fs/gfs2/glock.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2005 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 <asm/semaphore.h>
  20 #include <asm/uaccess.h>
  21
  22 #include "gfs2.h"
  23 #include "glock.h"
  24 #include "glops.h"
  25 #include "inode.h"
  26 #include "lm.h"
  27 #include "lops.h"
  28 #include "meta_io.h"
  29 #include "quota.h"
  30 #include "super.h"
  31
  32 /*  Must be kept in sync with the beginning of struct gfs2_glock  */
  33 struct glock_plug {
  34         struct list_head gl_list;
  35         unsigned long gl_flags;
  36 };
  37
  38 struct greedy {
  39         struct gfs2_holder gr_gh;
  40         struct work_struct gr_work;
  41 };
  42
  43 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  44
  45 /**
  46  * relaxed_state_ok - is a requested lock compatible with the current lock mode?
  47  * @actual: the current state of the lock
  48  * @requested: the lock state that was requested by the caller
  49  * @flags: the modifier flags passed in by the caller
  50  *
  51  * Returns: 1 if the locks are compatible, 0 otherwise
  52  */
  53
  54 static inline int relaxed_state_ok(unsigned int actual, unsigned requested,
  55                                    int flags)
  56 {
  57         if (actual == requested)
  58                 return 1;
  59
  60         if (flags & GL_EXACT)
  61                 return 0;
  62
  63         if (actual == LM_ST_EXCLUSIVE && requested == LM_ST_SHARED)
  64                 return 1;
  65
  66         if (actual != LM_ST_UNLOCKED && (flags & LM_FLAG_ANY))
  67                 return 1;
  68
  69         return 0;
  70 }
  71
  72 /**
  73  * gl_hash() - Turn glock number into hash bucket number
  74  * @lock: The glock number
  75  *
  76  * Returns: The number of the corresponding hash bucket
  77  */
  78
  79 static unsigned int gl_hash(struct lm_lockname *name)
  80 {
  81         unsigned int h;
  82
  83         h = jhash(&name->ln_number, sizeof(uint64_t), 0);
  84         h = jhash(&name->ln_type, sizeof(unsigned int), h);
  85         h &= GFS2_GL_HASH_MASK;
  86
  87         return h;
  88 }
  89
  90 /**
  91  * glock_free() - Perform a few checks and then release struct gfs2_glock
  92  * @gl: The glock to release
  93  *
  94  * Also calls lock module to release its internal structure for this glock.
  95  *
  96  */
  97
  98 static void glock_free(struct gfs2_glock *gl)
  99 {
 100         struct gfs2_sbd *sdp = gl->gl_sbd;
 101         struct inode *aspace = gl->gl_aspace;
 102
 103         gfs2_lm_put_lock(sdp, gl->gl_lock);
 104
 105         if (aspace)
 106                 gfs2_aspace_put(aspace);
 107
 108         kmem_cache_free(gfs2_glock_cachep, gl);
 109
 110         atomic_dec(&sdp->sd_glock_count);
 111 }
 112
 113 /**
 114  * gfs2_glock_hold() - increment reference count on glock
 115  * @gl: The glock to hold
 116  *
 117  */
 118
 119 void gfs2_glock_hold(struct gfs2_glock *gl)
 120 {
 121         kref_get(&gl->gl_ref);
 122 }
 123
 124 /* All work is done after the return from kref_put() so we
 125    can release the write_lock before the free. */
 126
 127 static void kill_glock(struct kref *kref)
 128 {
 129         struct gfs2_glock *gl = container_of(kref, struct gfs2_glock, gl_ref);
 130         struct gfs2_sbd *sdp = gl->gl_sbd;
 131
 132         gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
 133         gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
 134         gfs2_assert(sdp, list_empty(&gl->gl_holders));
 135         gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
 136         gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
 137         gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
 138 }
 139
 140 /**
 141  * gfs2_glock_put() - Decrement reference count on glock
 142  * @gl: The glock to put
 143  *
 144  */
 145
 146 int gfs2_glock_put(struct gfs2_glock *gl)
 147 {
 148         struct gfs2_sbd *sdp = gl->gl_sbd;
 149         struct gfs2_gl_hash_bucket *bucket = gl->gl_bucket;
 150         int rv = 0;
 151
 152         mutex_lock(&sdp->sd_invalidate_inodes_mutex);
 153
 154         write_lock(&bucket->hb_lock);
 155         if (kref_put(&gl->gl_ref, kill_glock)) {
 156                 list_del_init(&gl->gl_list);
 157                 write_unlock(&bucket->hb_lock);
 158                 glock_free(gl);
 159                 rv = 1;
 160                 goto out;
 161         }
 162         write_unlock(&bucket->hb_lock);
 163  out:
 164         mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
 165         return rv;
 166 }
 167
 168 /**
 169  * queue_empty - check to see if a glock's queue is empty
 170  * @gl: the glock
 171  * @head: the head of the queue to check
 172  *
 173  * This function protects the list in the event that a process already
 174  * has a holder on the list and is adding a second holder for itself.
 175  * The glmutex lock is what generally prevents processes from working
 176  * on the same glock at once, but the special case of adding a second
 177  * holder for yourself ("recursive" locking) doesn't involve locking
 178  * glmutex, making the spin lock necessary.
 179  *
 180  * Returns: 1 if the queue is empty
 181  */
 182
 183 static inline int queue_empty(struct gfs2_glock *gl, struct list_head *head)
 184 {
 185         int empty;
 186         spin_lock(&gl->gl_spin);
 187         empty = list_empty(head);
 188         spin_unlock(&gl->gl_spin);
 189         return empty;
 190 }
 191
 192 /**
 193  * search_bucket() - Find struct gfs2_glock by lock number
 194  * @bucket: the bucket to search
 195  * @name: The lock name
 196  *
 197  * Returns: NULL, or the struct gfs2_glock with the requested number
 198  */
 199
 200 static struct gfs2_glock *search_bucket(struct gfs2_gl_hash_bucket *bucket,
 201                                         struct lm_lockname *name)
 202 {
 203         struct gfs2_glock *gl;
 204
 205         list_for_each_entry(gl, &bucket->hb_list, gl_list) {
 206                 if (test_bit(GLF_PLUG, &gl->gl_flags))
 207                         continue;
 208                 if (!lm_name_equal(&gl->gl_name, name))
 209                         continue;
 210
 211                 kref_get(&gl->gl_ref);
 212
 213                 return gl;
 214         }
 215
 216         return NULL;
 217 }
 218
 219 /**
 220  * gfs2_glock_find() - Find glock by lock number
 221  * @sdp: The GFS2 superblock
 222  * @name: The lock name
 223  *
 224  * Returns: NULL, or the struct gfs2_glock with the requested number
 225  */
 226
 227 struct gfs2_glock *gfs2_glock_find(struct gfs2_sbd *sdp,
 228                                    struct lm_lockname *name)
 229 {
 230         struct gfs2_gl_hash_bucket *bucket = &sdp->sd_gl_hash[gl_hash(name)];
 231         struct gfs2_glock *gl;
 232
 233         read_lock(&bucket->hb_lock);
 234         gl = search_bucket(bucket, name);
 235         read_unlock(&bucket->hb_lock);
 236
 237         return gl;
 238 }
 239
 240 /**
 241  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 242  * @sdp: The GFS2 superblock
 243  * @number: the lock number
 244  * @glops: The glock_operations to use
 245  * @create: If 0, don't create the glock if it doesn't exist
 246  * @glp: the glock is returned here
 247  *
 248  * This does not lock a glock, just finds/creates structures for one.
 249  *
 250  * Returns: errno
 251  */
 252
 253 int gfs2_glock_get(struct gfs2_sbd *sdp, uint64_t number,
 254                    struct gfs2_glock_operations *glops, int create,
 255                    struct gfs2_glock **glp)
 256 {
 257         struct lm_lockname name;
 258         struct gfs2_glock *gl, *tmp;
 259         struct gfs2_gl_hash_bucket *bucket;
 260         int error;
 261
 262         name.ln_number = number;
 263         name.ln_type = glops->go_type;
 264         bucket = &sdp->sd_gl_hash[gl_hash(&name)];
 265
 266         read_lock(&bucket->hb_lock);
 267         gl = search_bucket(bucket, &name);
 268         read_unlock(&bucket->hb_lock);
 269
 270         if (gl || !create) {
 271                 *glp = gl;
 272                 return 0;
 273         }
 274
 275         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
 276         if (!gl)
 277                 return -ENOMEM;
 278
 279         memset(gl, 0, sizeof(struct gfs2_glock));
 280
 281         INIT_LIST_HEAD(&gl->gl_list);
 282         gl->gl_name = name;
 283         kref_init(&gl->gl_ref);
 284
 285         spin_lock_init(&gl->gl_spin);
 286
 287         gl->gl_state = LM_ST_UNLOCKED;
 288         INIT_LIST_HEAD(&gl->gl_holders);
 289         INIT_LIST_HEAD(&gl->gl_waiters1);
 290         INIT_LIST_HEAD(&gl->gl_waiters2);
 291         INIT_LIST_HEAD(&gl->gl_waiters3);
 292
 293         gl->gl_ops = glops;
 294
 295         gl->gl_bucket = bucket;
 296         INIT_LIST_HEAD(&gl->gl_reclaim);
 297
 298         gl->gl_sbd = sdp;
 299
 300         lops_init_le(&gl->gl_le, &gfs2_glock_lops);
 301         INIT_LIST_HEAD(&gl->gl_ail_list);
 302
 303         /* If this glock protects actual on-disk data or metadata blocks,
 304            create a VFS inode to manage the pages/buffers holding them. */
 305         if (glops == &gfs2_inode_glops ||
 306             glops == &gfs2_rgrp_glops ||
 307             glops == &gfs2_meta_glops) {
 308                 gl->gl_aspace = gfs2_aspace_get(sdp);
 309                 if (!gl->gl_aspace) {
 310                         error = -ENOMEM;
 311                         goto fail;
 312                 }
 313         }
 314
 315         error = gfs2_lm_get_lock(sdp, &name, &gl->gl_lock);
 316         if (error)
 317                 goto fail_aspace;
 318
 319         atomic_inc(&sdp->sd_glock_count);
 320
 321         write_lock(&bucket->hb_lock);
 322         tmp = search_bucket(bucket, &name);
 323         if (tmp) {
 324                 write_unlock(&bucket->hb_lock);
 325                 glock_free(gl);
 326                 gl = tmp;
 327         } else {
 328                 list_add_tail(&gl->gl_list, &bucket->hb_list);
 329                 write_unlock(&bucket->hb_lock);
 330         }
 331
 332         *glp = gl;
 333
 334         return 0;
 335
 336  fail_aspace:
 337         if (gl->gl_aspace)
 338                 gfs2_aspace_put(gl->gl_aspace);
 339
 340  fail:
 341         kmem_cache_free(gfs2_glock_cachep, gl);
 342
 343         return error;
 344 }
 345
 346 /**
 347  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 348  * @gl: the glock
 349  * @state: the state we're requesting
 350  * @flags: the modifier flags
 351  * @gh: the holder structure
 352  *
 353  */
 354
 355 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, int flags,
 356                       struct gfs2_holder *gh)
 357 {
 358         INIT_LIST_HEAD(&gh->gh_list);
 359         gh->gh_gl = gl;
 360         gh->gh_owner = (flags & GL_NEVER_RECURSE) ? NULL : current;
 361         gh->gh_state = state;
 362         gh->gh_flags = flags;
 363         gh->gh_error = 0;
 364         gh->gh_iflags = 0;
 365         init_completion(&gh->gh_wait);
 366
 367         if (gh->gh_state == LM_ST_EXCLUSIVE)
 368                 gh->gh_flags |= GL_LOCAL_EXCL;
 369
 370         gfs2_glock_hold(gl);
 371 }
 372
 373 /**
 374  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 375  * @state: the state we're requesting
 376  * @flags: the modifier flags
 377  * @gh: the holder structure
 378  *
 379  * Don't mess with the glock.
