SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 <linux/kallsyms.h>
  20 #include <linux/gfs2_ondisk.h>
  21 #include <asm/semaphore.h>
  22 #include <asm/uaccess.h>
  23
  24 #include "gfs2.h"
  25 #include "lm_interface.h"
  26 #include "incore.h"
  27 #include "glock.h"
  28 #include "glops.h"
  29 #include "inode.h"
  30 #include "lm.h"
  31 #include "lops.h"
  32 #include "meta_io.h"
  33 #include "quota.h"
  34 #include "super.h"
  35 #include "util.h"
  36
  37 /*  Must be kept in sync with the beginning of struct gfs2_glock  */
  38 struct glock_plug {
  39         struct list_head gl_list;
  40         unsigned long gl_flags;
  41 };
  42
  43 struct greedy {
  44         struct gfs2_holder gr_gh;
  45         struct work_struct gr_work;
  46 };
  47
  48 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  49
  50 /**
  51  * relaxed_state_ok - is a requested lock compatible with the current lock mode?
  52  * @actual: the current state of the lock
  53  * @requested: the lock state that was requested by the caller
  54  * @flags: the modifier flags passed in by the caller
  55  *
  56  * Returns: 1 if the locks are compatible, 0 otherwise
  57  */
  58
  59 static inline int relaxed_state_ok(unsigned int actual, unsigned requested,
  60                                    int flags)
  61 {
  62         if (actual == requested)
  63                 return 1;
  64
  65         if (flags & GL_EXACT)
  66                 return 0;
  67
  68         if (actual == LM_ST_EXCLUSIVE && requested == LM_ST_SHARED)
  69                 return 1;
  70
  71         if (actual != LM_ST_UNLOCKED && (flags & LM_FLAG_ANY))
  72                 return 1;
  73
  74         return 0;
  75 }
  76
  77 /**
  78  * gl_hash() - Turn glock number into hash bucket number
  79  * @lock: The glock number
  80  *
  81  * Returns: The number of the corresponding hash bucket
  82  */
  83
  84 static unsigned int gl_hash(struct lm_lockname *name)
  85 {
  86         unsigned int h;
  87
  88         h = jhash(&name->ln_number, sizeof(uint64_t), 0);
  89         h = jhash(&name->ln_type, sizeof(unsigned int), h);
  90         h &= GFS2_GL_HASH_MASK;
  91
  92         return h;
  93 }
  94
  95 /**
  96  * glock_free() - Perform a few checks and then release struct gfs2_glock
  97  * @gl: The glock to release
  98  *
  99  * Also calls lock module to release its internal structure for this glock.
 100  *
 101  */
 102
 103 static void glock_free(struct gfs2_glock *gl)
 104 {
 105         struct gfs2_sbd *sdp = gl->gl_sbd;
 106         struct inode *aspace = gl->gl_aspace;
 107
 108         gfs2_lm_put_lock(sdp, gl->gl_lock);
 109
 110         if (aspace)
 111                 gfs2_aspace_put(aspace);
 112
 113         kmem_cache_free(gfs2_glock_cachep, gl);
 114 }
 115
 116 /**
 117  * gfs2_glock_hold() - increment reference count on glock
 118  * @gl: The glock to hold
 119  *
 120  */
 121
 122 void gfs2_glock_hold(struct gfs2_glock *gl)
 123 {
 124         kref_get(&gl->gl_ref);
 125 }
 126
 127 /* All work is done after the return from kref_put() so we
 128    can release the write_lock before the free. */
 129
 130 static void kill_glock(struct kref *kref)
 131 {
 132         struct gfs2_glock *gl = container_of(kref, struct gfs2_glock, gl_ref);
 133         struct gfs2_sbd *sdp = gl->gl_sbd;
 134
 135         gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
 136         gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
 137         gfs2_assert(sdp, list_empty(&gl->gl_holders));
 138         gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
 139         gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
 140         gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
 141 }
 142
 143 /**
 144  * gfs2_glock_put() - Decrement reference count on glock
 145  * @gl: The glock to put
 146  *
 147  */
 148
 149 int gfs2_glock_put(struct gfs2_glock *gl)
 150 {
 151         struct gfs2_sbd *sdp = gl->gl_sbd;
 152         struct gfs2_gl_hash_bucket *bucket = gl->gl_bucket;
 153         int rv = 0;
 154
 155         mutex_lock(&sdp->sd_invalidate_inodes_mutex);
 156
 157         write_lock(&bucket->hb_lock);
 158         if (kref_put(&gl->gl_ref, kill_glock)) {
 159                 list_del_init(&gl->gl_list);
 160                 write_unlock(&bucket->hb_lock);
 161                 glock_free(gl);
 162                 rv = 1;
 163                 goto out;
 164         }
 165         write_unlock(&bucket->hb_lock);
 166  out:
 167         mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
 168         return rv;
 169 }
 170
 171 /**
 172  * queue_empty - check to see if a glock's queue is empty
 173  * @gl: the glock
 174  * @head: the head of the queue to check
 175  *
 176  * This function protects the list in the event that a process already
 177  * has a holder on the list and is adding a second holder for itself.
