh_gl = gl;
 366         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 367         gh->gh_owner = current;
 368         gh->gh_state = state;
 369         gh->gh_flags = flags;
 370         gh->gh_error = 0;
 371         gh->gh_iflags = 0;
 372         init_completion(&gh->gh_wait);
 373
 374         if (gh->gh_state == LM_ST_EXCLUSIVE)
 375                 gh->gh_flags |= GL_LOCAL_EXCL;
 376
 377         gfs2_glock_hold(gl);
 378 }
 379
 380 /**
 381  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 382  * @state: the state we're requesting
 383  * @flags: the modifier flags
 384  * @gh: the holder structure
 385  *
 386  * Don't mess with the glock.
 387  *
 388  */
 389
 390 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
 391 {
 392         gh->gh_state = state;
 393         gh->gh_flags = flags;
 394         if (gh->gh_state == LM_ST_EXCLUSIVE)
 395                 gh->gh_flags |= GL_LOCAL_EXCL;
 396
 397         gh->gh_iflags &= 1 << HIF_ALLOCED;
 398         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 399 }
 400
 401 /**
 402  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 403  * @gh: the holder structure
 404  *
 405  */
 406
 407 void gfs2_holder_uninit(struct gfs2_holder *gh)
 408 {
 409         gfs2_glock_put(gh->gh_gl);
 410         gh->gh_gl = NULL;
 411         gh->gh_ip = 0;
 412 }
 413
 414 /**
 415  * gfs2_holder_get - get a struct gfs2_holder structure
 416  * @gl: the glock
 417  * @state: the state we're requesting
 418  * @flags: the modifier flags
 419  * @gfp_flags: __GFP_NOFAIL
 420  *
 421  * Figure out how big an impact this function has.  Either:
 422  * 1) Replace it with a cache of structures hanging off the struct gfs2_sbd
 423  * 2) Leave it like it is
 424  *
 425  * Returns: the holder structure, NULL on ENOMEM
 426  */
 427
 428 static struct gfs2_holder *gfs2_holder_get(struct gfs2_glock *gl,
 429                                            unsigned int state,
 430                                            int flags, gfp_t gfp_flags)
 431 {
 432         struct gfs2_holder *gh;
 433
 434         gh = kmalloc(sizeof(struct gfs2_holder), gfp_flags);
 435         if (!gh)
 436                 return NULL;
 437
 438         gfs2_holder_init(gl, state, flags, gh);
 439         set_bit(HIF_ALLOCED, &gh->gh_iflags);
 440         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 441         return gh;
 442 }
 443
 444 /**
 445  * gfs2_holder_put - get rid of a struct gfs2_holder structure
 446  * @gh: the holder structure
 447  *
 448  */
 449
 450 static void gfs2_holder_put(struct gfs2_holder *gh)
 451 {
 452         gfs2_holder_uninit(gh);
 453         kfree(gh);
 454 }
 455
 456 /**
 457  * rq_mutex - process a mutex request in the queue
 458  * @gh: the glock holder
 459  *
 460  * Returns: 1 if the queue is blocked
 461  */
 462
 463 static int rq_mutex(struct gfs2_holder *gh)
 464 {
 465         struct gfs2_glock *gl = gh->gh_gl;
 466
 467         list_del_init(&gh->gh_list);
 468         /*  gh->gh_error never examined.  */
 469         set_bit(GLF_LOCK, &gl->gl_flags);
 470         complete(&gh->gh_wait);
 471
 472         return 1;
 473 }
 474
 475 /**
 476  * rq_promote - process a promote request in the queue
 477  * @gh: the glock holder
 478  *
 479  * Acquire a new inter-node lock, or change a lock state to more restrictive.
 480  *
 481  * Returns: 1 if the queue is blocked
 482  */
 483
 484 static int rq_promote(struct gfs2_holder *gh)
 485 {
 486         struct gfs2_glock *gl = gh->gh_gl;
 487         struct gfs2_sbd *sdp = gl->gl_sbd;
 488         struct gfs2_glock_operations *glops = gl->gl_ops;
 489
 490         if (!relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 491                 if (list_empty(&gl->gl_holders)) {
 492                         gl->gl_req_gh = gh;
 493                         set_bit(GLF_LOCK, &gl->gl_flags);
 494                         spin_unlock(&gl->gl_spin);
 495
 496                         if (atomic_read(&sdp->sd_reclaim_count) >
 497                             gfs2_tune_get(sdp, gt_reclaim_limit) &&
 498                             !(gh->gh_flags & LM_FLAG_PRIORITY)) {
 499                                 gfs2_reclaim_glock(sdp);
 500                                 gfs2_reclaim_glock(sdp);
 501                         }
 502
 503                         glops->go_xmote_th(gl, gh->gh_state,
 504                                            gh->gh_flags);
 505
 506                         spin_lock(&gl->gl_spin);
 507                 }
 508                 return 1;
 509         }
 510
 511         if (list_empty(&gl->gl_holders)) {
 512                 set_bit(HIF_FIRST, &gh->gh_iflags);
 513                 set_bit(GLF_LOCK, &gl->gl_flags);
 514         } else {
 515                 struct gfs2_holder *next_gh;
 516                 if (gh->gh_flags & GL_LOCAL_EXCL)
 517                         return 1;
 518                 next_gh = list_entry(gl->gl_holders.next, struct gfs2_holder,
 519                                      gh_list);
 520                 if (next_gh->gh_flags & GL_LOCAL_EXCL)
 521                          return 1;
 522         }
 523
 524         list_move_tail(&gh->gh_list, &gl->gl_holders);
 525         gh->gh_error = 0;
 526         set_bit(HIF_HOLDER, &gh->gh_iflags);
 527
 528         complete(&gh->gh_wait);
 529
 530         return 0;
 531 }
 532
 533 /**
 534  * rq_demote - process a demote request in the queue
 535  * @gh: the glock holder
 536  *
 537  * Returns: 1 if the queue is blocked
 538  */
 539
 540 static int rq_demote(struct gfs2_holder *gh)
 541 {
 542         struct gfs2_glock *gl = gh->gh_gl;
 543         struct gfs2_glock_operations *glops = gl->gl_ops;
 544
 545         if (!list_empty(&gl->gl_holders))
 546                 return 1;
 547
 548         if (gl->gl_state == gh->gh_state || gl->gl_state == LM_ST_UNLOCKED) {
 549                 list_del_init(&gh->gh_list);
 550                 gh->gh_error = 0;
 551                 spin_unlock(&gl->gl_spin);
 552                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 553                         gfs2_holder_put(gh);
 554                 else
 555                         complete(&gh->gh_wait);
 556                 spin_lock(&gl->gl_spin);
 557         } else {
 558                 gl->gl_req_gh = gh;
 559                 set_bit(GLF_LOCK, &gl->gl_flags);
 560                 spin_unlock(&gl->gl_spin);
 561
 562                 if (gh->gh_state == LM_ST_UNLOCKED ||
 563                     gl->gl_state != LM_ST_EXCLUSIVE)
 564                         glops->go_drop_th(gl);
 565                 else
 566                         glops->go_xmote_th(gl, gh->gh_state, gh->gh_flags);
 567
 568                 spin_lock(&gl->gl_spin);
 569         }
 570
 571         return 0;
 572 }
 573
 574 /**
 575  * rq_greedy - process a queued request to drop greedy status
 576  * @gh: the glock holder
 577  *
 578  * Returns: 1 if the queue is blocked
 579  */
 580
 581 static int rq_greedy(struct gfs2_holder *gh)
 582 {
 583         struct gfs2_glock *gl = gh->gh_gl;
 584
 585         list_del_init(&gh->gh_list);
 586         /*  gh->gh_error never examined.  */
 587         clear_bit(GLF_GREEDY, &gl->gl_flags);
 588         spin_unlock(&gl->gl_spin);
 589
 590         gfs2_holder_uninit(gh);
 591         kfree(container_of(gh, struct greedy, gr_gh));
 592
 593         spin_lock(&gl->gl_spin);
 594
 595         return 0;
 596 }
 597
 598 /**
 599  * run_queue - process holder structures on a glock
 600  * @gl: the glock
 601  *
 602  */
 603 static void run_queue(struct gfs2_glock *gl)
 604 {
 605         struct gfs2_holder *gh;
 606         int blocked = 1;
 607
 608         for (;;) {
 609                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 610                         break;
 611
 612                 if (!list_empty(&gl->gl_waiters1)) {
 613                         gh = list_entry(gl->gl_waiters1.next,
 614                                         struct gfs2_holder, gh_list);
 615
 616                         if (test_bit(HIF_MUTEX, &gh->gh_iflags))
 617                                 blocked = rq_mutex(gh);
 618                         else
 619                                 gfs2_assert_warn(gl->gl_sbd, 0);
 620
 621                 } else if (!list_empty(&gl->gl_waiters2) &&
 622                            !test_bit(GLF_SKIP_WAITERS2, &gl->gl_flags)) {
 623                         gh = list_entry(gl->gl_waiters2.next,
 624                                         struct gfs2_holder, gh_list);
 625
 626                         if (test_bit(HIF_DEMOTE, &gh->gh_iflags))
 627                                 blocked = rq_demote(gh);
 628                         else if (test_bit(HIF_GREEDY, &gh->gh_iflags))
 629                                 blocked = rq_greedy(gh);
 630                         else
 631                                 gfs2_assert_warn(gl->gl_sbd, 0);
 632
 633                 } else if (!list_empty(&gl->gl_waiters3)) {
 634                         gh = list_entry(gl->gl_waiters3.next,
 635                                         struct gfs2_holder, gh_list);
 636
 637                         if (test_bit(HIF_PROMOTE, &gh->gh_iflags))
 638                                 blocked = rq_promote(gh);
 639                         else
 640                                 gfs2_assert_warn(gl->gl_sbd, 0);
 641
 642                 } else
 643                         break;
 644
 645                 if (blocked)
 646                         break;
 647         }
 648 }
 649
 650 /**
 651  * gfs2_glmutex_lock - acquire a local lock on a glock
 652  * @gl: the glock
 653  *
 654  * Gives caller exclusive access to manipulate a glock structure.