 380  *
 381  */
 382
 383 void gfs2_holder_reinit(unsigned int state, int flags, struct gfs2_holder *gh)
 384 {
 385         gh->gh_state = state;
 386         gh->gh_flags = flags;
 387         if (gh->gh_state == LM_ST_EXCLUSIVE)
 388                 gh->gh_flags |= GL_LOCAL_EXCL;
 389
 390         gh->gh_iflags &= 1 << HIF_ALLOCED;
 391 }
 392
 393 /**
 394  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 395  * @gh: the holder structure
 396  *
 397  */
 398
 399 void gfs2_holder_uninit(struct gfs2_holder *gh)
 400 {
 401         gfs2_glock_put(gh->gh_gl);
 402         gh->gh_gl = NULL;
 403 }
 404
 405 /**
 406  * gfs2_holder_get - get a struct gfs2_holder structure
 407  * @gl: the glock
 408  * @state: the state we're requesting
 409  * @flags: the modifier flags
 410  * @gfp_flags: __GFP_NOFAIL
 411  *
 412  * Figure out how big an impact this function has.  Either:
 413  * 1) Replace it with a cache of structures hanging off the struct gfs2_sbd
 414  * 2) Leave it like it is
 415  *
 416  * Returns: the holder structure, NULL on ENOMEM
 417  */
 418
 419 struct gfs2_holder *gfs2_holder_get(struct gfs2_glock *gl, unsigned int state,
 420                                     int flags, gfp_t gfp_flags)
 421 {
 422         struct gfs2_holder *gh;
 423
 424         gh = kmalloc(sizeof(struct gfs2_holder), gfp_flags);
 425         if (!gh)
 426                 return NULL;
 427
 428         gfs2_holder_init(gl, state, flags, gh);
 429         set_bit(HIF_ALLOCED, &gh->gh_iflags);
 430
 431         return gh;
 432 }
 433
 434 /**
 435  * gfs2_holder_put - get rid of a struct gfs2_holder structure
 436  * @gh: the holder structure
 437  *
 438  */
 439
 440 void gfs2_holder_put(struct gfs2_holder *gh)
 441 {
 442         gfs2_holder_uninit(gh);
 443         kfree(gh);
 444 }
 445
 446 /**
 447  * handle_recurse - put other holder structures (marked recursive)
 448  *                  into the holders list
 449  * @gh: the holder structure
 450  *
 451  */
 452
 453 static void handle_recurse(struct gfs2_holder *gh)
 454 {
 455         struct gfs2_glock *gl = gh->gh_gl;
 456         struct gfs2_sbd *sdp = gl->gl_sbd;
 457         struct gfs2_holder *tmp_gh, *safe;
 458         int found = 0;
 459
 460         if (gfs2_assert_warn(sdp, gh->gh_owner))
 461                 return;
 462
 463         list_for_each_entry_safe(tmp_gh, safe, &gl->gl_waiters3, gh_list) {
 464                 if (tmp_gh->gh_owner != gh->gh_owner)
 465                         continue;
 466
 467                 gfs2_assert_warn(sdp,
 468                                  test_bit(HIF_RECURSE, &tmp_gh->gh_iflags));
 469
 470                 list_move_tail(&tmp_gh->gh_list, &gl->gl_holders);
 471                 tmp_gh->gh_error = 0;
 472                 set_bit(HIF_HOLDER, &tmp_gh->gh_iflags);
 473
 474                 complete(&tmp_gh->gh_wait);
 475
 476                 found = 1;
 477         }
 478
 479         gfs2_assert_warn(sdp, found);
 480 }
 481
 482 /**
 483  * do_unrecurse - a recursive holder was just dropped of the waiters3 list
 484  * @gh: the holder
 485  *
 486  * If there is only one other recursive holder, clear its HIF_RECURSE bit.
 487  * If there is more than one, leave them alone.
 488  *
 489  */
 490
 491 static void do_unrecurse(struct gfs2_holder *gh)
 492 {
 493         struct gfs2_glock *gl = gh->gh_gl;
 494         struct gfs2_sbd *sdp = gl->gl_sbd;
 495         struct gfs2_holder *tmp_gh, *last_gh = NULL;
 496         int found = 0;
 497
 498         if (gfs2_assert_warn(sdp, gh->gh_owner))
 499                 return;
 500
 501         list_for_each_entry(tmp_gh, &gl->gl_waiters3, gh_list) {
 502                 if (tmp_gh->gh_owner != gh->gh_owner)
 503                         continue;
 504
 505                 gfs2_assert_warn(sdp,
 506                                  test_bit(HIF_RECURSE, &tmp_gh->gh_iflags));
 507
 508                 if (found)
 509                         return;
 510
 511                 found = 1;
 512                 last_gh = tmp_gh;
 513         }
 514
 515         if (!gfs2_assert_warn(sdp, found))
 516                 clear_bit(HIF_RECURSE, &last_gh->gh_iflags);
 517 }
 518
 519 /**
 520  * rq_mutex - process a mutex request in the queue
 521  * @gh: the glock holder
 522  *
 523  * Returns: 1 if the queue is blocked
 524  */
 525
 526 static int rq_mutex(struct gfs2_holder *gh)
 527 {
 528         struct gfs2_glock *gl = gh->gh_gl;
 529
 530         list_del_init(&gh->gh_list);
 531         /*  gh->gh_error never examined.  */
 532         set_bit(GLF_LOCK, &gl->gl_flags);
 533         complete(&gh->gh_wait);
 534
 535         return 1;
 536 }
 537
 538 /**
 539  * rq_promote - process a promote request in the queue
 540  * @gh: the glock holder
 541  *
 542  * Acquire a new inter-node lock, or change a lock state to more restrictive.
 543  *
 544  * Returns: 1 if the queue is blocked
 545  */
 546
 547 static int rq_promote(struct gfs2_holder *gh)
 548 {
 549         struct gfs2_glock *gl = gh->gh_gl;
 550         struct gfs2_sbd *sdp = gl->gl_sbd;
 551         struct gfs2_glock_operations *glops = gl->gl_ops;
 552         int recurse;
 553
 554         if (!relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 555                 if (list_empty(&gl->gl_holders)) {
 556                         gl->gl_req_gh = gh;
 557                         set_bit(GLF_LOCK, &gl->gl_flags);
 558                         spin_unlock(&gl->gl_spin);
 559
 560                         if (atomic_read(&sdp->sd_reclaim_count) >
 561                             gfs2_tune_get(sdp, gt_reclaim_limit) &&
 562                             !(gh->gh_flags & LM_FLAG_PRIORITY)) {
 563                                 gfs2_reclaim_glock(sdp);
 564                                 gfs2_reclaim_glock(sdp);
 565                         }
 566
 567                         glops->go_xmote_th(gl, gh->gh_state,
 568                                            gh->gh_flags);
 569
 570                         spin_lock(&gl->gl_spin);
 571                 }
 572                 return 1;
 573         }
 574
 575         if (list_empty(&gl->gl_holders)) {
 576                 set_bit(HIF_FIRST, &gh->gh_iflags);
 577                 set_bit(GLF_LOCK, &gl->gl_flags);
 578                 recurse = 0;
 579         } else {
 580                 struct gfs2_holder *next_gh;
 581                 if (gh->gh_flags & GL_LOCAL_EXCL)
 582                         return 1;
 583                 next_gh = list_entry(gl->gl_holders.next, struct gfs2_holder,
 584                                      gh_list);
 585                 if (next_gh->gh_flags & GL_LOCAL_EXCL)
 586                          return 1;
 587                 recurse = test_bit(HIF_RECURSE, &gh->gh_iflags);
 588         }
 589
 590         list_move_tail(&gh->gh_list, &gl->gl_holders);
 591         gh->gh_error = 0;
 592         set_bit(HIF_HOLDER, &gh->gh_iflags);
 593
 594         if (recurse)
 595                 handle_recurse(gh);
 596
 597         complete(&gh->gh_wait);
 598
 599         return 0;
 600 }
 601
 602 /**
 603  * rq_demote - process a demote request in the queue
 604  * @gh: the glock holder
 605  *
 606  * Returns: 1 if the queue is blocked
 607  */
 608
 609 static int rq_demote(struct gfs2_holder *gh)
 610 {
 611         struct gfs2_glock *gl = gh->gh_gl;
 612         struct gfs2_glock_operations *glops = gl->gl_ops;
 613
 614         if (!list_empty(&gl->gl_holders))
 615                 return 1;
 616
 617         if (gl->gl_state == gh->gh_state || gl->gl_state == LM_ST_UNLOCKED) {
 618                 list_del_init(&gh->gh_list);
 619                 gh->gh_error = 0;
 620                 spin_unlock(&gl->gl_spin);
 621                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 622                         gfs2_holder_put(gh);
 623                 else
 624                         complete(&gh->gh_wait);
 625                 spin_lock(&gl->gl_spin);
 626         } else {
 627                 gl->gl_req_gh = gh;
 628                 set_bit(GLF_LOCK, &gl->gl_flags);
 629                 spin_unlock(&gl->gl_spin);
 630