 178  * The glmutex lock is what generally prevents processes from working
 179  * on the same glock at once, but the special case of adding a second
 180  * holder for yourself ("recursive" locking) doesn't involve locking
 181  * glmutex, making the spin lock necessary.
 182  *
 183  * Returns: 1 if the queue is empty
 184  */
 185
 186 static inline int queue_empty(struct gfs2_glock *gl, struct list_head *head)
 187 {
 188         int empty;
 189         spin_lock(&gl->gl_spin);
 190         empty = list_empty(head);
 191         spin_unlock(&gl->gl_spin);
 192         return empty;
 193 }
 194
 195 /**
 196  * search_bucket() - Find struct gfs2_glock by lock number
 197  * @bucket: the bucket to search
 198  * @name: The lock name
 199  *
 200  * Returns: NULL, or the struct gfs2_glock with the requested number
 201  */
 202
 203 static struct gfs2_glock *search_bucket(struct gfs2_gl_hash_bucket *bucket,
 204                                         struct lm_lockname *name)
 205 {
 206         struct gfs2_glock *gl;
 207
 208         list_for_each_entry(gl, &bucket->hb_list, gl_list) {
 209                 if (test_bit(GLF_PLUG, &gl->gl_flags))
 210                         continue;
 211                 if (!lm_name_equal(&gl->gl_name, name))
 212                         continue;
 213
 214                 kref_get(&gl->gl_ref);
 215
 216                 return gl;
 217         }
 218
 219         return NULL;
 220 }
 221
 222 /**
 223  * gfs2_glock_find() - Find glock by lock number
 224  * @sdp: The GFS2 superblock
 225  * @name: The lock name
 226  *
 227  * Returns: NULL, or the struct gfs2_glock with the requested number
 228  */
 229
 230 struct gfs2_glock *gfs2_glock_find(struct gfs2_sbd *sdp,
 231                                    struct lm_lockname *name)
 232 {
 233         struct gfs2_gl_hash_bucket *bucket = &sdp->sd_gl_hash[gl_hash(name)];
 234         struct gfs2_glock *gl;
 235
 236         read_lock(&bucket->hb_lock);
 237         gl = search_bucket(bucket, name);
 238         read_unlock(&bucket->hb_lock);
 239
 240         return gl;
 241 }
 242
 243 /**
 244  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 245  * @sdp: The GFS2 superblock
 246  * @number: the lock number
 247  * @glops: The glock_operations to use
 248  * @create: If 0, don't create the glock if it doesn't exist
 249  * @glp: the glock is returned here
 250  *
 251  * This does not lock a glock, just finds/creates structures for one.