 655  */
 656
 657 void gfs2_glmutex_lock(struct gfs2_glock *gl)
 658 {
 659         struct gfs2_holder gh;
 660
 661         gfs2_holder_init(gl, 0, 0, &gh);
 662         set_bit(HIF_MUTEX, &gh.gh_iflags);
 663
 664         spin_lock(&gl->gl_spin);
 665         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 666                 list_add_tail(&gh.gh_list, &gl->gl_waiters1);
 667         else {
 668                 gl->gl_owner = current;
 669                 gl->gl_ip = (unsigned long)__builtin_return_address(0);
 670                 complete(&gh.gh_wait);
 671         }
 672         spin_unlock(&gl->gl_spin);
 673
 674         wait_for_completion(&gh.gh_wait);
 675         gfs2_holder_uninit(&gh);
 676 }
 677
 678 /**
 679  * gfs2_glmutex_trylock - try to acquire a local lock on a glock
 680  * @gl: the glock
 681  *
 682  * Returns: 1 if the glock is acquired
 683  */
 684
 685 static int gfs2_glmutex_trylock(struct gfs2_glock *gl)
 686 {
 687         int acquired = 1;
 688
 689         spin_lock(&gl->gl_spin);
 690         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 691                 acquired = 0;
 692         else {
 693                 gl->gl_owner = current;
 694                 gl->gl_ip = (unsigned long)__builtin_return_address(0);
 695         }
 696         spin_unlock(&gl->gl_spin);
 697
 698         return acquired;
 699 }
 700
 701 /**
 702  * gfs2_glmutex_unlock - release a local lock on a glock
 703  * @gl: the glock
 704  *
 705  */
 706
 707 void gfs2_glmutex_unlock(struct gfs2_glock *gl)
 708 {
 709         spin_lock(&gl->gl_spin);
 710         clear_bit(GLF_LOCK, &gl->gl_flags);
 711         gl->gl_owner = NULL;
 712         gl->gl_ip = 0;
 713         run_queue(gl);
 714         BUG_ON(!spin_is_locked(&gl->gl_spin));
 715         spin_unlock(&gl->gl_spin);
 716 }
 717
 718 /**
 719  * handle_callback - add a demote request to a lock's queue
 720  * @gl: the glock
 721  * @state: the state the caller wants us to change to
 722  *
 723  */
 724
 725 static void handle_callback(struct gfs2_glock *gl, unsigned int state)
 726 {
 727         struct gfs2_holder *gh, *new_gh = NULL;
 728
 729  restart:
 730         spin_lock(&gl->gl_spin);
 731
 732         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
 733                 if (test_bit(HIF_DEMOTE, &gh->gh_iflags) &&
 734                     gl->gl_req_gh != gh) {
 735                         if (gh->gh_state != state)
 736                                 gh->gh_state = LM_ST_UNLOCKED;
 737                         goto out;
 738                 }
 739         }
 740
 741         if (new_gh) {
 742                 list_add_tail(&new_gh->gh_list, &gl->gl_waiters2);
 743                 new_gh = NULL;
 744         } else {
 745                 spin_unlock(&gl->gl_spin);
 746
 747                 new_gh = gfs2_holder_get(gl, state, LM_FLAG_TRY,
 748                                          GFP_KERNEL | __GFP_NOFAIL),
 749                 set_bit(HIF_DEMOTE, &new_gh->gh_iflags);
 750                 set_bit(HIF_DEALLOC, &new_gh->gh_iflags);
 751
 752                 goto restart;
 753         }
 754
 755  out:
 756         spin_unlock(&gl->gl_spin);
 757
 758         if (new_gh)
 759                 gfs2_holder_put(new_gh);
 760 }
 761
 762 /**
 763  * state_change - record that the glock is now in a different state
 764  * @gl: the glock
 765  * @new_state the new state
 766  *
 767  */
 768
 769 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 770 {
 771         int held1, held2;
 772
 773         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 774         held2 = (new_state != LM_ST_UNLOCKED);
 775
 776         if (held1 != held2) {
 777                 if (held2)
 778                         gfs2_glock_hold(gl);
 779                 else
 780                         gfs2_glock_put(gl);
 781         }
 782
 783         gl->gl_state = new_state;
 784 }
 785
 786 /**
 787  * xmote_bh - Called after the lock module is done acquiring a lock
 788  * @gl: The glock in question
 789  * @ret: the int returned from the lock module
 790  *
 791  */
 792
 793 static void xmote_bh(struct gfs2_glock *gl, unsigned int ret)
 794 {
 795         struct gfs2_sbd *sdp = gl->gl_sbd;
 796         struct gfs2_glock_operations *glops = gl->gl_ops;
 797         struct gfs2_holder *gh = gl->gl_req_gh;
 798         int prev_state = gl->gl_state;
 799         int op_done = 1;
 800
 801         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 802         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 803         gfs2_assert_warn(sdp, !(ret & LM_OUT_ASYNC));
 804
 805         state_change(gl, ret & LM_OUT_ST_MASK);
 806
 807         if (prev_state != LM_ST_UNLOCKED && !(ret & LM_OUT_CACHEABLE)) {
 808                 if (glops->go_inval)
 809                         glops->go_inval(gl, DIO_METADATA | DIO_DATA);
 810         } else if (gl->gl_state == LM_ST_DEFERRED) {
 811                 /* We might not want to do this here.