 631                 if (gh->gh_state == LM_ST_UNLOCKED ||
 632                     gl->gl_state != LM_ST_EXCLUSIVE)
 633                         glops->go_drop_th(gl);
 634                 else
 635                         glops->go_xmote_th(gl, gh->gh_state, gh->gh_flags);
 636
 637                 spin_lock(&gl->gl_spin);
 638         }
 639
 640         return 0;
 641 }
 642
 643 /**
 644  * rq_greedy - process a queued request to drop greedy status
 645  * @gh: the glock holder
 646  *
 647  * Returns: 1 if the queue is blocked
 648  */
 649
 650 static int rq_greedy(struct gfs2_holder *gh)
 651 {
 652         struct gfs2_glock *gl = gh->gh_gl;
 653
 654         list_del_init(&gh->gh_list);
 655         /*  gh->gh_error never examined.  */
 656         clear_bit(GLF_GREEDY, &gl->gl_flags);
 657         spin_unlock(&gl->gl_spin);
 658
 659         gfs2_holder_uninit(gh);
 660         kfree(container_of(gh, struct greedy, gr_gh));
 661
 662         spin_lock(&gl->gl_spin);
 663
 664         return 0;
 665 }
 666
 667 /**
 668  * run_queue - process holder structures on a glock
 669  * @gl: the glock
 670  *
 671  */
 672
 673 static void run_queue(struct gfs2_glock *gl)
 674 {
 675         struct gfs2_holder *gh;
 676         int blocked = 1;
 677
 678         for (;;) {
 679                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 680                         break;
 681
 682                 if (!list_empty(&gl->gl_waiters1)) {
 683                         gh = list_entry(gl->gl_waiters1.next,
 684                                         struct gfs2_holder, gh_list);
 685
 686                         if (test_bit(HIF_MUTEX, &gh->gh_iflags))
 687                                 blocked = rq_mutex(gh);
 688                         else
 689                                 gfs2_assert_warn(gl->gl_sbd, 0);
 690
 691                 } else if (!list_empty(&gl->gl_waiters2) &&
 692                            !test_bit(GLF_SKIP_WAITERS2, &gl->gl_flags)) {
 693                         gh = list_entry(gl->gl_waiters2.next,
 694                                         struct gfs2_holder, gh_list);
 695
 696                         if (test_bit(HIF_DEMOTE, &gh->gh_iflags))
 697                                 blocked = rq_demote(gh);
 698                         else if (test_bit(HIF_GREEDY, &gh->gh_iflags))
 699                                 blocked = rq_greedy(gh);
 700                         else
 701                                 gfs2_assert_warn(gl->gl_sbd, 0);
 702
 703                 } else if (!list_empty(&gl->gl_waiters3)) {
 704                         gh = list_entry(gl->gl_waiters3.next,
 705                                         struct gfs2_holder, gh_list);
 706
 707                         if (test_bit(HIF_PROMOTE, &gh->gh_iflags))
 708                                 blocked = rq_promote(gh);
 709                         else
 710                                 gfs2_assert_warn(gl->gl_sbd, 0);
 711
 712                 } else
 713                         break;
 714
 715                 if (blocked)
 716                         break;
 717         }
 718 }
 719
 720 /**
 721  * gfs2_glmutex_lock - acquire a local lock on a glock
 722  * @gl: the glock
 723  *
 724  * Gives caller exclusive access to manipulate a glock structure.
 725  */
 726
 727 void gfs2_glmutex_lock(struct gfs2_glock *gl)
 728 {
 729         struct gfs2_holder gh;
 730
 731         gfs2_holder_init(gl, 0, 0, &gh);
 732         set_bit(HIF_MUTEX, &gh.gh_iflags);
 733
 734         spin_lock(&gl->gl_spin);
 735         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 736                 list_add_tail(&gh.gh_list, &gl->gl_waiters1);
 737         else
 738                 complete(&gh.gh_wait);
 739         spin_unlock(&gl->gl_spin);
 740
 741         wait_for_completion(&gh.gh_wait);
 742         gfs2_holder_uninit(&gh);
 743 }
 744
 745 /**
 746  * gfs2_glmutex_trylock - try to acquire a local lock on a glock
 747  * @gl: the glock
 748  *
 749  * Returns: 1 if the glock is acquired
 750  */
 751
 752 int gfs2_glmutex_trylock(struct gfs2_glock *gl)
 753 {
 754         int acquired = 1;
 755
 756         spin_lock(&gl->gl_spin);
 757         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 758                 acquired = 0;
 759         spin_unlock(&gl->gl_spin);
 760
 761         return acquired;
 762 }
 763
 764 /**
 765  * gfs2_glmutex_unlock - release a local lock on a glock
 766  * @gl: the glock
 767  *
 768  */
 769
 770 void gfs2_glmutex_unlock(struct gfs2_glock *gl)
 771 {
 772         spin_lock(&gl->gl_spin);
 773         clear_bit(GLF_LOCK, &gl->gl_flags);
 774         run_queue(gl);
 775         spin_unlock(&gl->gl_spin);
 776 }
 777
 778 /**
 779  * handle_callback - add a demote request to a lock's queue
 780  * @gl: the glock
 781  * @state: the state the caller wants us to change to
 782  *
 783  */
 784
 785 static void handle_callback(struct gfs2_glock *gl, unsigned int state)
 786 {
 787         struct gfs2_holder *gh, *new_gh = NULL;
 788
 789  restart:
 790         spin_lock(&gl->gl_spin);
 791
 792         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
 793                 if (test_bit(HIF_DEMOTE, &gh->gh_iflags) &&
 794                     gl->gl_req_gh != gh) {
 795                         if (gh->gh_state != state)
 796                                 gh->gh_state = LM_ST_UNLOCKED;
 797                         goto out;
 798                 }
 799         }
 800
 801         if (new_gh) {
 802                 list_add_tail(&new_gh->gh_list, &gl->gl_waiters2);
 803                 new_gh = NULL;
 804         } else {
 805                 spin_unlock(&gl->gl_spin);
 806
 807                 new_gh = gfs2_holder_get(gl, state,
 808                                          LM_FLAG_TRY | GL_NEVER_RECURSE,
 809                                          GFP_KERNEL | __GFP_NOFAIL),
 810                 set_bit(HIF_DEMOTE, &new_gh->gh_iflags);
 811                 set_bit(HIF_DEALLOC, &new_gh->gh_iflags);
 812
 813                 goto restart;
 814         }
 815
 816  out:
 817         spin_unlock(&gl->gl_spin);
 818
 819         if (new_gh)
 820                 gfs2_holder_put(new_gh);
 821 }
 822
 823 /**
 824  * state_change - record that the glock is now in a different state
 825  * @gl: the glock
 826  * @new_state the new state
 827  *
 828  */
 829
 830 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 831 {
 832         struct gfs2_sbd *sdp = gl->gl_sbd;
 833         int held1, held2;
 834
 835         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 836         held2 = (new_state != LM_ST_UNLOCKED);
 837
 838         if (held1 != held2) {
 839                 if (held2) {
 840                         atomic_inc(&sdp->sd_glock_held_count);
 841                         gfs2_glock_hold(gl);
 842                 } else {
 843                         atomic_dec(&sdp->sd_glock_held_count);
 844                         gfs2_glock_put(gl);
 845                 }
 846         }
 847
 848         gl->gl_state = new_state;
 849 }
 850
 851 /**
 852  * xmote_bh - Called after the lock module is done acquiring a lock
 853  * @gl: The glock in question
 854  * @ret: the int returned from the lock module
 855  *
 856  */
 857
 858 static void xmote_bh(struct gfs2_glock *gl, unsigned int ret)
 859 {
 860         struct gfs2_sbd *sdp = gl->gl_sbd;
 861         struct gfs2_glock_operations *glops = gl->gl_ops;
 862         struct gfs2_holder *gh = gl->gl_req_gh;
 863         int prev_state = gl->gl_state;
 864         int op_done = 1;
 865
 866         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 867         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 868         gfs2_assert_warn(sdp, !(ret & LM_OUT_ASYNC));
 869
 870         state_change(gl, ret & LM_OUT_ST_MASK);
 871
 872         if (prev_state != LM_ST_UNLOCKED && !(ret & LM_OUT_CACHEABLE)) {
 873                 if (glops->go_inval)
 874                         glops->go_inval(gl, DIO_METADATA | DIO_DATA);
 875         } else if (gl->gl_state == LM_ST_DEFERRED) {
 876                 /* We might not want to do this here.