 252  *
 253  * Returns: errno
 254  */
 255
 256 int gfs2_glock_get(struct gfs2_sbd *sdp, uint64_t number,
 257                    struct gfs2_glock_operations *glops, int create,
 258                    struct gfs2_glock **glp)
 259 {
 260         struct lm_lockname name;
 261         struct gfs2_glock *gl, *tmp;
 262         struct gfs2_gl_hash_bucket *bucket;
 263         int error;
 264
 265         name.ln_number = number;
 266         name.ln_type = glops->go_type;
 267         bucket = &sdp->sd_gl_hash[gl_hash(&name)];
 268
 269         read_lock(&bucket->hb_lock);
 270         gl = search_bucket(bucket, &name);
 271         read_unlock(&bucket->hb_lock);
 272
 273         if (gl || !create) {
 274                 *glp = gl;
 275                 return 0;
 276         }
 277
 278         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
 279         if (!gl)
 280                 return -ENOMEM;
 281
 282         memset(gl, 0, sizeof(struct gfs2_glock));
 283
 284         INIT_LIST_HEAD(&gl->gl_list);
 285         gl->gl_name = name;
 286         kref_init(&gl->gl_ref);
 287
 288         spin_lock_init(&gl->gl_spin);
 289
 290         gl->gl_state = LM_ST_UNLOCKED;
 291         INIT_LIST_HEAD(&gl->gl_holders);
 292         INIT_LIST_HEAD(&gl->gl_waiters1);
 293         INIT_LIST_HEAD(&gl->gl_waiters2);
 294         INIT_LIST_HEAD(&gl->gl_waiters3);
 295
 296         gl->gl_ops = glops;
 297
 298         gl->gl_bucket = bucket;
 299         INIT_LIST_HEAD(&gl->gl_reclaim);
 300
 301         gl->gl_sbd = sdp;
 302
 303         lops_init_le(&gl->gl_le, &gfs2_glock_lops);
 304         INIT_LIST_HEAD(&gl->gl_ail_list);
 305
 306         /* If this glock protects actual on-disk data or metadata blocks,
 307            create a VFS inode to manage the pages/buffers holding them. */
 308         if (glops == &gfs2_inode_glops ||
 309             glops == &gfs2_rgrp_glops ||
 310             glops == &gfs2_meta_glops) {
 311                 gl->gl_aspace = gfs2_aspace_get(sdp);
 312                 if (!gl->gl_aspace) {
 313                         error = -ENOMEM;
 314                         goto fail;
 315                 }
 316         }
 317
 318         error = gfs2_lm_get_lock(sdp, &name, &gl->gl_lock);
 319         if (error)
 320                 goto fail_aspace;
 321
 322         write_lock(&bucket->hb_lock);
 323         tmp = search_bucket(bucket, &name);
 324         if (tmp) {
 325                 write_unlock(&bucket->hb_lock);
 326                 glock_free(gl);
 327                 gl = tmp;
 328         } else {
 329                 list_add_tail(&gl->gl_list, &bucket->hb_list);
 330                 write_unlock(&bucket->hb_lock);
 331         }
 332
 333         *glp = gl;
 334
 335         return 0;
 336
 337  fail_aspace:
 338         if (gl->gl_aspace)
 339                 gfs2_aspace_put(gl->gl_aspace);
 340
 341  fail:
 342         kmem_cache_free(gfs2_glock_cachep, gl);
 343
 344         return error;
 345 }
 346
 347 /**
 348  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 349  * @gl: the glock
 350  * @state: the state we're requesting
 351  * @flags: the modifier flags
 352  * @gh: the holder structure
 353  *
 354  */
 355
 356 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, int flags,
 357                       struct gfs2_holder *gh)
 358 {
 359         INIT_LIST_HEAD(&gh->gh_list);
 360         gh->gh_gl = gl;
 361         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 362         gh->gh_owner = (flags & GL_NEVER_RECURSE) ? NULL : current;
 363         gh->gh_state = state;
 364         gh->gh_flags = flags;
 365         gh->gh_error = 0;
 366         gh->gh_iflags = 0;
 367         init_completion(&gh->gh_wait);
 368
 369         if (gh->gh_state == LM_ST_EXCLUSIVE)
 370                 gh->gh_flags |= GL_LOCAL_EXCL;
 371
 372         gfs2_glock_hold(gl);
 373 }
 374
 375 /**
 376  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 377  * @state: the state we're requesting
 378  * @flags: the modifier flags
 379  * @gh: the holder structure
 380  *
 381  * Don't mess with the glock.