 812                    Look at moving to the inode glops. */
 813                 if (glops->go_inval)
 814                         glops->go_inval(gl, DIO_DATA);
 815         }
 816
 817         /*  Deal with each possible exit condition  */
 818
 819         if (!gh)
 820                 gl->gl_stamp = jiffies;
 821
 822         else if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
 823                 spin_lock(&gl->gl_spin);
 824                 list_del_init(&gh->gh_list);
 825                 gh->gh_error = -EIO;
 826                 spin_unlock(&gl->gl_spin);
 827
 828         } else if (test_bit(HIF_DEMOTE, &gh->gh_iflags)) {
 829                 spin_lock(&gl->gl_spin);
 830                 list_del_init(&gh->gh_list);
 831                 if (gl->gl_state == gh->gh_state ||
 832                     gl->gl_state == LM_ST_UNLOCKED)
 833                         gh->gh_error = 0;
 834                 else {
 835                         if (gfs2_assert_warn(sdp, gh->gh_flags &
 836                                         (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) == -1)
 837                                 fs_warn(sdp, "ret = 0x%.8X\n", ret);
 838                         gh->gh_error = GLR_TRYFAILED;
 839                 }
 840                 spin_unlock(&gl->gl_spin);
 841
 842                 if (ret & LM_OUT_CANCELED)
 843                         handle_callback(gl, LM_ST_UNLOCKED); /* Lame */
 844
 845         } else if (ret & LM_OUT_CANCELED) {
 846                 spin_lock(&gl->gl_spin);
 847                 list_del_init(&gh->gh_list);
 848                 gh->gh_error = GLR_CANCELED;
 849                 spin_unlock(&gl->gl_spin);
 850
 851         } else if (relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 852                 spin_lock(&gl->gl_spin);
 853                 list_move_tail(&gh->gh_list, &gl->gl_holders);
 854                 gh->gh_error = 0;
 855                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 856                 spin_unlock(&gl->gl_spin);
 857
 858                 set_bit(HIF_FIRST, &gh->gh_iflags);
 859
 860                 op_done = 0;
 861
 862         } else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 863                 spin_lock(&gl->gl_spin);
 864                 list_del_init(&gh->gh_list);
 865                 gh->gh_error = GLR_TRYFAILED;
 866                 spin_unlock(&gl->gl_spin);
 867
 868         } else {
 869                 if (gfs2_assert_withdraw(sdp, 0) == -1)
 870                         fs_err(sdp, "ret = 0x%.8X\n", ret);
 871         }
 872
 873         if (glops->go_xmote_bh)
 874                 glops->go_xmote_bh(gl);
 875
 876         if (op_done) {
 877                 spin_lock(&gl->gl_spin);
 878                 gl->gl_req_gh = NULL;
 879                 gl->gl_req_bh = NULL;
 880                 clear_bit(GLF_LOCK, &gl->gl_flags);
 881                 run_queue(gl);
 882                 spin_unlock(&gl->gl_spin);
 883         }
 884
 885         gfs2_glock_put(gl);
 886
 887         if (gh) {
 888                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 889                         gfs2_holder_put(gh);
 890                 else
 891                         complete(&gh->gh_wait);
 892         }
 893 }
 894
 895 /**
 896  * gfs2_glock_xmote_th - Call into the lock module to acquire or change a glock
 897  * @gl: The glock in question
 898  * @state: the requested state
 899  * @flags: modifier flags to the lock call
 900  *
 901  */
 902
 903 void gfs2_glock_xmote_th(struct gfs2_glock *gl, unsigned int state, int flags)
 904 {
 905         struct gfs2_sbd *sdp = gl->gl_sbd;
 906         struct gfs2_glock_operations *glops = gl->gl_ops;
 907         int lck_flags = flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB |
 908                                  LM_FLAG_NOEXP | LM_FLAG_ANY |
 909                                  LM_FLAG_PRIORITY);
 910         unsigned int lck_ret;
 911
 912         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 913         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 914         gfs2_assert_warn(sdp, state != LM_ST_UNLOCKED);
 915         gfs2_assert_warn(sdp, state != gl->gl_state);
 916
 917         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 918                 if (glops->go_sync)
 919                         glops->go_sync(gl,
 920                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
 921         }
 922
 923         gfs2_glock_hold(gl);
 924         gl->gl_req_bh = xmote_bh;
 925
 926         lck_ret = gfs2_lm_lock(sdp, gl->gl_lock, gl->gl_state, state,
 927                                lck_flags);
 928
 929         if (gfs2_assert_withdraw(sdp, !(lck_ret & LM_OUT_ERROR)))
 930                 return;
 931
 932         if (lck_ret & LM_OUT_ASYNC)
 933                 gfs2_assert_warn(sdp, lck_ret == LM_OUT_ASYNC);
 934         else
 935                 xmote_bh(gl, lck_ret);
 936 }
 937
 938 /**
 939  * drop_bh - Called after a lock module unlock completes
 940  * @gl: the glock
 941  * @ret: the return status
 942  *
 943  * Doesn't wake up the process waiting on the struct gfs2_holder (if any)
 944  * Doesn't drop the reference on the glock the top half took out
 945  *
 946  */
 947
 948 static void drop_bh(struct gfs2_glock *gl, unsigned int ret)
 949 {
 950         struct gfs2_sbd *sdp = gl->gl_sbd;
 951         struct gfs2_glock_operations *glops = gl->gl_ops;
 952         struct gfs2_holder *gh = gl->gl_req_gh;
 953
 954         clear_bit(GLF_PREFETCH, &gl->gl_flags);
 955
 956         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 957         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 958         gfs2_assert_warn(sdp, !ret);
 959
 960         state_change(gl, LM_ST_UNLOCKED);
 961
 962         if (glops->go_inval)
 963                 glops->go_inval(gl, DIO_METADATA | DIO_DATA);
 964
 965         if (gh) {
 966                 spin_lock(&gl->gl_spin);
 967                 list_del_init(&gh->gh_list);
 968                 gh->gh_error = 0;
 969                 spin_unlock(&gl->gl_spin);
 970         }
 971
 972         if (glops->go_drop_bh)
 973                 glops->go_drop_bh(gl);
 974
 975         spin_lock(&gl->gl_spin);
 976         gl->gl_req_gh = NULL;
 977         gl->gl_req_bh = NULL;
 978         clear_bit(GLF_LOCK, &gl->gl_flags);
 979         run_queue(gl);
 980         spin_unlock(&gl->gl_spin);
 981
 982         gfs2_glock_put(gl);
 983
 984         if (gh) {
 985                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 986                         gfs2_holder_put(gh);
 987                 else
 988                         complete(&gh->gh_wait);
 989         }
 990 }
 991
 992 /**
 993  * gfs2_glock_drop_th - call into the lock module to unlock a lock
 994  * @gl: the glock
 995  *
 996  */
 997
 998 void gfs2_glock_drop_th(struct gfs2_glock *gl)
 999 {
1000         struct gfs2_sbd *sdp = gl->gl_sbd;
1001         struct gfs2_glock_operations *glops = gl->gl_ops;
1002         unsigned int ret;
1003
1004         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1005         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1006         gfs2_assert_warn(sdp, gl->gl_state != LM_ST_UNLOCKED);
1007
1008         if (gl->gl_state == LM_ST_EXCLUSIVE) {
1009                 if (glops->go_sync)
1010                         glops->go_sync(gl,
1011                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
1012         }
1013
1014         gfs2_glock_hold(gl);
1015         gl->gl_req_bh = drop_bh;
1016
1017         ret = gfs2_lm_unlock(sdp, gl->gl_lock, gl->gl_state);
1018
1019         if (gfs2_assert_withdraw(sdp, !(ret & LM_OUT_ERROR)))
1020                 return;
1021
1022         if (!ret)
1023                 drop_bh(gl, ret);
1024         else
1025                 gfs2_assert_warn(sdp, ret == LM_OUT_ASYNC);
1026 }
1027
1028 /**
1029  * do_cancels - cancel requests for locks stuck waiting on an expire flag
1030  * @gh: the LM_FLAG_PRIORITY holder waiting to acquire the lock
1031  *
1032  * Don't cancel GL_NOCANCEL requests.