 877                    Look at moving to the inode glops. */
 878                 if (glops->go_inval)
 879                         glops->go_inval(gl, DIO_DATA);
 880         }
 881
 882         /*  Deal with each possible exit condition  */
 883
 884         if (!gh)
 885                 gl->gl_stamp = jiffies;
 886
 887         else if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
 888                 spin_lock(&gl->gl_spin);
 889                 list_del_init(&gh->gh_list);
 890                 gh->gh_error = -EIO;
 891                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 892                         do_unrecurse(gh);
 893                 spin_unlock(&gl->gl_spin);
 894
 895         } else if (test_bit(HIF_DEMOTE, &gh->gh_iflags)) {
 896                 spin_lock(&gl->gl_spin);
 897                 list_del_init(&gh->gh_list);
 898                 if (gl->gl_state == gh->gh_state ||
 899                     gl->gl_state == LM_ST_UNLOCKED)
 900                         gh->gh_error = 0;
 901                 else {
 902                         if (gfs2_assert_warn(sdp, gh->gh_flags &
 903                                         (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) == -1)
 904                                 fs_warn(sdp, "ret = 0x%.8X\n", ret);
 905                         gh->gh_error = GLR_TRYFAILED;
 906                 }
 907                 spin_unlock(&gl->gl_spin);
 908
 909                 if (ret & LM_OUT_CANCELED)
 910                         handle_callback(gl, LM_ST_UNLOCKED); /* Lame */
 911
 912         } else if (ret & LM_OUT_CANCELED) {
 913                 spin_lock(&gl->gl_spin);
 914                 list_del_init(&gh->gh_list);
 915                 gh->gh_error = GLR_CANCELED;
 916                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 917                         do_unrecurse(gh);
 918                 spin_unlock(&gl->gl_spin);
 919
 920         } else if (relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 921                 spin_lock(&gl->gl_spin);
 922                 list_move_tail(&gh->gh_list, &gl->gl_holders);
 923                 gh->gh_error = 0;
 924                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 925                 spin_unlock(&gl->gl_spin);
 926
 927                 set_bit(HIF_FIRST, &gh->gh_iflags);
 928
 929                 op_done = 0;
 930
 931         } else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 932                 spin_lock(&gl->gl_spin);
 933                 list_del_init(&gh->gh_list);
 934                 gh->gh_error = GLR_TRYFAILED;
 935                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 936                         do_unrecurse(gh);
 937                 spin_unlock(&gl->gl_spin);
 938
 939         } else {
 940                 if (gfs2_assert_withdraw(sdp, 0) == -1)
 941                         fs_err(sdp, "ret = 0x%.8X\n", ret);
 942         }
 943
 944         if (glops->go_xmote_bh)
 945                 glops->go_xmote_bh(gl);
 946
 947         if (op_done) {
 948                 spin_lock(&gl->gl_spin);
 949                 gl->gl_req_gh = NULL;
 950                 gl->gl_req_bh = NULL;
 951                 clear_bit(GLF_LOCK, &gl->gl_flags);
 952                 run_queue(gl);
 953                 spin_unlock(&gl->gl_spin);
 954         }
 955
 956         gfs2_glock_put(gl);
 957
 958         if (gh) {
 959                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 960                         gfs2_holder_put(gh);
 961                 else
 962                         complete(&gh->gh_wait);
 963         }
 964 }
 965
 966 /**
 967  * gfs2_glock_xmote_th - Call into the lock module to acquire or change a glock
 968  * @gl: The glock in question
 969  * @state: the requested state
 970  * @flags: modifier flags to the lock call
 971  *
 972  */
 973
 974 void gfs2_glock_xmote_th(struct gfs2_glock *gl, unsigned int state, int flags)
 975 {
 976         struct gfs2_sbd *sdp = gl->gl_sbd;
 977         struct gfs2_glock_operations *glops = gl->gl_ops;
 978         int lck_flags = flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB |
 979                                  LM_FLAG_NOEXP | LM_FLAG_ANY |
 980                                  LM_FLAG_PRIORITY);
 981         unsigned int lck_ret;
 982
 983         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 984         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 985         gfs2_assert_warn(sdp, state != LM_ST_UNLOCKED);
 986         gfs2_assert_warn(sdp, state != gl->gl_state);
 987
 988         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 989                 if (glops->go_sync)
 990                         glops->go_sync(gl,
 991                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
 992         }
 993
 994         gfs2_glock_hold(gl);
 995         gl->gl_req_bh = xmote_bh;
 996
 997         atomic_inc(&sdp->sd_lm_lock_calls);
 998
 999         lck_ret = gfs2_lm_lock(sdp, gl->gl_lock, gl->gl_state, state,
1000                                lck_flags);
1001
1002         if (gfs2_assert_withdraw(sdp, !(lck_ret & LM_OUT_ERROR)))
1003                 return;
1004
1005         if (lck_ret & LM_OUT_ASYNC)
1006                 gfs2_assert_warn(sdp, lck_ret == LM_OUT_ASYNC);
1007         else
1008                 xmote_bh(gl, lck_ret);
1009 }
1010
1011 /**
1012  * drop_bh - Called after a lock module unlock completes
1013  * @gl: the glock
1014  * @ret: the return status
1015  *
1016  * Doesn't wake up the process waiting on the struct gfs2_holder (if any)
1017  * Doesn't drop the reference on the glock the top half took out
1018  *
1019  */
1020
1021 static void drop_bh(struct gfs2_glock *gl, unsigned int ret)
1022 {
1023         struct gfs2_sbd *sdp = gl->gl_sbd;
1024         struct gfs2_glock_operations *glops = gl->gl_ops;
1025         struct gfs2_holder *gh = gl->gl_req_gh;
1026
1027         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1028
1029         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1030         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1031         gfs2_assert_warn(sdp, !ret);
1032
1033         state_change(gl, LM_ST_UNLOCKED);
1034
1035         if (glops->go_inval)
1036                 glops->go_inval(gl, DIO_METADATA | DIO_DATA);
1037
1038         if (gh) {
1039                 spin_lock(&gl->gl_spin);
1040                 list_del_init(&gh->gh_list);
1041                 gh->gh_error = 0;
1042                 spin_unlock(&gl->gl_spin);
1043         }
1044
1045         if (glops->go_drop_bh)
1046                 glops->go_drop_bh(gl);
1047
1048         spin_lock(&gl->gl_spin);
1049         gl->gl_req_gh = NULL;
1050         gl->gl_req_bh = NULL;
1051         clear_bit(GLF_LOCK, &gl->gl_flags);
1052         run_queue(gl);
1053         spin_unlock(&gl->gl_spin);
1054
1055         gfs2_glock_put(gl);
1056
1057         if (gh) {
1058                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
1059                         gfs2_holder_put(gh);
1060                 else
1061                         complete(&gh->gh_wait);
1062         }
1063 }
1064
1065 /**
1066  * gfs2_glock_drop_th - call into the lock module to unlock a lock
1067  * @gl: the glock
1068  *
1069  */
1070
1071 void gfs2_glock_drop_th(struct gfs2_glock *gl)
1072 {
1073         struct gfs2_sbd *sdp = gl->gl_sbd;
1074         struct gfs2_glock_operations *glops = gl->gl_ops;
1075         unsigned int ret;
1076
1077         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1078         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1079         gfs2_assert_warn(sdp, gl->gl_state != LM_ST_UNLOCKED);
1080
1081         if (gl->gl_state == LM_ST_EXCLUSIVE) {
1082                 if (glops->go_sync)
1083                         glops->go_sync(gl,
1084                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
1085         }
1086
1087         gfs2_glock_hold(gl);
1088         gl->gl_req_bh = drop_bh;
1089
1090         atomic_inc(&sdp->sd_lm_unlock_calls);
1091
1092         ret = gfs2_lm_unlock(sdp, gl->gl_lock, gl->gl_state);
1093
1094         if (gfs2_assert_withdraw(sdp, !(ret & LM_OUT_ERROR)))
1095                 return;
1096
1097         if (!ret)
1098                 drop_bh(gl, ret);
1099         else
1100                 gfs2_assert_warn(sdp, ret == LM_OUT_ASYNC);
1101 }
1102
1103 /**
1104  * do_cancels - cancel requests for locks stuck waiting on an expire flag
1105  * @gh: the LM_FLAG_PRIORITY holder waiting to acquire the lock
1106  *
1107  * Don't cancel GL_NOCANCEL requests.
1108  */
1109
1110 static void do_cancels(struct gfs2_holder *gh)
1111 {
1112         struct gfs2_glock *gl = gh->gh_gl;
1113
1114         spin_lock(&gl->gl_spin);
1115
1116         while (gl->gl_req_gh != gh &&
1117                !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1118                !list_empty(&gh->gh_list)) {
1119                 if (gl->gl_req_bh &&
1120                     !(gl->gl_req_gh &&
1121                       (gl->gl_req_gh->gh_flags & GL_NOCANCEL))) {
1122                         spin_unlock(&gl->gl_spin);
1123                         gfs2_lm_cancel(gl->gl_sbd, gl->gl_lock);
1124                         msleep(100);
1125                         spin_lock(&gl->gl_spin);
1126                 } else {
1127                         spin_unlock(&gl->gl_spin);
1128                         msleep(100);
1129                         spin_lock(&gl->gl_spin);
1130                 }
1131         }
1132
1133         spin_unlock(&gl->gl_spin);
1134 }
1135
1136 /**
1137  * glock_wait_internal - wait on a glock acquisition
1138  * @gh: the glock holder
1139  *
1140  * Returns: 0 on success
1141  */
1142
1143 static int glock_wait_internal(struct gfs2_holder *gh)
1144 {
1145         struct gfs2_glock *gl = gh->gh_gl;
1146         struct gfs2_sbd *sdp = gl->gl_sbd;
1147         struct gfs2_glock_operations *glops = gl->gl_ops;
1148
1149         if (test_bit(HIF_ABORTED, &gh->gh_iflags))
1150                 return -EIO;
1151
1152         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1153                 spin_lock(&gl->gl_spin);
1154                 if (gl->gl_req_gh != gh &&
1155                     !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1156                     !list_empty(&gh->gh_list)) {
1157                         list_del_init(&gh->gh_list);
1158                         gh->gh_error = GLR_TRYFAILED;
1159                         if (test_bit(HIF_RECURSE, &gh->gh_iflags))
1160                                 do_unrecurse(gh);
1161                         run_queue(gl);
1162                         spin_unlock(&gl->gl_spin);
1163                         return gh->gh_error;
1164                 }
1165                 spin_unlock(&gl->gl_spin);
1166         }
1167
1168         if (gh->gh_flags & LM_FLAG_PRIORITY)
1169                 do_cancels(gh);
1170
1171         wait_for_completion(&gh->gh_wait);
1172
1173         if (gh->gh_error)
1174                 return gh->gh_error;
1175
1176         gfs2_assert_withdraw(sdp, test_bit(HIF_HOLDER, &gh->gh_iflags));
1177         gfs2_assert_withdraw(sdp, relaxed_state_ok(gl->gl_state,
1178                                                    gh->gh_state,
1179                                                    gh->gh_flags));
1180
1181         if (test_bit(HIF_FIRST, &gh->gh_iflags)) {
1182                 gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1183
1184                 if (glops->go_lock) {
1185                         gh->gh_error = glops->go_lock(gh);
1186                         if (gh->gh_error) {
1187                                 spin_lock(&gl->gl_spin);
1188                                 list_del_init(&gh->gh_list);
1189                                 if (test_and_clear_bit(HIF_RECURSE,
1190                                                        &gh->gh_iflags))
1191                                         do_unrecurse(gh);
1192                                 spin_unlock(&gl->gl_spin);
1193                         }
1194                 }
1195
1196                 spin_lock(&gl->gl_spin);
1197                 gl->gl_req_gh = NULL;
1198                 gl->gl_req_bh = NULL;
1199                 clear_bit(GLF_LOCK, &gl->gl_flags);
1200                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
1201                         handle_recurse(gh);
1202                 run_queue(gl);
1203                 spin_unlock(&gl->gl_spin);
1204         }
1205
1206         return gh->gh_error;
1207 }
1208
1209 static inline struct gfs2_holder *
1210 find_holder_by_owner(struct list_head *head, struct task_struct *owner)
1211 {
1212         struct gfs2_holder *gh;
1213
1214         list_for_each_entry(gh, head, gh_list) {
1215                 if (gh->gh_owner == owner)
1216                         return gh;
1217         }
1218
1219         return NULL;
1220 }
1221
1222 /**
1223  * recurse_check -
1224  *
1225  * Make sure the new holder is compatible with the pre-existing one.