 382  *
 383  */
 384
 385 void gfs2_holder_reinit(unsigned int state, int flags, struct gfs2_holder *gh)
 386 {
 387         gh->gh_state = state;
 388         gh->gh_flags = flags;
 389         if (gh->gh_state == LM_ST_EXCLUSIVE)
 390                 gh->gh_flags |= GL_LOCAL_EXCL;
 391
 392         gh->gh_iflags &= 1 << HIF_ALLOCED;
 393         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 394 }
 395
 396 /**
 397  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 398  * @gh: the holder structure
 399  *
 400  */
 401
 402 void gfs2_holder_uninit(struct gfs2_holder *gh)
 403 {
 404         gfs2_glock_put(gh->gh_gl);
 405         gh->gh_gl = NULL;
 406         gh->gh_ip = 0;
 407 }
 408
 409 /**
 410  * gfs2_holder_get - get a struct gfs2_holder structure
 411  * @gl: the glock
 412  * @state: the state we're requesting
 413  * @flags: the modifier flags
 414  * @gfp_flags: __GFP_NOFAIL
 415  *
 416  * Figure out how big an impact this function has.  Either:
 417  * 1) Replace it with a cache of structures hanging off the struct gfs2_sbd
 418  * 2) Leave it like it is
 419  *
 420  * Returns: the holder structure, NULL on ENOMEM
 421  */
 422
 423 struct gfs2_holder *gfs2_holder_get(struct gfs2_glock *gl, unsigned int state,
 424                                     int flags, gfp_t gfp_flags)
 425 {
 426         struct gfs2_holder *gh;
 427
 428         gh = kmalloc(sizeof(struct gfs2_holder), gfp_flags);
 429         if (!gh)
 430                 return NULL;
 431
 432         gfs2_holder_init(gl, state, flags, gh);
 433         set_bit(HIF_ALLOCED, &gh->gh_iflags);
 434         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 435         return gh;
 436 }
 437
 438 /**
 439  * gfs2_holder_put - get rid of a struct gfs2_holder structure
 440  * @gh: the holder structure
 441  *
 442  */
 443
 444 void gfs2_holder_put(struct gfs2_holder *gh)
 445 {
 446         gfs2_holder_uninit(gh);
 447         kfree(gh);
 448 }
 449
 450 /**
 451  * handle_recurse - put other holder structures (marked recursive)
 452  *                  into the holders list
 453  * @gh: the holder structure
 454  *
 455  */
 456
 457 static void handle_recurse(struct gfs2_holder *gh)
 458 {
 459         struct gfs2_glock *gl = gh->gh_gl;
 460         struct gfs2_sbd *sdp = gl->gl_sbd;
 461         struct gfs2_holder *tmp_gh, *safe;
 462         int found = 0;
 463
 464         printk(KERN_INFO "recursion %016llx, %u\n", gl->gl_name.ln_number,
 465                 gl->gl_name.ln_type);
 466
 467         if (gfs2_assert_warn(sdp, gh->gh_owner))
 468                 return;
 469
 470         list_for_each_entry_safe(tmp_gh, safe, &gl->gl_waiters3, gh_list) {
 471                 if (tmp_gh->gh_owner != gh->gh_owner)
 472                         continue;
 473
 474                 gfs2_assert_warn(sdp,
 475                                  test_bit(HIF_RECURSE, &tmp_gh->gh_iflags));
 476
 477                 list_move_tail(&tmp_gh->gh_list, &gl->gl_holders);
 478                 tmp_gh->gh_error = 0;
 479                 set_bit(HIF_HOLDER, &tmp_gh->gh_iflags);
 480
 481                 complete(&tmp_gh->gh_wait);
 482
 483                 found = 1;
 484         }
 485
 486         gfs2_assert_warn(sdp, found);
 487 }
 488
 489 /**
 490  * do_unrecurse - a recursive holder was just dropped of the waiters3 list
 491  * @gh: the holder
 492  *
 493  * If there is only one other recursive holder, clear its HIF_RECURSE bit.