1033  */
1034
1035 static void do_cancels(struct gfs2_holder *gh)
1036 {
1037         struct gfs2_glock *gl = gh->gh_gl;
1038
1039         spin_lock(&gl->gl_spin);
1040
1041         while (gl->gl_req_gh != gh &&
1042                !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1043                !list_empty(&gh->gh_list)) {
1044                 if (gl->gl_req_bh &&
1045                     !(gl->gl_req_gh &&
1046                       (gl->gl_req_gh->gh_flags & GL_NOCANCEL))) {
1047                         spin_unlock(&gl->gl_spin);
1048                         gfs2_lm_cancel(gl->gl_sbd, gl->gl_lock);
1049                         msleep(100);
1050                         spin_lock(&gl->gl_spin);
1051                 } else {
1052                         spin_unlock(&gl->gl_spin);
1053                         msleep(100);
1054                         spin_lock(&gl->gl_spin);
1055                 }
1056         }
1057
1058         spin_unlock(&gl->gl_spin);
1059 }
1060
1061 /**
1062  * glock_wait_internal - wait on a glock acquisition
1063  * @gh: the glock holder
1064  *
1065  * Returns: 0 on success
1066  */
1067
1068 static int glock_wait_internal(struct gfs2_holder *gh)
1069 {
1070         struct gfs2_glock *gl = gh->gh_gl;
1071         struct gfs2_sbd *sdp = gl->gl_sbd;
1072         struct gfs2_glock_operations *glops = gl->gl_ops;
1073
1074         if (test_bit(HIF_ABORTED, &gh->gh_iflags))
1075                 return -EIO;
1076
1077         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1078                 spin_lock(&gl->gl_spin);
1079                 if (gl->gl_req_gh != gh &&
1080                     !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1081                     !list_empty(&gh->gh_list)) {
1082                         list_del_init(&gh->gh_list);
1083                         gh->gh_error = GLR_TRYFAILED;
1084                         run_queue(gl);
1085                         spin_unlock(&gl->gl_spin);
1086                         return gh->gh_error;
1087                 }
1088                 spin_unlock(&gl->gl_spin);
1089         }
1090
1091         if (gh->gh_flags & LM_FLAG_PRIORITY)
1092                 do_cancels(gh);
1093
1094         wait_for_completion(&gh->gh_wait);
1095
1096         if (gh->gh_error)
1097                 return gh->gh_error;
1098
1099         gfs2_assert_withdraw(sdp, test_bit(HIF_HOLDER, &gh->gh_iflags));
1100         gfs2_assert_withdraw(sdp, relaxed_state_ok(gl->gl_state,
1101                                                    gh->gh_state,
1102                                                    gh->gh_flags));
1103
1104         if (test_bit(HIF_FIRST, &gh->gh_iflags)) {
1105                 gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1106
1107                 if (glops->go_lock) {
1108                         gh->gh_error = glops->go_lock(gh);
1109                         if (gh->gh_error) {
1110                                 spin_lock(&gl->gl_spin);
1111                                 list_del_init(&gh->gh_list);
1112                                 spin_unlock(&gl->gl_spin);
1113                         }
1114                 }
1115
1116                 spin_lock(&gl->gl_spin);
1117                 gl->gl_req_gh = NULL;
1118                 gl->gl_req_bh = NULL;
1119                 clear_bit(GLF_LOCK, &gl->gl_flags);
1120                 run_queue(gl);
1121                 spin_unlock(&gl->gl_spin);
1122         }
1123
1124         return gh->gh_error;
1125 }
1126
1127 static inline struct gfs2_holder *
1128 find_holder_by_owner(struct list_head *head, struct task_struct *owner)
1129 {
1130         struct gfs2_holder *gh;
1131
1132         list_for_each_entry(gh, head, gh_list) {
1133                 if (gh->gh_owner == owner)
1134                         return gh;
1135         }
1136
1137         return NULL;
1138 }
1139
1140 /**
1141  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1142  * @gh: the holder structure to add
1143  *
1144  */
1145
1146 static void add_to_queue(struct gfs2_holder *gh)
1147 {
1148         struct gfs2_glock *gl = gh->gh_gl;
1149         struct gfs2_holder *existing;
1150
1151         BUG_ON(!gh->gh_owner);
1152
1153         existing = find_holder_by_owner(&gl->gl_holders, gh->gh_owner);
1154         if (existing) {
1155                 print_symbol(KERN_WARNING "original: %s\n", existing->gh_ip);
1156                 print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
1157                 BUG();
1158         }
1159
1160         existing = find_holder_by_owner(&gl->gl_waiters3, gh->gh_owner);
1161         if (existing) {
1162                 print_symbol(KERN_WARNING "original: %s\n", existing->gh_ip);
1163                 print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
1164                 BUG();
1165         }
1166
1167         if (gh->gh_flags & LM_FLAG_PRIORITY)
1168                 list_add(&gh->gh_list, &gl->gl_waiters3);
1169         else
1170                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1171 }
1172
1173 /**
1174  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1175  * @gh: the holder structure
1176  *
1177  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1178  *
1179  * Returns: 0, GLR_TRYFAILED, or errno on failure
1180  */
1181
1182 int gfs2_glock_nq(struct gfs2_holder *gh)
1183 {
1184         struct gfs2_glock *gl = gh->gh_gl;
1185         struct gfs2_sbd *sdp = gl->gl_sbd;
1186         int error = 0;
1187
1188 restart:
1189         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
1190                 set_bit(HIF_ABORTED, &gh->gh_iflags);
1191                 return -EIO;
1192         }
1193
1194         set_bit(HIF_PROMOTE, &gh->gh_iflags);
1195
1196         spin_lock(&gl->gl_spin);
1197         add_to_queue(gh);
1198         run_queue(gl);
1199         spin_unlock(&gl->gl_spin);
1200
1201         if (!(gh->gh_flags & GL_ASYNC)) {
1202                 error = glock_wait_internal(gh);
1203                 if (error == GLR_CANCELED) {
1204                         msleep(100);
1205                         goto restart;
1206                 }
1207         }
1208
1209         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1210
1211         if (error == GLR_TRYFAILED && (gh->gh_flags & GL_DUMP))
1212                 dump_glock(gl);
1213
1214         return error;
1215 }
1216
1217 /**
1218  * gfs2_glock_poll - poll to see if an async request has been completed
1219  * @gh: the holder
1220  *
1221  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1222  */
1223
1224 int gfs2_glock_poll(struct gfs2_holder *gh)
1225 {
1226         struct gfs2_glock *gl = gh->gh_gl;
1227         int ready = 0;
1228
1229         spin_lock(&gl->gl_spin);
1230
1231         if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1232                 ready = 1;
1233         else if (list_empty(&gh->gh_list)) {
1234                 if (gh->gh_error == GLR_CANCELED) {
1235                         spin_unlock(&gl->gl_spin);
1236                         msleep(100);
1237                         if (gfs2_glock_nq(gh))
1238                                 return 1;
1239                         return 0;
1240                 } else
1241                         ready = 1;
1242         }
1243
1244         spin_unlock(&gl->gl_spin);
1245
1246         return ready;
1247 }
1248
1249 /**
1250  * gfs2_glock_wait - wait for a lock acquisition that ended in a GLR_ASYNC
1251  * @gh: the holder structure
1252  *
1253  * Returns: 0, GLR_TRYFAILED, or errno on failure
1254  */
1255
1256 int gfs2_glock_wait(struct gfs2_holder *gh)
1257 {
1258         int error;
1259
1260         error = glock_wait_internal(gh);
1261         if (error == GLR_CANCELED) {
1262                 msleep(100);
1263                 gh->gh_flags &= ~GL_ASYNC;
1264                 error = gfs2_glock_nq(gh);
1265         }
1266
1267         return error;
1268 }
1269
1270 /**
1271  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1272  * @gh: the glock holder
1273  *
1274  */
1275
1276 void gfs2_glock_dq(struct gfs2_holder *gh)
1277 {
1278         struct gfs2_glock *gl = gh->gh_gl;
1279         struct gfs2_glock_operations *glops = gl->gl_ops;
1280
1281         if (gh->gh_flags & GL_SYNC)
1282                 set_bit(GLF_SYNC, &gl->gl_flags);
1283
1284         if (gh->gh_flags & GL_NOCACHE)
1285                 handle_callback(gl, LM_ST_UNLOCKED);
1286
1287         gfs2_glmutex_lock(gl);
1288
1289         spin_lock(&gl->gl_spin);
1290         list_del_init(&gh->gh_list);
1291
1292         if (list_empty(&gl->gl_holders)) {
1293                 spin_unlock(&gl->gl_spin);
1294
1295                 if (glops->go_unlock)
1296                         glops->go_unlock(gh);
1297
1298                 if (test_bit(GLF_SYNC, &gl->gl_flags)) {
1299                         if (glops->go_sync)
1300                                 glops->go_sync(gl, DIO_METADATA | DIO_DATA);
1301                 }
1302
1303                 gl->gl_stamp = jiffies;
1304
1305                 spin_lock(&gl->gl_spin);
1306         }
1307
1308         clear_bit(GLF_LOCK, &gl->gl_flags);
1309         run_queue(gl);
1310         spin_unlock(&gl->gl_spin);
1311 }
1312
1313 /**
1314  * gfs2_glock_prefetch - Try to prefetch a glock
1315  * @gl: the glock
1316  * @state: the state to prefetch in
1317  * @flags: flags passed to go_xmote_th()
1318  *
1319  */
1320
1321 static void gfs2_glock_prefetch(struct gfs2_glock *gl, unsigned int state,