1226  *
1227  */
1228
1229 static int recurse_check(struct gfs2_holder *existing, struct gfs2_holder *new,
1230                          unsigned int state)
1231 {
1232         struct gfs2_sbd *sdp = existing->gh_gl->gl_sbd;
1233
1234         if (gfs2_assert_warn(sdp, (new->gh_flags & LM_FLAG_ANY) ||
1235                                   !(existing->gh_flags & LM_FLAG_ANY)))
1236                 goto fail;
1237
1238         if (gfs2_assert_warn(sdp, (existing->gh_flags & GL_LOCAL_EXCL) ||
1239                                   !(new->gh_flags & GL_LOCAL_EXCL)))
1240                 goto fail;
1241
1242         if (gfs2_assert_warn(sdp, relaxed_state_ok(state, new->gh_state,
1243                                                    new->gh_flags)))
1244                 goto fail;
1245
1246         return 0;
1247
1248  fail:
1249         set_bit(HIF_ABORTED, &new->gh_iflags);
1250         return -EINVAL;
1251 }
1252
1253 /**
1254  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1255  * @gh: the holder structure to add
1256  *
1257  */
1258
1259 static void add_to_queue(struct gfs2_holder *gh)
1260 {
1261         struct gfs2_glock *gl = gh->gh_gl;
1262         struct gfs2_holder *existing;
1263
1264         if (!gh->gh_owner)
1265                 goto out;
1266
1267         existing = find_holder_by_owner(&gl->gl_holders, gh->gh_owner);
1268         if (existing) {
1269                 if (recurse_check(existing, gh, gl->gl_state))
1270                         return;
1271
1272                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1273                 set_bit(HIF_HOLDER, &gh->gh_iflags);
1274
1275                 gh->gh_error = 0;
1276                 complete(&gh->gh_wait);
1277
1278                 return;
1279         }
1280
1281         existing = find_holder_by_owner(&gl->gl_waiters3, gh->gh_owner);
1282         if (existing) {
1283                 if (recurse_check(existing, gh, existing->gh_state))
1284                         return;
1285
1286                 set_bit(HIF_RECURSE, &gh->gh_iflags);
1287                 set_bit(HIF_RECURSE, &existing->gh_iflags);
1288
1289                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1290
1291                 return;
1292         }
1293
1294  out:
1295         if (gh->gh_flags & LM_FLAG_PRIORITY)
1296                 list_add(&gh->gh_list, &gl->gl_waiters3);
1297         else
1298                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1299 }
1300
1301 /**
1302  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1303  * @gh: the holder structure
1304  *
1305  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1306  *
1307  * Returns: 0, GLR_TRYFAILED, or errno on failure
1308  */
1309
1310 int gfs2_glock_nq(struct gfs2_holder *gh)
1311 {
1312         struct gfs2_glock *gl = gh->gh_gl;
1313         struct gfs2_sbd *sdp = gl->gl_sbd;
1314         int error = 0;
1315
1316         atomic_inc(&sdp->sd_glock_nq_calls);
1317
1318  restart:
1319         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
1320                 set_bit(HIF_ABORTED, &gh->gh_iflags);
1321                 return -EIO;
1322         }
1323
1324         set_bit(HIF_PROMOTE, &gh->gh_iflags);
1325
1326         spin_lock(&gl->gl_spin);
1327         add_to_queue(gh);
1328         run_queue(gl);
1329         spin_unlock(&gl->gl_spin);
1330
1331         if (!(gh->gh_flags & GL_ASYNC)) {
1332                 error = glock_wait_internal(gh);
1333                 if (error == GLR_CANCELED) {
1334                         msleep(1000);
1335                         goto restart;
1336                 }
1337         }
1338
1339         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1340
1341         return error;
1342 }
1343
1344 /**
1345  * gfs2_glock_poll - poll to see if an async request has been completed
1346  * @gh: the holder
1347  *
1348  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1349  */
1350
1351 int gfs2_glock_poll(struct gfs2_holder *gh)
1352 {
1353         struct gfs2_glock *gl = gh->gh_gl;
1354         int ready = 0;
1355
1356         spin_lock(&gl->gl_spin);
1357
1358         if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1359                 ready = 1;
1360         else if (list_empty(&gh->gh_list)) {
1361                 if (gh->gh_error == GLR_CANCELED) {
1362                         spin_unlock(&gl->gl_spin);
1363                         msleep(1000);
1364                         if (gfs2_glock_nq(gh))
1365                                 return 1;
1366                         return 0;
1367                 } else
1368                         ready = 1;
1369         }
1370
1371         spin_unlock(&gl->gl_spin);
1372
1373         return ready;
1374 }
1375
1376 /**
1377  * gfs2_glock_wait - wait for a lock acquisition that ended in a GLR_ASYNC
1378  * @gh: the holder structure
1379  *
1380  * Returns: 0, GLR_TRYFAILED, or errno on failure
1381  */
1382
1383 int gfs2_glock_wait(struct gfs2_holder *gh)
1384 {
1385         int error;
1386
1387         error = glock_wait_internal(gh);
1388         if (error == GLR_CANCELED) {
1389                 msleep(1000);
1390                 gh->gh_flags &= ~GL_ASYNC;
1391                 error = gfs2_glock_nq(gh);
1392         }
1393
1394         return error;
1395 }
1396
1397 /**
1398  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1399  * @gh: the glock holder
1400  *
1401  */
1402
1403 void gfs2_glock_dq(struct gfs2_holder *gh)
1404 {
1405         struct gfs2_glock *gl = gh->gh_gl;
1406         struct gfs2_sbd *sdp = gl->gl_sbd;
1407         struct gfs2_glock_operations *glops = gl->gl_ops;
1408
1409         atomic_inc(&sdp->sd_glock_dq_calls);
1410
1411         if (gh->gh_flags & GL_SYNC)
1412                 set_bit(GLF_SYNC, &gl->gl_flags);
1413
1414         if (gh->gh_flags & GL_NOCACHE)
1415                 handle_callback(gl, LM_ST_UNLOCKED);
1416
1417         gfs2_glmutex_lock(gl);
1418
1419         spin_lock(&gl->gl_spin);
1420         list_del_init(&gh->gh_list);
1421
1422         if (list_empty(&gl->gl_holders)) {
1423                 spin_unlock(&gl->gl_spin);
1424
1425                 if (glops->go_unlock)
1426                         glops->go_unlock(gh);
1427
1428                 if (test_bit(GLF_SYNC, &gl->gl_flags)) {
1429                         if (glops->go_sync)
1430                                 glops->go_sync(gl, DIO_METADATA | DIO_DATA);
1431                 }
1432
1433                 gl->gl_stamp = jiffies;
1434
1435                 spin_lock(&gl->gl_spin);
1436         }
1437
1438         clear_bit(GLF_LOCK, &gl->gl_flags);
1439         run_queue(gl);
1440         spin_unlock(&gl->gl_spin);
1441 }
1442
1443 /**
1444  * gfs2_glock_prefetch - Try to prefetch a glock
1445  * @gl: the glock
1446  * @state: the state to prefetch in
1447  * @flags: flags passed to go_xmote_th()
1448  *
1449  */
1450
1451 void gfs2_glock_prefetch(struct gfs2_glock *gl, unsigned int state, int flags)
1452 {
1453         struct gfs2_glock_operations *glops = gl->gl_ops;
1454
1455         spin_lock(&gl->gl_spin);
1456
1457         if (test_bit(GLF_LOCK, &gl->gl_flags) ||
1458             !list_empty(&gl->gl_holders) ||
1459             !list_empty(&gl->gl_waiters1) ||
1460             !list_empty(&gl->gl_waiters2) ||
1461             !list_empty(&gl->gl_waiters3) ||
1462             relaxed_state_ok(gl->gl_state, state, flags)) {
1463                 spin_unlock(&gl->gl_spin);
1464                 return;
1465         }
1466
1467         set_bit(GLF_PREFETCH, &gl->gl_flags);
1468         set_bit(GLF_LOCK, &gl->gl_flags);
1469         spin_unlock(&gl->gl_spin);
1470
1471         glops->go_xmote_th(gl, state, flags);
1472
1473         atomic_inc(&gl->gl_sbd->sd_glock_prefetch_calls);
1474 }
1475
1476 /**
1477  * gfs2_glock_force_drop - Force a glock to be uncached
1478  * @gl: the glock
1479  *
1480  */
1481
1482 void gfs2_glock_force_drop(struct gfs2_glock *gl)
1483 {
1484         struct gfs2_holder gh;
1485
1486         gfs2_holder_init(gl, LM_ST_UNLOCKED, GL_NEVER_RECURSE, &gh);
1487         set_bit(HIF_DEMOTE, &gh.gh_iflags);
1488
1489         spin_lock(&gl->gl_spin);
1490         list_add_tail(&gh.gh_list, &gl->gl_waiters2);
1491         run_queue(gl);
1492         spin_unlock(&gl->gl_spin);
1493
1494         wait_for_completion(&gh.gh_wait);
1495         gfs2_holder_uninit(&gh);
1496 }
1497
1498 static void greedy_work(void *data)
1499 {
1500         struct greedy *gr = (struct greedy *)data;
1501         struct gfs2_holder *gh = &gr->gr_gh;
1502         struct gfs2_glock *gl = gh->gh_gl;
1503         struct gfs2_glock_operations *glops = gl->gl_ops;
1504
1505         clear_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1506
1507         if (glops->go_greedy)
1508                 glops->go_greedy(gl);
1509
1510         spin_lock(&gl->gl_spin);
1511
1512         if (list_empty(&gl->gl_waiters2)) {
1513                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1514                 spin_unlock(&gl->gl_spin);
1515                 gfs2_holder_uninit(gh);
1516                 kfree(gr);
1517         } else {
1518                 gfs2_glock_hold(gl);
1519                 list_add_tail(&gh->gh_list, &gl->gl_waiters2);
1520                 run_queue(gl);
1521                 spin_unlock(&gl->gl_spin);
1522                 gfs2_glock_put(gl);
1523         }
1524 }
1525
1526 /**
1527  * gfs2_glock_be_greedy -
1528  * @gl:
1529  * @time:
1530  *
1531  * Returns: 0 if go_greedy will be called, 1 otherwise
1532  */
1533
1534 int gfs2_glock_be_greedy(struct gfs2_glock *gl, unsigned int time)
1535 {
1536         struct greedy *gr;
1537         struct gfs2_holder *gh;
1538
1539         if (!time ||
1540             gl->gl_sbd->sd_args.ar_localcaching ||
1541             test_and_set_bit(GLF_GREEDY, &gl->gl_flags))
1542                 return 1;
1543
1544         gr = kmalloc(sizeof(struct greedy), GFP_KERNEL);
1545         if (!gr) {
1546                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1547                 return 1;
1548         }
1549         gh = &gr->gr_gh;
1550
1551         gfs2_holder_init(gl, 0, GL_NEVER_RECURSE, gh);
1552         set_bit(HIF_GREEDY, &gh->gh_iflags);
1553         INIT_WORK(&gr->gr_work, greedy_work, gr);
1554
1555         set_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1556         schedule_delayed_work(&gr->gr_work, time);
1557
1558         return 0;
1559 }
1560
1561 /**
1562  * gfs2_glock_nq_init - intialize a holder and enqueue it on a glock
1563  * @gl: the glock
1564  * @state: the state we're requesting
1565  * @flags: the modifier flags
1566  * @gh: the holder structure
1567  *
1568  * Returns: 0, GLR_*, or errno