 494  * If there is more than one, leave them alone.
 495  *
 496  */
 497
 498 static void do_unrecurse(struct gfs2_holder *gh)
 499 {
 500         struct gfs2_glock *gl = gh->gh_gl;
 501         struct gfs2_sbd *sdp = gl->gl_sbd;
 502         struct gfs2_holder *tmp_gh, *last_gh = NULL;
 503         int found = 0;
 504
 505         if (gfs2_assert_warn(sdp, gh->gh_owner))
 506                 return;
 507
 508         list_for_each_entry(tmp_gh, &gl->gl_waiters3, gh_list) {
 509                 if (tmp_gh->gh_owner != gh->gh_owner)
 510                         continue;
 511
 512                 gfs2_assert_warn(sdp,
 513                                  test_bit(HIF_RECURSE, &tmp_gh->gh_iflags));
 514
 515                 if (found)
 516                         return;
 517
 518                 found = 1;
 519                 last_gh = tmp_gh;
 520         }
 521
 522         if (!gfs2_assert_warn(sdp, found))
 523                 clear_bit(HIF_RECURSE, &last_gh->gh_iflags);
 524 }
 525
 526 /**
 527  * rq_mutex - process a mutex request in the queue
 528  * @gh: the glock holder
 529  *
 530  * Returns: 1 if the queue is blocked
 531  */
 532
 533 static int rq_mutex(struct gfs2_holder *gh)
 534 {
 535         struct gfs2_glock *gl = gh->gh_gl;
 536
 537         list_del_init(&gh->gh_list);
 538         /*  gh->gh_error never examined.  */
 539         set_bit(GLF_LOCK, &gl->gl_flags);
 540         complete(&gh->gh_wait);
 541
 542         return 1;
 543 }
 544
 545 /**
 546  * rq_promote - process a promote request in the queue
 547  * @gh: the glock holder
 548  *
 549  * Acquire a new inter-node lock, or change a lock state to more restrictive.
 550  *
 551  * Returns: 1 if the queue is blocked
 552  */
 553
 554 static int rq_promote(struct gfs2_holder *gh)
 555 {
 556         struct gfs2_glock *gl = gh->gh_gl;
 557         struct gfs2_sbd *sdp = gl->gl_sbd;
 558         struct gfs2_glock_operations *glops = gl->gl_ops;
 559         int recurse;
 560
 561         if (!relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 562                 if (list_empty(&gl->gl_holders)) {
 563                         gl->gl_req_gh = gh;
 564                         set_bit(GLF_LOCK, &gl->gl_flags);
 565                         spin_unlock(&gl->gl_spin);
 566
 567                         if (atomic_read(&sdp->sd_reclaim_count) >
 568                             gfs2_tune_get(sdp, gt_reclaim_limit) &&
 569                             !(gh->gh_flags & LM_FLAG_PRIORITY)) {
 570                                 gfs2_reclaim_glock(sdp);
 571                                 gfs2_reclaim_glock(sdp);
 572                         }
 573
 574                         glops->go_xmote_th(gl, gh->gh_state,
 575                                            gh->gh_flags);
 576
 577                         spin_lock(&gl->gl_spin);
 578                 }
 579                 return 1;
 580         }
 581
 582         if (list_empty(&gl->gl_holders)) {
 583                 set_bit(HIF_FIRST, &gh->gh_iflags);
 584                 set_bit(GLF_LOCK, &gl->gl_flags);
 585                 recurse = 0;
 586         } else {
 587                 struct gfs2_holder *next_gh;
 588                 if (gh->gh_flags & GL_LOCAL_EXCL)
 589                         return 1;
 590                 next_gh = list_entry(gl->gl_holders.next, struct gfs2_holder,
 591                                      gh_list);
 592                 if (next_gh->gh_flags & GL_LOCAL_EXCL)
 593                          return 1;