1322                                 int flags)
1323 {
1324         struct gfs2_glock_operations *glops = gl->gl_ops;
1325
1326         spin_lock(&gl->gl_spin);
1327
1328         if (test_bit(GLF_LOCK, &gl->gl_flags) ||
1329             !list_empty(&gl->gl_holders) ||
1330             !list_empty(&gl->gl_waiters1) ||
1331             !list_empty(&gl->gl_waiters2) ||
1332             !list_empty(&gl->gl_waiters3) ||
1333             relaxed_state_ok(gl->gl_state, state, flags)) {
1334                 spin_unlock(&gl->gl_spin);
1335                 return;
1336         }
1337
1338         set_bit(GLF_PREFETCH, &gl->gl_flags);
1339         set_bit(GLF_LOCK, &gl->gl_flags);
1340         spin_unlock(&gl->gl_spin);
1341
1342         glops->go_xmote_th(gl, state, flags);
1343 }
1344
1345 static void greedy_work(void *data)
1346 {
1347         struct greedy *gr = data;
1348         struct gfs2_holder *gh = &gr->gr_gh;
1349         struct gfs2_glock *gl = gh->gh_gl;
1350         struct gfs2_glock_operations *glops = gl->gl_ops;
1351
1352         clear_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1353
1354         if (glops->go_greedy)
1355                 glops->go_greedy(gl);
1356
1357         spin_lock(&gl->gl_spin);
1358
1359         if (list_empty(&gl->gl_waiters2)) {
1360                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1361                 spin_unlock(&gl->gl_spin);
1362                 gfs2_holder_uninit(gh);
1363                 kfree(gr);
1364         } else {
1365                 gfs2_glock_hold(gl);
1366                 list_add_tail(&gh->gh_list, &gl->gl_waiters2);
1367                 run_queue(gl);
1368                 spin_unlock(&gl->gl_spin);
1369                 gfs2_glock_put(gl);
1370         }
1371 }
1372
1373 /**
1374  * gfs2_glock_be_greedy -
1375  * @gl:
1376  * @time:
1377  *
1378  * Returns: 0 if go_greedy will be called, 1 otherwise
1379  */
1380
1381 int gfs2_glock_be_greedy(struct gfs2_glock *gl, unsigned int time)
1382 {
1383         struct greedy *gr;
1384         struct gfs2_holder *gh;
1385
1386         if (!time ||
1387             gl->gl_sbd->sd_args.ar_localcaching ||
1388             test_and_set_bit(GLF_GREEDY, &gl->gl_flags))
1389                 return 1;
1390
1391         gr = kmalloc(sizeof(struct greedy), GFP_KERNEL);
1392         if (!gr) {
1393                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1394                 return 1;
1395         }
1396         gh = &gr->gr_gh;
1397
1398         gfs2_holder_init(gl, 0, 0, gh);
1399         set_bit(HIF_GREEDY, &gh->gh_iflags);
1400         INIT_WORK(&gr->gr_work, greedy_work, gr);
1401
1402         set_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1403         schedule_delayed_work(&gr->gr_work, time);
1404
1405         return 0;
1406 }
1407
1408 /**
1409  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1410  * @gh: the holder structure
1411  *
1412  */
1413
1414 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1415 {
1416         gfs2_glock_dq(gh);
1417         gfs2_holder_uninit(gh);
1418 }
1419
1420 /**
1421  * gfs2_glock_nq_num - acquire a glock based on lock number
1422  * @sdp: the filesystem
1423  * @number: the lock number
1424  * @glops: the glock operations for the type of glock
1425  * @state: the state to acquire the glock in
1426  * @flags: modifier flags for the aquisition
1427  * @gh: the struct gfs2_holder
1428  *
1429  * Returns: errno
1430  */
1431
1432 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, uint64_t number,
1433                       struct gfs2_glock_operations *glops, unsigned int state,
1434                       int flags, struct gfs2_holder *gh)
1435 {
1436         struct gfs2_glock *gl;
1437         int error;
1438
1439         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1440         if (!error) {
1441                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1442                 gfs2_glock_put(gl);
1443         }
1444
1445         return error;
1446 }
1447
1448 /**
1449  * glock_compare - Compare two struct gfs2_glock structures for sorting
1450  * @arg_a: the first structure
1451  * @arg_b: the second structure
1452  *
1453  */
1454
1455 static int glock_compare(const void *arg_a, const void *arg_b)
1456 {
1457         struct gfs2_holder *gh_a = *(struct gfs2_holder **)arg_a;
1458         struct gfs2_holder *gh_b = *(struct gfs2_holder **)arg_b;
1459         struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1460         struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1461         int ret = 0;
1462
1463         if (a->ln_number > b->ln_number)
1464                 ret = 1;
1465         else if (a->ln_number < b->ln_number)
1466                 ret = -1;
1467         else {
1468                 if (gh_a->gh_state == LM_ST_SHARED &&
1469                     gh_b->gh_state == LM_ST_EXCLUSIVE)
1470                         ret = 1;
1471                 else if (!(gh_a->gh_flags & GL_LOCAL_EXCL) &&
1472                          (gh_b->gh_flags & GL_LOCAL_EXCL))
1473                         ret = 1;
1474         }
1475
1476         return ret;
1477 }
1478
1479 /**
1480  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1481  * @num_gh: the number of structures
1482  * @ghs: an array of struct gfs2_holder structures
1483  *
1484  * Returns: 0 on success (all glocks acquired),
1485  *          errno on failure (no glocks acquired)
1486  */
1487
1488 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1489                      struct gfs2_holder **p)
1490 {
1491         unsigned int x;
1492         int error = 0;
1493
1494         for (x = 0; x < num_gh; x++)
1495                 p[x] = &ghs[x];
1496
1497         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1498
1499         for (x = 0; x < num_gh; x++) {
1500                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1501
1502                 error = gfs2_glock_nq(p[x]);
1503                 if (error) {
1504                         while (x--)
1505                                 gfs2_glock_dq(p[x]);
1506                         break;
1507                 }
1508         }
1509
1510         return error;
1511 }
1512
1513 /**
1514  * gfs2_glock_nq_m - acquire multiple glocks
1515  * @num_gh: the number of structures
1516  * @ghs: an array of struct gfs2_holder structures
1517  *
1518  * Figure out how big an impact this function has.  Either:
1519  * 1) Replace this code with code that calls gfs2_glock_prefetch()
1520  * 2) Forget async stuff and just call nq_m_sync()
1521  * 3) Leave it like it is
1522  *
1523  * Returns: 0 on success (all glocks acquired),
1524  *          errno on failure (no glocks acquired)
1525  */
1526
1527 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1528 {
1529         int *e;
1530         unsigned int x;
1531         int borked = 0, serious = 0;
1532         int error = 0;
1533
1534         if (!num_gh)
1535                 return 0;
1536
1537         if (num_gh == 1) {
1538                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1539                 return gfs2_glock_nq(ghs);
1540         }
1541
1542         e = kcalloc(num_gh, sizeof(struct gfs2_holder *), GFP_KERNEL);
1543         if (!e)
1544                 return -ENOMEM;
1545
1546         for (x = 0; x < num_gh; x++) {
1547                 ghs[x].gh_flags |= LM_FLAG_TRY | GL_ASYNC;
1548                 error = gfs2_glock_nq(&ghs[x]);
1549                 if (error) {
1550                         borked = 1;
1551                         serious = error;
1552                         num_gh = x;
1553                         break;
1554                 }
1555         }
1556
1557         for (x = 0; x < num_gh; x++) {
1558                 error = e[x] = glock_wait_internal(&ghs[x]);
1559                 if (error) {
1560                         borked = 1;
1561                         if (error != GLR_TRYFAILED && error != GLR_CANCELED)
1562                                 serious = error;
1563                 }
1564         }
1565
1566         if (!borked) {
1567                 kfree(e);
1568                 return 0;
1569         }
1570
1571         for (x = 0; x < num_gh; x++)
1572                 if (!e[x])
1573                         gfs2_glock_dq(&ghs[x]);
1574
1575         if (serious)
1576                 error = serious;
1577         else {
1578                 for (x = 0; x < num_gh; x++)
1579                         gfs2_holder_reinit(ghs[x].gh_state, ghs[x].gh_flags,
1580                                           &ghs[x]);
1581                 error = nq_m_sync(num_gh, ghs, (struct gfs2_holder **)e);
1582         }
1583
1584         kfree(e);
1585
1586         return error;
1587 }
1588
1589 /**
1590  * gfs2_glock_dq_m - release multiple glocks
1591  * @num_gh: the number of structures
1592  * @ghs: an array of struct gfs2_holder structures
1593  *
1594  */
1595
1596 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1597 {
1598         unsigned int x;
1599
1600         for (x = 0; x < num_gh; x++)
1601                 gfs2_glock_dq(&ghs[x]);
1602 }
1603
1604 /**
1605  * gfs2_glock_dq_uninit_m - release multiple glocks
1606  * @num_gh: the number of structures
1607  * @ghs: an array of struct gfs2_holder structures
1608  *
1609  */
1610
1611 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1612 {
1613         unsigned int x;
1614
1615         for (x = 0; x < num_gh; x++)
1616                 gfs2_glock_dq_uninit(&ghs[x]);