1569  */
1570
1571 int gfs2_glock_nq_init(struct gfs2_glock *gl, unsigned int state, int flags,
1572                        struct gfs2_holder *gh)
1573 {
1574         int error;
1575
1576         gfs2_holder_init(gl, state, flags, gh);
1577
1578         error = gfs2_glock_nq(gh);
1579         if (error)
1580                 gfs2_holder_uninit(gh);
1581
1582         return error;
1583 }
1584
1585 /**
1586  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1587  * @gh: the holder structure
1588  *
1589  */
1590
1591 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1592 {
1593         gfs2_glock_dq(gh);
1594         gfs2_holder_uninit(gh);
1595 }
1596
1597 /**
1598  * gfs2_glock_nq_num - acquire a glock based on lock number
1599  * @sdp: the filesystem
1600  * @number: the lock number
1601  * @glops: the glock operations for the type of glock
1602  * @state: the state to acquire the glock in
1603  * @flags: modifier flags for the aquisition
1604  * @gh: the struct gfs2_holder
1605  *
1606  * Returns: errno
1607  */
1608
1609 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, uint64_t number,
1610                       struct gfs2_glock_operations *glops, unsigned int state,
1611                       int flags, struct gfs2_holder *gh)
1612 {
1613         struct gfs2_glock *gl;
1614         int error;
1615
1616         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1617         if (!error) {
1618                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1619                 gfs2_glock_put(gl);
1620         }
1621
1622         return error;
1623 }
1624
1625 /**
1626  * glock_compare - Compare two struct gfs2_glock structures for sorting
1627  * @arg_a: the first structure
1628  * @arg_b: the second structure
1629  *
1630  */
1631
1632 static int glock_compare(const void *arg_a, const void *arg_b)
1633 {
1634         struct gfs2_holder *gh_a = *(struct gfs2_holder **)arg_a;
1635         struct gfs2_holder *gh_b = *(struct gfs2_holder **)arg_b;
1636         struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1637         struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1638         int ret = 0;
1639
1640         if (a->ln_number > b->ln_number)
1641                 ret = 1;
1642         else if (a->ln_number < b->ln_number)
1643                 ret = -1;
1644         else {
1645                 if (gh_a->gh_state == LM_ST_SHARED &&
1646                     gh_b->gh_state == LM_ST_EXCLUSIVE)
1647                         ret = 1;
1648                 else if (!(gh_a->gh_flags & GL_LOCAL_EXCL) &&
1649                          (gh_b->gh_flags & GL_LOCAL_EXCL))
1650                         ret = 1;
1651         }
1652
1653         return ret;
1654 }
1655
1656 /**
1657  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1658  * @num_gh: the number of structures
1659  * @ghs: an array of struct gfs2_holder structures
1660  *
1661  * Returns: 0 on success (all glocks acquired),
1662  *          errno on failure (no glocks acquired)
1663  */
1664
1665 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1666                      struct gfs2_holder **p)
1667 {
1668         unsigned int x;
1669         int error = 0;
1670
1671         for (x = 0; x < num_gh; x++)
1672                 p[x] = &ghs[x];
1673
1674         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1675
1676         for (x = 0; x < num_gh; x++) {
1677                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1678
1679                 error = gfs2_glock_nq(p[x]);
1680                 if (error) {
1681                         while (x--)
1682                                 gfs2_glock_dq(p[x]);
1683                         break;
1684                 }
1685         }
1686
1687         return error;
1688 }
1689
1690 /**
1691  * gfs2_glock_nq_m - acquire multiple glocks
1692  * @num_gh: the number of structures
1693  * @ghs: an array of struct gfs2_holder structures
1694  *
1695  * Figure out how big an impact this function has.  Either:
1696  * 1) Replace this code with code that calls gfs2_glock_prefetch()
1697  * 2) Forget async stuff and just call nq_m_sync()
1698  * 3) Leave it like it is
1699  *
1700  * Returns: 0 on success (all glocks acquired),
1701  *          errno on failure (no glocks acquired)
1702  */
1703
1704 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1705 {
1706         int *e;
1707         unsigned int x;
1708         int borked = 0, serious = 0;
1709         int error = 0;
1710
1711         if (!num_gh)
1712                 return 0;
1713
1714         if (num_gh == 1) {
1715                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1716                 return gfs2_glock_nq(ghs);
1717         }
1718
1719         e = kcalloc(num_gh, sizeof(struct gfs2_holder *), GFP_KERNEL);
1720         if (!e)
1721                 return -ENOMEM;
1722
1723         for (x = 0; x < num_gh; x++) {
1724                 ghs[x].gh_flags |= LM_FLAG_TRY | GL_ASYNC;
1725                 error = gfs2_glock_nq(&ghs[x]);
1726                 if (error) {
1727                         borked = 1;
1728                         serious = error;
1729                         num_gh = x;
1730                         break;
1731                 }
1732         }
1733
1734         for (x = 0; x < num_gh; x++) {
1735                 error = e[x] = glock_wait_internal(&ghs[x]);
1736                 if (error) {
1737                         borked = 1;
1738                         if (error != GLR_TRYFAILED && error != GLR_CANCELED)
1739                                 serious = error;
1740                 }
1741         }
1742
1743         if (!borked) {
1744                 kfree(e);
1745                 return 0;
1746         }
1747
1748         for (x = 0; x < num_gh; x++)
1749                 if (!e[x])
1750                         gfs2_glock_dq(&ghs[x]);
1751
1752         if (serious)
1753                 error = serious;
1754         else {
1755                 for (x = 0; x < num_gh; x++)
1756                         gfs2_holder_reinit(ghs[x].gh_state, ghs[x].gh_flags,
1757                                           &ghs[x]);
1758                 error = nq_m_sync(num_gh, ghs, (struct gfs2_holder **)e);
1759         }
1760
1761         kfree(e);
1762
1763         return error;
1764 }
1765
1766 /**
1767  * gfs2_glock_dq_m - release multiple glocks
1768  * @num_gh: the number of structures
1769  * @ghs: an array of struct gfs2_holder structures
1770  *
1771  */
1772
1773 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1774 {
1775         unsigned int x;
1776
1777         for (x = 0; x < num_gh; x++)
1778                 gfs2_glock_dq(&ghs[x]);
1779 }
1780
1781 /**
1782  * gfs2_glock_dq_uninit_m - release multiple glocks
1783  * @num_gh: the number of structures
1784  * @ghs: an array of struct gfs2_holder structures
1785  *
1786  */
1787
1788 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1789 {
1790         unsigned int x;
1791
1792         for (x = 0; x < num_gh; x++)
1793                 gfs2_glock_dq_uninit(&ghs[x]);
1794 }
1795
1796 /**
1797  * gfs2_glock_prefetch_num - prefetch a glock based on lock number
1798  * @sdp: the filesystem
1799  * @number: the lock number
1800  * @glops: the glock operations for the type of glock
1801  * @state: the state to acquire the glock in
1802  * @flags: modifier flags for the aquisition
1803  *
1804  * Returns: errno
1805  */
1806
1807 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, uint64_t number,
1808                              struct gfs2_glock_operations *glops,
1809                              unsigned int state, int flags)
1810 {
1811         struct gfs2_glock *gl;
1812         int error;
1813
1814         if (atomic_read(&sdp->sd_reclaim_count) <
1815             gfs2_tune_get(sdp, gt_reclaim_limit)) {
1816                 error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1817                 if (!error) {
1818                         gfs2_glock_prefetch(gl, state, flags);
1819                         gfs2_glock_put(gl);
1820                 }
1821         }
1822 }
1823
1824 /**
1825  * gfs2_lvb_hold - attach a LVB from a glock
1826  * @gl: The glock in question
1827  *
1828  */
1829
1830 int gfs2_lvb_hold(struct gfs2_glock *gl)
1831 {
1832         int error;
1833
1834         gfs2_glmutex_lock(gl);
1835
1836         if (!atomic_read(&gl->gl_lvb_count)) {
1837                 error = gfs2_lm_hold_lvb(gl->gl_sbd, gl->gl_lock, &gl->gl_lvb);
1838                 if (error) {
1839                         gfs2_glmutex_unlock(gl);
1840                         return error;
1841                 }
1842                 gfs2_glock_hold(gl);
1843         }
1844         atomic_inc(&gl->gl_lvb_count);
1845
1846         gfs2_glmutex_unlock(gl);
1847
1848         return 0;
1849 }
1850
1851 /**
1852  * gfs2_lvb_unhold - detach a LVB from a glock
1853  * @gl: The glock in question
1854  *
1855  */
1856
1857 void gfs2_lvb_unhold(struct gfs2_glock *gl)
1858 {
1859         gfs2_glock_hold(gl);
1860         gfs2_glmutex_lock(gl);
1861
1862         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count) > 0);
1863         if (atomic_dec_and_test(&gl->gl_lvb_count)) {
1864                 gfs2_lm_unhold_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1865                 gl->gl_lvb = NULL;
1866                 gfs2_glock_put(gl);
1867         }
1868
1869         gfs2_glmutex_unlock(gl);
1870         gfs2_glock_put(gl);
1871 }
1872
1873 void gfs2_lvb_sync(struct gfs2_glock *gl)
1874 {
1875         gfs2_glmutex_lock(gl);
1876
1877         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count));
1878         if (!gfs2_assert_warn(gl->gl_sbd, gfs2_glock_is_held_excl(gl)))
1879                 gfs2_lm_sync_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1880
1881         gfs2_glmutex_unlock(gl);
1882 }
1883
1884 static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
1885                         unsigned int state)
1886 {
1887         struct gfs2_glock *gl;
1888
1889         gl = gfs2_glock_find(sdp, name);
1890         if (!gl)
1891                 return;
1892
1893         if (gl->gl_ops->go_callback)
1894                 gl->gl_ops->go_callback(gl, state);
1895         handle_callback(gl, state);
1896
1897         spin_lock(&gl->gl_spin);
1898         run_queue(gl);
1899         spin_unlock(&gl->gl_spin);
1900
1901         gfs2_glock_put(gl);
1902 }
1903
1904 /**
1905  * gfs2_glock_cb - Callback used by locking module
1906  * @fsdata: Pointer to the superblock
1907  * @type: Type of callback
1908  * @data: Type dependent data pointer
1909  *
1910  * Called by the locking module when it wants to tell us something.