 594                 recurse = test_bit(HIF_RECURSE, &gh->gh_iflags);
 595         }
 596
 597         list_move_tail(&gh->gh_list, &gl->gl_holders);
 598         gh->gh_error = 0;
 599         set_bit(HIF_HOLDER, &gh->gh_iflags);
 600
 601         if (recurse)
 602                 handle_recurse(gh);
 603
 604         complete(&gh->gh_wait);
 605
 606         return 0;
 607 }
 608
 609 /**
 610  * rq_demote - process a demote request in the queue
 611  * @gh: the glock holder
 612  *
 613  * Returns: 1 if the queue is blocked
 614  */
 615
 616 static int rq_demote(struct gfs2_holder *gh)
 617 {
 618         struct gfs2_glock *gl = gh->gh_gl;
 619         struct gfs2_glock_operations *glops = gl->gl_ops;
 620
 621         if (!list_empty(&gl->gl_holders))
 622                 return 1;
 623
 624         if (gl->gl_state == gh->gh_state || gl->gl_state == LM_ST_UNLOCKED) {
 625                 list_del_init(&gh->gh_list);
 626                 gh->gh_error = 0;
 627                 spin_unlock(&gl->gl_spin);
 628                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 629                         gfs2_holder_put(gh);
 630                 else
 631                         complete(&gh->gh_wait);
 632                 spin_lock(&gl->gl_spin);
 633         } else {
 634                 gl->gl_req_gh = gh;
 635                 set_bit(GLF_LOCK, &gl->gl_flags);
 636                 spin_unlock(&gl->gl_spin);
 637
 638                 if (gh->gh_state == LM_ST_UNLOCKED ||
 639                     gl->gl_state != LM_ST_EXCLUSIVE)
 640                         glops->go_drop_th(gl);
 641                 else
 642                         glops->go_xmote_th(gl, gh->gh_state, gh->gh_flags);
 643
 644                 spin_lock(&gl->gl_spin);
 645         }
 646
 647         return 0;
 648 }
 649
 650 /**
 651  * rq_greedy - process a queued request to drop greedy status
 652  * @gh: the glock holder
 653  *
 654  * Returns: 1 if the queue is blocked
 655  */
 656
 657 static int rq_greedy(struct gfs2_holder *gh)
 658 {
 659         struct gfs2_glock *gl = gh->gh_gl;
 660
 661         list_del_init(&gh->gh_list);
 662         /*  gh->gh_error never examined.  */
 663         clear_bit(GLF_GREEDY, &gl->gl_flags);
 664         spin_unlock(&gl->gl_spin);
 665
 666         gfs2_holder_uninit(gh);
 667         kfree(container_of(gh, struct greedy, gr_gh));
 668
 669         spin_lock(&gl->gl_spin);
 670
 671         return 0;
 672 }
 673
 674 /**
 675  * run_queue - process holder structures on a glock
 676  * @gl: the glock
 677  *
 678  */
 679
 680 static void run_queue(struct gfs2_glock *gl)
 681 {
 682         struct gfs2_holder *gh;
 683         int blocked = 1;
 684
 685         for (;;) {
 686                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 687                         break;
 688
 689                 if (!list_empty(&gl->gl_waiters1)) {
 690                         gh = list_entry(gl->gl_waiters1.next,
 691                                         struct gfs2_holder, gh_list);
 692
 693                         if (test_bit(HIF_MUTEX, &gh->gh_iflags))
 694                                 blocked = rq_mutex(gh);
 695                         else
 696                                 gfs2_assert_warn(gl->gl_sbd, 0);
 697
 698                 } else if (!list_empty(&gl->gl_waiters2) &&
 699                            !test_bit(GLF_SKIP_WAITERS2, &gl->gl_flags)) {
 700                         gh = list_entry(gl->gl_waiters2.next,