1617 }
1618
1619 /**
1620  * gfs2_glock_prefetch_num - prefetch a glock based on lock number
1621  * @sdp: the filesystem
1622  * @number: the lock number
1623  * @glops: the glock operations for the type of glock
1624  * @state: the state to acquire the glock in
1625  * @flags: modifier flags for the aquisition
1626  *
1627  * Returns: errno
1628  */
1629
1630 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, uint64_t number,
1631                              struct gfs2_glock_operations *glops,
1632                              unsigned int state, int flags)
1633 {
1634         struct gfs2_glock *gl;
1635         int error;
1636
1637         if (atomic_read(&sdp->sd_reclaim_count) <
1638             gfs2_tune_get(sdp, gt_reclaim_limit)) {
1639                 error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1640                 if (!error) {
1641                         gfs2_glock_prefetch(gl, state, flags);
1642                         gfs2_glock_put(gl);
1643                 }
1644         }
1645 }
1646
1647 /**
1648  * gfs2_lvb_hold - attach a LVB from a glock
1649  * @gl: The glock in question
1650  *
1651  */
1652
1653 int gfs2_lvb_hold(struct gfs2_glock *gl)
1654 {
1655         int error;
1656
1657         gfs2_glmutex_lock(gl);
1658
1659         if (!atomic_read(&gl->gl_lvb_count)) {
1660                 error = gfs2_lm_hold_lvb(gl->gl_sbd, gl->gl_lock, &gl->gl_lvb);
1661                 if (error) {
1662                         gfs2_glmutex_unlock(gl);
1663                         return error;
1664                 }
1665                 gfs2_glock_hold(gl);
1666         }
1667         atomic_inc(&gl->gl_lvb_count);
1668
1669         gfs2_glmutex_unlock(gl);
1670
1671         return 0;
1672 }
1673
1674 /**
1675  * gfs2_lvb_unhold - detach a LVB from a glock
1676  * @gl: The glock in question
1677  *
1678  */
1679
1680 void gfs2_lvb_unhold(struct gfs2_glock *gl)
1681 {
1682         gfs2_glock_hold(gl);
1683         gfs2_glmutex_lock(gl);
1684
1685         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count) > 0);
1686         if (atomic_dec_and_test(&gl->gl_lvb_count)) {
1687                 gfs2_lm_unhold_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1688                 gl->gl_lvb = NULL;
1689                 gfs2_glock_put(gl);
1690         }
1691
1692         gfs2_glmutex_unlock(gl);
1693         gfs2_glock_put(gl);
1694 }
1695
1696 #if 0
1697 void gfs2_lvb_sync(struct gfs2_glock *gl)
1698 {
1699         gfs2_glmutex_lock(gl);
1700
1701         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count));
1702         if (!gfs2_assert_warn(gl->gl_sbd, gfs2_glock_is_held_excl(gl)))
1703                 gfs2_lm_sync_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1704
1705         gfs2_glmutex_unlock(gl);
1706 }
1707 #endif  /*  0  */
1708
1709 static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
1710                         unsigned int state)
1711 {
1712         struct gfs2_glock *gl;
1713
1714         gl = gfs2_glock_find(sdp, name);
1715         if (!gl)
1716                 return;
1717
1718         if (gl->gl_ops->go_callback)
1719                 gl->gl_ops->go_callback(gl, state);
1720         handle_callback(gl, state);
1721
1722         spin_lock(&gl->gl_spin);
1723         run_queue(gl);
1724         spin_unlock(&gl->gl_spin);
1725
1726         gfs2_glock_put(gl);
1727 }
1728
1729 /**
1730  * gfs2_glock_cb - Callback used by locking module
1731  * @fsdata: Pointer to the superblock
1732  * @type: Type of callback
1733  * @data: Type dependent data pointer
1734  *
1735  * Called by the locking module when it wants to tell us something.
1736  * Either we need to drop a lock, one of our ASYNC requests completed, or
1737  * a journal from another client needs to be recovered.
1738  */
1739
1740 void gfs2_glock_cb(lm_fsdata_t *fsdata, unsigned int type, void *data)
1741 {
1742         struct gfs2_sbd *sdp = (struct gfs2_sbd *)fsdata;
1743
1744         switch (type) {
1745         case LM_CB_NEED_E:
1746                 blocking_cb(sdp, data, LM_ST_UNLOCKED);
1747                 return;
1748
1749         case LM_CB_NEED_D:
1750                 blocking_cb(sdp, data, LM_ST_DEFERRED);
1751                 return;
1752
1753         case LM_CB_NEED_S:
1754                 blocking_cb(sdp, data, LM_ST_SHARED);
1755                 return;
1756
1757         case LM_CB_ASYNC: {
1758                 struct lm_async_cb *async = data;
1759                 struct gfs2_glock *gl;
1760
1761                 gl = gfs2_glock_find(sdp, &async->lc_name);
1762                 if (gfs2_assert_warn(sdp, gl))
1763                         return;
1764                 if (!gfs2_assert_warn(sdp, gl->gl_req_bh))
1765                         gl->gl_req_bh(gl, async->lc_ret);
1766                 gfs2_glock_put(gl);
1767                 return;
1768         }
1769
1770         case LM_CB_NEED_RECOVERY:
1771                 gfs2_jdesc_make_dirty(sdp, *(unsigned int *)data);
1772                 if (sdp->sd_recoverd_process)
1773                         wake_up_process(sdp->sd_recoverd_process);
1774                 return;
1775
1776         case LM_CB_DROPLOCKS:
1777                 gfs2_gl_hash_clear(sdp, NO_WAIT);
1778                 gfs2_quota_scan(sdp);
1779                 return;
1780
1781         default:
1782                 gfs2_assert_warn(sdp, 0);
1783                 return;
1784         }
1785 }
1786
1787 /**
1788  * gfs2_try_toss_inode - try to remove a particular inode struct from cache
1789  * sdp: the filesystem
1790  * inum: the inode number
1791  *
1792  */
1793
1794 void gfs2_try_toss_inode(struct gfs2_sbd *sdp, struct gfs2_inum *inum)
1795 {
1796         struct gfs2_glock *gl;
1797         struct gfs2_inode *ip;
1798         int error;
1799
1800         error = gfs2_glock_get(sdp, inum->no_addr, &gfs2_inode_glops,
1801                                NO_CREATE, &gl);
1802         if (error || !gl)
1803                 return;
1804
1805         if (!gfs2_glmutex_trylock(gl))
1806                 goto out;
1807
1808         ip = gl->gl_object;
1809         if (!ip)
1810                 goto out_unlock;
1811
1812         if (atomic_read(&ip->i_count))
1813                 goto out_unlock;
1814
1815         gfs2_inode_destroy(ip, 1);
1816
1817  out_unlock:
1818         gfs2_glmutex_unlock(gl);
1819
1820  out:
1821         gfs2_glock_put(gl);
1822 }
1823
1824 /**
1825  * gfs2_iopen_go_callback - Try to kick the inode/vnode associated with an
1826  *                          iopen glock from memory
1827  * @io_gl: the iopen glock
1828  * @state: the state into which the glock should be put
1829  *
1830  */
1831
1832 void gfs2_iopen_go_callback(struct gfs2_glock *io_gl, unsigned int state)
1833 {
1834         struct gfs2_glock *i_gl;
1835
1836         if (state != LM_ST_UNLOCKED)
1837                 return;
1838
1839         spin_lock(&io_gl->gl_spin);
1840         i_gl = io_gl->gl_object;
1841         if (i_gl) {
1842                 gfs2_glock_hold(i_gl);
1843                 spin_unlock(&io_gl->gl_spin);
1844         } else {
1845                 spin_unlock(&io_gl->gl_spin);
1846                 return;
1847         }
1848
1849         if (gfs2_glmutex_trylock(i_gl)) {
1850                 struct gfs2_inode *ip = i_gl->gl_object;
1851                 if (ip) {
1852                         gfs2_try_toss_vnode(ip);
1853                         gfs2_glmutex_unlock(i_gl);
1854                         gfs2_glock_schedule_for_reclaim(i_gl);
1855                         goto out;
1856                 }
1857                 gfs2_glmutex_unlock(i_gl);
1858         }
1859
1860  out:
1861         gfs2_glock_put(i_gl);
1862 }
1863
1864 /**
1865  * demote_ok - Check to see if it's ok to unlock a glock
1866  * @gl: the glock
1867  *
1868  * Returns: 1 if it's ok
1869  */
1870
1871 static int demote_ok(struct gfs2_glock *gl)
1872 {
1873         struct gfs2_sbd *sdp = gl->gl_sbd;
1874         struct gfs2_glock_operations *glops = gl->gl_ops;
1875         int demote = 1;
1876
1877         if (test_bit(GLF_STICKY, &gl->gl_flags))
1878                 demote = 0;
1879         else if (test_bit(GLF_PREFETCH, &gl->gl_flags))
1880                 demote = time_after_eq(jiffies,
1881                                     gl->gl_stamp +
1882                                     gfs2_tune_get(sdp, gt_prefetch_secs) * HZ);
1883         else if (glops->go_demote_ok)
1884                 demote = glops->go_demote_ok(gl);
1885
1886         return demote;
1887 }
1888
1889 /**
1890  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
1891  * @gl: the glock
1892  *
1893  */
1894
1895 void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
1896 {
1897         struct gfs2_sbd *sdp = gl->gl_sbd;
1898
1899         spin_lock(&sdp->sd_reclaim_lock);
1900         if (list_empty(&gl->gl_reclaim)) {
1901                 gfs2_glock_hold(gl);
1902                 list_add(&gl->gl_reclaim, &sdp->sd_reclaim_list);
1903                 atomic_inc(&sdp->sd_reclaim_count);
1904         }
1905         spin_unlock(&sdp->sd_reclaim_lock);
1906
1907         wake_up(&sdp->sd_reclaim_wq);
1908 }
1909
1910 /**
1911  * gfs2_reclaim_glock - process the next glock on the filesystem's reclaim list
1912  * @sdp: the filesystem
1913  *
1914  * Called from gfs2_glockd() glock reclaim daemon, or when promoting a
1915  * different glock and we notice that there are a lot of glocks in the
1916  * reclaim list.