1911  * Either we need to drop a lock, one of our ASYNC requests completed, or
1912  * a journal from another client needs to be recovered.
1913  */
1914
1915 void gfs2_glock_cb(lm_fsdata_t *fsdata, unsigned int type, void *data)
1916 {
1917         struct gfs2_sbd *sdp = (struct gfs2_sbd *)fsdata;
1918
1919         atomic_inc(&sdp->sd_lm_callbacks);
1920
1921         switch (type) {
1922         case LM_CB_NEED_E:
1923                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_UNLOCKED);
1924                 return;
1925
1926         case LM_CB_NEED_D:
1927                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_DEFERRED);
1928                 return;
1929
1930         case LM_CB_NEED_S:
1931                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_SHARED);
1932                 return;
1933
1934         case LM_CB_ASYNC: {
1935                 struct lm_async_cb *async = (struct lm_async_cb *)data;
1936                 struct gfs2_glock *gl;
1937
1938                 gl = gfs2_glock_find(sdp, &async->lc_name);
1939                 if (gfs2_assert_warn(sdp, gl))
1940                         return;
1941                 if (!gfs2_assert_warn(sdp, gl->gl_req_bh))
1942                         gl->gl_req_bh(gl, async->lc_ret);
1943                 gfs2_glock_put(gl);
1944
1945                 return;
1946         }
1947
1948         case LM_CB_NEED_RECOVERY:
1949                 gfs2_jdesc_make_dirty(sdp, *(unsigned int *)data);
1950                 if (sdp->sd_recoverd_process)
1951                         wake_up_process(sdp->sd_recoverd_process);
1952                 return;
1953
1954         case LM_CB_DROPLOCKS:
1955                 gfs2_gl_hash_clear(sdp, NO_WAIT);
1956                 gfs2_quota_scan(sdp);
1957                 return;
1958
1959         default:
1960                 gfs2_assert_warn(sdp, 0);
1961                 return;
1962         }
1963 }
1964
1965 /**
1966  * gfs2_try_toss_inode - try to remove a particular inode struct from cache
1967  * sdp: the filesystem
1968  * inum: the inode number
1969  *
1970  */
1971
1972 void gfs2_try_toss_inode(struct gfs2_sbd *sdp, struct gfs2_inum *inum)
1973 {
1974         struct gfs2_glock *gl;
1975         struct gfs2_inode *ip;
1976         int error;
1977
1978         error = gfs2_glock_get(sdp, inum->no_addr, &gfs2_inode_glops,
1979                                NO_CREATE, &gl);
1980         if (error || !gl)
1981                 return;
1982
1983         if (!gfs2_glmutex_trylock(gl))
1984                 goto out;
1985
1986         ip = get_gl2ip(gl);
1987         if (!ip)
1988                 goto out_unlock;
1989
1990         if (atomic_read(&ip->i_count))
1991                 goto out_unlock;
1992
1993         gfs2_inode_destroy(ip);
1994
1995  out_unlock:
1996         gfs2_glmutex_unlock(gl);
1997
1998  out:
1999         gfs2_glock_put(gl);
2000 }
2001
2002 /**
2003  * gfs2_iopen_go_callback - Try to kick the inode/vnode associated with an
2004  *                          iopen glock from memory
2005  * @io_gl: the iopen glock
2006  * @state: the state into which the glock should be put
2007  *
2008  */
2009
2010 void gfs2_iopen_go_callback(struct gfs2_glock *io_gl, unsigned int state)
2011 {
2012         struct gfs2_glock *i_gl;
2013
2014         if (state != LM_ST_UNLOCKED)
2015                 return;
2016
2017         spin_lock(&io_gl->gl_spin);
2018         i_gl = get_gl2gl(io_gl);
2019         if (i_gl) {
2020                 gfs2_glock_hold(i_gl);
2021                 spin_unlock(&io_gl->gl_spin);
2022         } else {
2023                 spin_unlock(&io_gl->gl_spin);
2024                 return;
2025         }
2026
2027         if (gfs2_glmutex_trylock(i_gl)) {
2028                 struct gfs2_inode *ip = get_gl2ip(i_gl);
2029                 if (ip) {
2030                         gfs2_try_toss_vnode(ip);
2031                         gfs2_glmutex_unlock(i_gl);
2032                         gfs2_glock_schedule_for_reclaim(i_gl);
2033                         goto out;
2034                 }
2035                 gfs2_glmutex_unlock(i_gl);
2036         }
2037
2038  out:
2039         gfs2_glock_put(i_gl);
2040 }
2041
2042 /**
2043  * demote_ok - Check to see if it's ok to unlock a glock
2044  * @gl: the glock
2045  *
2046  * Returns: 1 if it's ok
2047  */
2048
2049 static int demote_ok(struct gfs2_glock *gl)
2050 {
2051         struct gfs2_sbd *sdp = gl->gl_sbd;
2052         struct gfs2_glock_operations *glops = gl->gl_ops;
2053         int demote = 1;
2054
2055         if (test_bit(GLF_STICKY, &gl->gl_flags))
2056                 demote = 0;
2057         else if (test_bit(GLF_PREFETCH, &gl->gl_flags))
2058                 demote = time_after_eq(jiffies,
2059                                     gl->gl_stamp +
2060                                     gfs2_tune_get(sdp, gt_prefetch_secs) * HZ);
2061         else if (glops->go_demote_ok)
2062                 demote = glops->go_demote_ok(gl);
2063
2064         return demote;
2065 }
2066
2067 /**
2068  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
2069  * @gl: the glock
2070  *
2071  */
2072
2073 void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
2074 {
2075         struct gfs2_sbd *sdp = gl->gl_sbd;
2076
2077         spin_lock(&sdp->sd_reclaim_lock);
2078         if (list_empty(&gl->gl_reclaim)) {
2079                 gfs2_glock_hold(gl);
2080                 list_add(&gl->gl_reclaim, &sdp->sd_reclaim_list);
2081                 atomic_inc(&sdp->sd_reclaim_count);
2082         }
2083         spin_unlock(&sdp->sd_reclaim_lock);
2084
2085         wake_up(&sdp->sd_reclaim_wq);
2086 }
2087
2088 /**
2089  * gfs2_reclaim_glock - process the next glock on the filesystem's reclaim list
2090  * @sdp: the filesystem
2091  *
2092  * Called from gfs2_glockd() glock reclaim daemon, or when promoting a
2093  * different glock and we notice that there are a lot of glocks in the
2094  * reclaim list.
2095  *
2096  */
2097
2098 void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
2099 {
2100         struct gfs2_glock *gl;
2101
2102         spin_lock(&sdp->sd_reclaim_lock);
2103         if (list_empty(&sdp->sd_reclaim_list)) {
2104                 spin_unlock(&sdp->sd_reclaim_lock);
2105                 return;
2106         }
2107         gl = list_entry(sdp->sd_reclaim_list.next,
2108                         struct gfs2_glock, gl_reclaim);
2109         list_del_init(&gl->gl_reclaim);
2110         spin_unlock(&sdp->sd_reclaim_lock);
2111
2112         atomic_dec(&sdp->sd_reclaim_count);
2113         atomic_inc(&sdp->sd_reclaimed);
2114
2115         if (gfs2_glmutex_trylock(gl)) {
2116                 if (gl->gl_ops == &gfs2_inode_glops) {
2117                         struct gfs2_inode *ip = get_gl2ip(gl);
2118                         if (ip && !atomic_read(&ip->i_count))
2119                                 gfs2_inode_destroy(ip);
2120                 }
2121                 if (queue_empty(gl, &gl->gl_holders) &&
2122                     gl->gl_state != LM_ST_UNLOCKED &&
2123                     demote_ok(gl))
2124                         handle_callback(gl, LM_ST_UNLOCKED);
2125                 gfs2_glmutex_unlock(gl);
2126         }
2127
2128         gfs2_glock_put(gl);
2129 }
2130
2131 /**
2132  * examine_bucket - Call a function for glock in a hash bucket
2133  * @examiner: the function
2134  * @sdp: the filesystem
2135  * @bucket: the bucket
2136  *
2137  * Returns: 1 if the bucket has entries
2138  */
2139
2140 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
2141                           struct gfs2_gl_hash_bucket *bucket)
2142 {
2143         struct glock_plug plug;
2144         struct list_head *tmp;
2145         struct gfs2_glock *gl;
2146         int entries;
2147
2148         /* Add "plug" to end of bucket list, work back up list from there */
2149         memset(&plug.gl_flags, 0, sizeof(unsigned long));
2150         set_bit(GLF_PLUG, &plug.gl_flags);
2151
2152         write_lock(&bucket->hb_lock);
2153         list_add(&plug.gl_list, &bucket->hb_list);
2154         write_unlock(&bucket->hb_lock);
2155
2156         for (;;) {
2157                 write_lock(&bucket->hb_lock);
2158
2159                 for (;;) {
2160                         tmp = plug.gl_list.next;
2161
2162                         if (tmp == &bucket->hb_list) {
2163                                 list_del(&plug.gl_list);
2164                                 entries = !list_empty(&bucket->hb_list);
2165                                 write_unlock(&bucket->hb_lock);
2166                                 return entries;
2167                         }
2168                         gl = list_entry(tmp, struct gfs2_glock, gl_list);
2169
2170                         /* Move plug up list */
2171                         list_move(&plug.gl_list, &gl->gl_list);
2172
2173                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2174                                 continue;
2175
2176                         /* examiner() must glock_put() */
2177                         gfs2_glock_hold(gl);
2178
2179                         break;
2180                 }
2181
2182                 write_unlock(&bucket->hb_lock);
2183
2184                 examiner(gl);
2185         }
2186 }
2187
2188 /**
2189  * scan_glock - look at a glock and see if we can reclaim it
2190  * @gl: the glock to look at
2191  *
2192  */
2193
2194 static void scan_glock(struct gfs2_glock *gl)
2195 {
2196         if (gfs2_glmutex_trylock(gl)) {
2197                 if (gl->gl_ops == &gfs2_inode_glops) {
2198                         struct gfs2_inode *ip = get_gl2ip(gl);
2199                         if (ip && !atomic_read(&ip->i_count))
2200                                 goto out_schedule;
2201                 }
2202                 if (queue_empty(gl, &gl->gl_holders) &&
2203                     gl->gl_state != LM_ST_UNLOCKED &&
2204                     demote_ok(gl))
2205                         goto out_schedule;
2206
2207                 gfs2_glmutex_unlock(gl);
2208         }
2209
2210         gfs2_glock_put(gl);
2211
2212         return;
2213
2214  out_schedule:
2215         gfs2_glmutex_unlock(gl);
2216         gfs2_glock_schedule_for_reclaim(gl);
2217         gfs2_glock_put(gl);
2218 }
2219
2220 /**
2221  * gfs2_scand_internal - Look for glocks and inodes to toss from memory
2222  * @sdp: the filesystem
2223  *
2224  */
2225
2226 void gfs2_scand_internal(struct gfs2_sbd *sdp)
2227 {
2228         unsigned int x;
2229
2230         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2231                 examine_bucket(scan_glock, sdp, &sdp->sd_gl_hash[x]);
2232                 cond_resched();
2233         }
2234 }
2235
2236 /**
2237  * clear_glock - look at a glock and see if we can free it from glock cache
2238  * @gl: the glock to look at
2239  *
2240  */
2241
2242 static void clear_glock(struct gfs2_glock *gl)
2243 {
2244         struct gfs2_sbd *sdp = gl->gl_sbd;
2245         int released;
2246
2247         spin_lock(&sdp->sd_reclaim_lock);
2248         if (!list_empty(&gl->gl_reclaim)) {
2249                 list_del_init(&gl->gl_reclaim);
2250                 atomic_dec(&sdp->sd_reclaim_count);
2251                 released = gfs2_glock_put(gl);
2252                 gfs2_assert(sdp, !released);
2253         }
2254         spin_unlock(&sdp->sd_reclaim_lock);
2255
2256         if (gfs2_glmutex_trylock(gl)) {
2257                 if (gl->gl_ops == &gfs2_inode_glops) {
2258                         struct gfs2_inode *ip = get_gl2ip(gl);
2259                         if (ip && !atomic_read(&ip->i_count))
2260                                 gfs2_inode_destroy(ip);
2261                 }
2262                 if (queue_empty(gl, &gl->gl_holders) &&
2263                     gl->gl_state != LM_ST_UNLOCKED)
2264                         handle_callback(gl, LM_ST_UNLOCKED);
2265
2266                 gfs2_glmutex_unlock(gl);
2267         }
2268
2269         gfs2_glock_put(gl);
2270 }
2271
2272 /**
2273  * gfs2_gl_hash_clear - Empty out the glock hash table
2274  * @sdp: the filesystem
2275  * @wait: wait until it's all gone
2276  *
2277  * Called when unmounting the filesystem, or when inter-node lock manager
2278  * requests DROPLOCKS because it is running out of capacity.
