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