1917  *
1918  */
1919
1920 void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
1921 {
1922         struct gfs2_glock *gl;
1923
1924         spin_lock(&sdp->sd_reclaim_lock);
1925         if (list_empty(&sdp->sd_reclaim_list)) {
1926                 spin_unlock(&sdp->sd_reclaim_lock);
1927                 return;
1928         }
1929         gl = list_entry(sdp->sd_reclaim_list.next,
1930                         struct gfs2_glock, gl_reclaim);
1931         list_del_init(&gl->gl_reclaim);
1932         spin_unlock(&sdp->sd_reclaim_lock);
1933
1934         atomic_dec(&sdp->sd_reclaim_count);
1935         atomic_inc(&sdp->sd_reclaimed);
1936
1937         if (gfs2_glmutex_trylock(gl)) {
1938                 if (gl->gl_ops == &gfs2_inode_glops) {
1939                         struct gfs2_inode *ip = gl->gl_object;
1940                         if (ip && !atomic_read(&ip->i_count))
1941                                 gfs2_inode_destroy(ip, 1);
1942                 }
1943                 if (queue_empty(gl, &gl->gl_holders) &&
1944                     gl->gl_state != LM_ST_UNLOCKED &&
1945                     demote_ok(gl))
1946                         handle_callback(gl, LM_ST_UNLOCKED);
1947                 gfs2_glmutex_unlock(gl);
1948         }
1949
1950         gfs2_glock_put(gl);
1951 }
1952
1953 /**
1954  * examine_bucket - Call a function for glock in a hash bucket
1955  * @examiner: the function
1956  * @sdp: the filesystem
1957  * @bucket: the bucket
1958  *
1959  * Returns: 1 if the bucket has entries
1960  */
1961
1962 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
1963                           struct gfs2_gl_hash_bucket *bucket)
1964 {
1965         struct glock_plug plug;
1966         struct list_head *tmp;
1967         struct gfs2_glock *gl;
1968         int entries;
1969
1970         /* Add "plug" to end of bucket list, work back up list from there */
1971         memset(&plug.gl_flags, 0, sizeof(unsigned long));
1972         set_bit(GLF_PLUG, &plug.gl_flags);
1973
1974         write_lock(&bucket->hb_lock);
1975         list_add(&plug.gl_list, &bucket->hb_list);
1976         write_unlock(&bucket->hb_lock);
1977
1978         for (;;) {
1979                 write_lock(&bucket->hb_lock);
1980
1981                 for (;;) {
1982                         tmp = plug.gl_list.next;
1983
1984                         if (tmp == &bucket->hb_list) {
1985                                 list_del(&plug.gl_list);
1986                                 entries = !list_empty(&bucket->hb_list);
1987                                 write_unlock(&bucket->hb_lock);
1988                                 return entries;
1989                         }
1990                         gl = list_entry(tmp, struct gfs2_glock, gl_list);
1991
1992                         /* Move plug up list */
1993                         list_move(&plug.gl_list, &gl->gl_list);
1994
1995                         if (test_bit(GLF_PLUG, &gl->gl_flags))
1996                                 continue;
1997
1998                         /* examiner() must glock_put() */
1999                         gfs2_glock_hold(gl);
2000
2001                         break;
2002                 }
2003
2004                 write_unlock(&bucket->hb_lock);
2005
2006                 examiner(gl);
2007         }
2008 }
2009
2010 /**
2011  * scan_glock - look at a glock and see if we can reclaim it
2012  * @gl: the glock to look at
2013  *
2014  */
2015
2016 static void scan_glock(struct gfs2_glock *gl)
2017 {
2018         if (gfs2_glmutex_trylock(gl)) {
2019                 if (gl->gl_ops == &gfs2_inode_glops) {
2020                         struct gfs2_inode *ip = gl->gl_object;
2021                         if (ip && !atomic_read(&ip->i_count))
2022                                 goto out_schedule;
2023                 }
2024                 if (queue_empty(gl, &gl->gl_holders) &&
2025                     gl->gl_state != LM_ST_UNLOCKED &&
2026                     demote_ok(gl))
2027                         goto out_schedule;
2028
2029                 gfs2_glmutex_unlock(gl);
2030         }
2031
2032         gfs2_glock_put(gl);
2033
2034         return;
2035
2036  out_schedule:
2037         gfs2_glmutex_unlock(gl);
2038         gfs2_glock_schedule_for_reclaim(gl);
2039         gfs2_glock_put(gl);
2040 }
2041
2042 /**
2043  * gfs2_scand_internal - Look for glocks and inodes to toss from memory
2044  * @sdp: the filesystem
2045  *
2046  */
2047
2048 void gfs2_scand_internal(struct gfs2_sbd *sdp)
2049 {
2050         unsigned int x;
2051
2052         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2053                 examine_bucket(scan_glock, sdp, &sdp->sd_gl_hash[x]);
2054                 cond_resched();
2055         }
2056 }
2057
2058 /**
2059  * clear_glock - look at a glock and see if we can free it from glock cache
2060  * @gl: the glock to look at
2061  *
2062  */
2063
2064 static void clear_glock(struct gfs2_glock *gl)
2065 {
2066         struct gfs2_sbd *sdp = gl->gl_sbd;
2067         int released;
2068
2069         spin_lock(&sdp->sd_reclaim_lock);
2070         if (!list_empty(&gl->gl_reclaim)) {
2071                 list_del_init(&gl->gl_reclaim);
2072                 atomic_dec(&sdp->sd_reclaim_count);
2073                 spin_unlock(&sdp->sd_reclaim_lock);
2074                 released = gfs2_glock_put(gl);
2075                 gfs2_assert(sdp, !released);
2076         } else {
2077                 spin_unlock(&sdp->sd_reclaim_lock);
2078         }
2079
2080         if (gfs2_glmutex_trylock(gl)) {
2081                 if (gl->gl_ops == &gfs2_inode_glops) {
2082                         struct gfs2_inode *ip = gl->gl_object;
2083                         if (ip && !atomic_read(&ip->i_count))
2084                                 gfs2_inode_destroy(ip, 1);
2085                 }
2086                 if (queue_empty(gl, &gl->gl_holders) &&
2087                     gl->gl_state != LM_ST_UNLOCKED)
2088                         handle_callback(gl, LM_ST_UNLOCKED);
2089
2090                 gfs2_glmutex_unlock(gl);
2091         }
2092
2093         gfs2_glock_put(gl);
2094 }
2095
2096 /**
2097  * gfs2_gl_hash_clear - Empty out the glock hash table
2098  * @sdp: the filesystem
2099  * @wait: wait until it's all gone
2100  *
2101  * Called when unmounting the filesystem, or when inter-node lock manager
2102  * requests DROPLOCKS because it is running out of capacity.