2279  */
2280
2281 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
2282 {
2283         unsigned long t;
2284         unsigned int x;
2285         int cont;
2286
2287         t = jiffies;
2288
2289         for (;;) {
2290                 cont = 0;
2291
2292                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
2293                         if (examine_bucket(clear_glock, sdp,
2294                                            &sdp->sd_gl_hash[x]))
2295                                 cont = 1;
2296
2297                 if (!wait || !cont)
2298                         break;
2299
2300                 if (time_after_eq(jiffies,
2301                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
2302                         fs_warn(sdp, "Unmount seems to be stalled. "
2303                                      "Dumping lock state...\n");
2304                         gfs2_dump_lockstate(sdp);
2305                         t = jiffies;
2306                 }
2307
2308                 /* invalidate_inodes() requires that the sb inodes list
2309                    not change, but an async completion callback for an
2310                    unlock can occur which does glock_put() which
2311                    can call iput() which will change the sb inodes list.
2312                    invalidate_inodes_mutex prevents glock_put()'s during
2313                    an invalidate_inodes() */
2314
2315                 mutex_lock(&sdp->sd_invalidate_inodes_mutex);
2316                 invalidate_inodes(sdp->sd_vfs);
2317                 mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
2318                 yield();
2319         }
2320 }
2321
2322 /*
2323  *  Diagnostic routines to help debug distributed deadlock
2324  */
2325
2326 /**
2327  * dump_holder - print information about a glock holder
2328  * @str: a string naming the type of holder
2329  * @gh: the glock holder
2330  *
2331  * Returns: 0 on success, -ENOBUFS when we run out of space
2332  */
2333
2334 static int dump_holder(char *str, struct gfs2_holder *gh)
2335 {
2336         unsigned int x;
2337         int error = -ENOBUFS;
2338
2339         printk("  %s\n", str);
2340         printk("    owner = %ld\n",
2341                    (gh->gh_owner) ? (long)gh->gh_owner->pid : -1);
2342         printk("    gh_state = %u\n", gh->gh_state);
2343         printk("    gh_flags =");
2344         for (x = 0; x < 32; x++)
2345                 if (gh->gh_flags & (1 << x))
2346                         printk(" %u", x);
2347         printk(" \n");
2348         printk("    error = %d\n", gh->gh_error);
2349         printk("    gh_iflags =");
2350         for (x = 0; x < 32; x++)
2351                 if (test_bit(x, &gh->gh_iflags))
2352                         printk(" %u", x);
2353         printk(" \n");
2354
2355         error = 0;
2356
2357         return error;
2358 }
2359
2360 /**
2361  * dump_inode - print information about an inode
2362  * @ip: the inode
2363  *
2364  * Returns: 0 on success, -ENOBUFS when we run out of space
2365  */
2366
2367 static int dump_inode(struct gfs2_inode *ip)
2368 {
2369         unsigned int x;
2370         int error = -ENOBUFS;
2371
2372         printk("  Inode:\n");
2373         printk("    num = %llu %llu\n",
2374                     ip->i_num.no_formal_ino, ip->i_num.no_addr);
2375         printk("    type = %u\n", IF2DT(ip->i_di.di_mode));
2376         printk("    i_count = %d\n", atomic_read(&ip->i_count));
2377         printk("    i_flags =");
2378         for (x = 0; x < 32; x++)
2379                 if (test_bit(x, &ip->i_flags))
2380                         printk(" %u", x);
2381         printk(" \n");
2382         printk("    vnode = %s\n", (ip->i_vnode) ? "yes" : "no");
2383
2384         error = 0;
2385
2386         return error;
2387 }
2388
2389 /**
2390  * dump_glock - print information about a glock
2391  * @gl: the glock
2392  * @count: where we are in the buffer
2393  *
2394  * Returns: 0 on success, -ENOBUFS when we run out of space
2395  */
2396
2397 static int dump_glock(struct gfs2_glock *gl)
2398 {
2399         struct gfs2_holder *gh;
2400         unsigned int x;
2401         int error = -ENOBUFS;
2402
2403         spin_lock(&gl->gl_spin);
2404
2405         printk("Glock (%u, %llu)\n",
2406                     gl->gl_name.ln_type,
2407                     gl->gl_name.ln_number);
2408         printk("  gl_flags =");
2409         for (x = 0; x < 32; x++)
2410                 if (test_bit(x, &gl->gl_flags))
2411                         printk(" %u", x);
2412         printk(" \n");
2413         printk("  gl_ref = %d\n", atomic_read(&gl->gl_ref.refcount));
2414         printk("  gl_state = %u\n", gl->gl_state);
2415         printk("  req_gh = %s\n", (gl->gl_req_gh) ? "yes" : "no");
2416         printk("  req_bh = %s\n", (gl->gl_req_bh) ? "yes" : "no");
2417         printk("  lvb_count = %d\n", atomic_read(&gl->gl_lvb_count));
2418         printk("  object = %s\n", (gl->gl_object) ? "yes" : "no");
2419         printk("  le = %s\n",
2420                    (list_empty(&gl->gl_le.le_list)) ? "no" : "yes");
2421         printk("  reclaim = %s\n",
2422                     (list_empty(&gl->gl_reclaim)) ? "no" : "yes");
2423         if (gl->gl_aspace)
2424                 printk("  aspace = %lu\n",
2425                             gl->gl_aspace->i_mapping->nrpages);
2426         else
2427                 printk("  aspace = no\n");
2428         printk("  ail = %d\n", atomic_read(&gl->gl_ail_count));
2429         if (gl->gl_req_gh) {
2430                 error = dump_holder("Request", gl->gl_req_gh);
2431                 if (error)
2432                         goto out;
2433         }
2434         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
2435                 error = dump_holder("Holder", gh);
2436                 if (error)
2437                         goto out;
2438         }
2439         list_for_each_entry(gh, &gl->gl_waiters1, gh_list) {
2440                 error = dump_holder("Waiter1", gh);
2441                 if (error)
2442                         goto out;
2443         }
2444         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
2445                 error = dump_holder("Waiter2", gh);
2446                 if (error)
2447                         goto out;
2448         }
2449         list_for_each_entry(gh, &gl->gl_waiters3, gh_list) {
2450                 error = dump_holder("Waiter3", gh);
2451                 if (error)
2452                         goto out;
2453         }
2454         if (gl->gl_ops == &gfs2_inode_glops && get_gl2ip(gl)) {
2455                 if (!test_bit(GLF_LOCK, &gl->gl_flags) &&
2456                     list_empty(&gl->gl_holders)) {
2457                         error = dump_inode(get_gl2ip(gl));
2458                         if (error)
2459                                 goto out;
2460                 } else {
2461                         error = -ENOBUFS;
2462                         printk("  Inode: busy\n");
2463                 }
2464         }
2465
2466         error = 0;
2467
2468  out:
2469         spin_unlock(&gl->gl_spin);
2470
2471         return error;
2472 }
2473
2474 /**
2475  * gfs2_dump_lockstate - print out the current lockstate
2476  * @sdp: the filesystem
2477  * @ub: the buffer to copy the information into
2478  *
2479  * If @ub is NULL, dump the lockstate to the console.
2480  *
2481  */
2482
2483 int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
2484 {
2485         struct gfs2_gl_hash_bucket *bucket;
2486         struct gfs2_glock *gl;
2487         unsigned int x;
2488         int error = 0;
2489
2490         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2491                 bucket = &sdp->sd_gl_hash[x];
2492
2493                 read_lock(&bucket->hb_lock);
2494
2495                 list_for_each_entry(gl, &bucket->hb_list, gl_list) {
2496                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2497                                 continue;
2498
2499                         error = dump_glock(gl);
2500                         if (error)
2501                                 break;
2502                 }
2503
2504                 read_unlock(&bucket->hb_lock);
2505
2506                 if (error)
2507                         break;
2508         }
2509
2510
2511         return error;
2512 }
2513