2103  */
2104
2105 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
2106 {
2107         unsigned long t;
2108         unsigned int x;
2109         int cont;
2110
2111         t = jiffies;
2112
2113         for (;;) {
2114                 cont = 0;
2115
2116                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
2117                         if (examine_bucket(clear_glock, sdp,
2118                                            &sdp->sd_gl_hash[x]))
2119                                 cont = 1;
2120
2121                 if (!wait || !cont)
2122                         break;
2123
2124                 if (time_after_eq(jiffies,
2125                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
2126                         fs_warn(sdp, "Unmount seems to be stalled. "
2127                                      "Dumping lock state...\n");
2128                         gfs2_dump_lockstate(sdp);
2129                         t = jiffies;
2130                 }
2131
2132                 /* invalidate_inodes() requires that the sb inodes list
2133                    not change, but an async completion callback for an
2134                    unlock can occur which does glock_put() which
2135                    can call iput() which will change the sb inodes list.
2136                    invalidate_inodes_mutex prevents glock_put()'s during
2137                    an invalidate_inodes() */
2138
2139                 mutex_lock(&sdp->sd_invalidate_inodes_mutex);
2140                 invalidate_inodes(sdp->sd_vfs);
2141                 mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
2142                 msleep(10);
2143         }
2144 }
2145
2146 /*
2147  *  Diagnostic routines to help debug distributed deadlock
2148  */
2149
2150 /**
2151  * dump_holder - print information about a glock holder
2152  * @str: a string naming the type of holder
2153  * @gh: the glock holder
2154  *
2155  * Returns: 0 on success, -ENOBUFS when we run out of space
2156  */
2157
2158 static int dump_holder(char *str, struct gfs2_holder *gh)
2159 {
2160         unsigned int x;
2161         int error = -ENOBUFS;
2162
2163         printk(KERN_INFO "  %s\n", str);
2164         printk(KERN_INFO "    owner = %ld\n",
2165                    (gh->gh_owner) ? (long)gh->gh_owner->pid : -1);
2166         printk(KERN_INFO "    gh_state = %u\n", gh->gh_state);
2167         printk(KERN_INFO "    gh_flags =");
2168         for (x = 0; x < 32; x++)
2169                 if (gh->gh_flags & (1 << x))
2170                         printk(" %u", x);
2171         printk(" \n");
2172         printk(KERN_INFO "    error = %d\n", gh->gh_error);
2173         printk(KERN_INFO "    gh_iflags =");
2174         for (x = 0; x < 32; x++)
2175                 if (test_bit(x, &gh->gh_iflags))
2176                         printk(" %u", x);
2177         printk(" \n");
2178         print_symbol(KERN_INFO "    initialized at: %s\n", gh->gh_ip);
2179
2180         error = 0;
2181
2182         return error;
2183 }
2184
2185 /**
2186  * dump_inode - print information about an inode
2187  * @ip: the inode
2188  *
2189  * Returns: 0 on success, -ENOBUFS when we run out of space
2190  */
2191
2192 static int dump_inode(struct gfs2_inode *ip)
2193 {
2194         unsigned int x;
2195         int error = -ENOBUFS;
2196
2197         printk(KERN_INFO "  Inode:\n");
2198         printk(KERN_INFO "    num = %llu %llu\n",
2199                     ip->i_num.no_formal_ino, ip->i_num.no_addr);
2200         printk(KERN_INFO "    type = %u\n", IF2DT(ip->i_di.di_mode));
2201         printk(KERN_INFO "    i_count = %d\n", atomic_read(&ip->i_count));
2202         printk(KERN_INFO "    i_flags =");
2203         for (x = 0; x < 32; x++)
2204                 if (test_bit(x, &ip->i_flags))
2205                         printk(" %u", x);
2206         printk(" \n");
2207         printk(KERN_INFO "    vnode = %s\n", (ip->i_vnode) ? "yes" : "no");
2208
2209         error = 0;
2210
2211         return error;
2212 }
2213
2214 /**
2215  * dump_glock - print information about a glock
2216  * @gl: the glock
2217  * @count: where we are in the buffer
2218  *
2219  * Returns: 0 on success, -ENOBUFS when we run out of space
2220  */
2221
2222 static int dump_glock(struct gfs2_glock *gl)
2223 {
2224         struct gfs2_holder *gh;
2225         unsigned int x;
2226         int error = -ENOBUFS;
2227
2228         spin_lock(&gl->gl_spin);
2229
2230         printk(KERN_INFO "Glock (%u, %llu)\n", gl->gl_name.ln_type,
2231                gl->gl_name.ln_number);
2232         printk(KERN_INFO "  gl_flags =");
2233         for (x = 0; x < 32; x++)
2234                 if (test_bit(x, &gl->gl_flags))
2235                         printk(" %u", x);
2236         printk(" \n");
2237         printk(KERN_INFO "  gl_ref = %d\n", atomic_read(&gl->gl_ref.refcount));
2238         printk(KERN_INFO "  gl_state = %u\n", gl->gl_state);
2239         printk(KERN_INFO "  gl_owner = %s\n", gl->gl_owner->comm);
2240         print_symbol(KERN_INFO "  gl_ip = %s\n", gl->gl_ip);
2241         printk(KERN_INFO "  req_gh = %s\n", (gl->gl_req_gh) ? "yes" : "no");
2242         printk(KERN_INFO "  req_bh = %s\n", (gl->gl_req_bh) ? "yes" : "no");
2243         printk(KERN_INFO "  lvb_count = %d\n", atomic_read(&gl->gl_lvb_count));
2244         printk(KERN_INFO "  object = %s\n", (gl->gl_object) ? "yes" : "no");
2245         printk(KERN_INFO "  le = %s\n",
2246                    (list_empty(&gl->gl_le.le_list)) ? "no" : "yes");
2247         printk(KERN_INFO "  reclaim = %s\n",
2248                     (list_empty(&gl->gl_reclaim)) ? "no" : "yes");
2249         if (gl->gl_aspace)
2250                 printk(KERN_INFO "  aspace = %lu\n",
2251                             gl->gl_aspace->i_mapping->nrpages);
2252         else
2253                 printk(KERN_INFO "  aspace = no\n");
2254         printk(KERN_INFO "  ail = %d\n", atomic_read(&gl->gl_ail_count));
2255         if (gl->gl_req_gh) {
2256                 error = dump_holder("Request", gl->gl_req_gh);
2257                 if (error)
2258                         goto out;
2259         }
2260         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
2261                 error = dump_holder("Holder", gh);
2262                 if (error)
2263                         goto out;
2264         }
2265         list_for_each_entry(gh, &gl->gl_waiters1, gh_list) {
2266                 error = dump_holder("Waiter1", gh);
2267                 if (error)
2268                         goto out;
2269         }
2270         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
2271                 error = dump_holder("Waiter2", gh);
2272                 if (error)
2273                         goto out;
2274         }
2275         list_for_each_entry(gh, &gl->gl_waiters3, gh_list) {
2276                 error = dump_holder("Waiter3", gh);
2277                 if (error)
2278                         goto out;
2279         }
2280         if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object) {
2281                 if (!test_bit(GLF_LOCK, &gl->gl_flags) &&
2282                     list_empty(&gl->gl_holders)) {
2283                         error = dump_inode(gl->gl_object);
2284                         if (error)
2285                                 goto out;
2286                 } else {
2287                         error = -ENOBUFS;
2288                         printk(KERN_INFO "  Inode: busy\n");
2289                 }
2290         }
2291
2292         error = 0;
2293
2294  out:
2295         spin_unlock(&gl->gl_spin);
2296
2297         return error;
2298 }
2299
2300 /**
2301  * gfs2_dump_lockstate - print out the current lockstate
2302  * @sdp: the filesystem
2303  * @ub: the buffer to copy the information into
2304  *
2305  * If @ub is NULL, dump the lockstate to the console.
2306  *
2307  */
2308
2309 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
2310 {
2311         struct gfs2_gl_hash_bucket *bucket;
2312         struct gfs2_glock *gl;
2313         unsigned int x;
2314         int error = 0;
2315
2316         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2317                 bucket = &sdp->sd_gl_hash[x];
2318
2319                 read_lock(&bucket->hb_lock);
2320
2321                 list_for_each_entry(gl, &bucket->hb_list, gl_list) {
2322                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2323                                 continue;
2324
2325                         error = dump_glock(gl);
2326                         if (error)
2327                                 break;
2328                 }
2329
2330                 read_unlock(&bucket->hb_lock);
2331
2332                 if (error)
2333                         break;
2334         }
2335
2336
2337         return error;
2338 }
2339