 701                                         struct gfs2_holder, gh_list);
 702
 703                         if (test_bit(HIF_DEMOTE, &gh->gh_iflags))
 704                                 blocked = rq_demote(gh);
 705                         else if (test_bit(HIF_GREEDY, &gh->gh_iflags))
 706                                 blocked = rq_greedy(gh);
 707                         else
 708                                 gfs2_assert_warn(gl->gl_sbd, 0);
 709
 710                 } else if (!list_empty(&gl->gl_waiters3)) {
 711                         gh = list_entry(gl->gl_waiters3.next,
 712                                         struct gfs2_holder, gh_list);
 713
 714                         if (test_bit(HIF_PROMOTE, &gh->gh_iflags))
 715                                 blocked = rq_promote(gh);
 716                         else
 717                                 gfs2_assert_warn(gl->gl_sbd, 0);
 718
 719                 } else
 720                         break;
 721
 722                 if (blocked)
 723                         break;
 724         }
 725 }
 726
 727 /**
 728  * gfs2_glmutex_lock - acquire a local lock on a glock
 729  * @gl: the glock
 730  *
 731  * Gives caller exclusive access to manipulate a glock structure.
 732  */
 733
 734 void gfs2_glmutex_lock(struct gfs2_glock *gl)
 735 {
 736         struct gfs2_holder gh;
 737
 738         gfs2_holder_init(gl, 0, 0, &gh);
 739         set_bit(HIF_MUTEX, &gh.gh_iflags);
 740
 741         spin_lock(&gl->gl_spin);
 742         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 743                 list_add_tail(&gh.gh_list, &gl->gl_waiters1);
 744         else
 745                 complete(&gh.gh_wait);
 746         spin_unlock(&gl->gl_spin);
 747
 748         wait_for_completion(&gh.gh_wait);
 749         gfs2_holder_uninit(&gh);
 750 }
 751
 752 /**
 753  * gfs2_glmutex_trylock - try to acquire a local lock on a glock
 754  * @gl: the glock
 755  *
 756  * Returns: 1 if the glock is acquired
 757  */
 758
 759 int gfs2_glmutex_trylock(struct gfs2_glock *gl)
 760 {
 761         int acquired = 1;
 762
 763         spin_lock(&gl->gl_spin);
 764         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 765                 acquired = 0;
 766         spin_unlock(&gl->gl_spin);
 767
 768         return acquired;
 769 }
 770
 771 /**
 772  * gfs2_glmutex_unlock - release a local lock on a glock
 773  * @gl: the glock
 774  *
 775  */
 776
 777 void gfs2_glmutex_unlock(struct gfs2_glock *gl)
 778 {
 779         spin_lock(&gl->gl_spin);
 780         clear_bit(GLF_LOCK, &gl->gl_flags);
 781         run_queue(gl);
 782         spin_unlock(&gl->gl_spin);
 783 }
 784
 785 /**
 786  * handle_callback - add a demote request to a lock's queue
 787  * @gl: the glock
 788  * @state: the state the caller wants us to change to
 789  *
 790  */
 791
 792 static void handle_callback(struct gfs2_glock *gl, unsigned int state)
 793 {
 794         struct gfs2_holder *gh, *new_gh = NULL;
 795
 796  restart:
 797         spin_lock(&gl->gl_spin);
 798
 799         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
 800                 if (test_bit(HIF_DEMOTE, &gh->gh_iflags) &&
 801                     gl->gl_req_gh != gh) {
 802                         if (gh->gh_state != state)
 803                                 gh->gh_state = LM_ST_UNLOCKED;
 804                         goto out;
 805                 }
 806         }
 807
 808         if (new_gh) {
 809                 list_add_tail(&new_gh->gh_list, &gl->gl_waiters2);
 810                 new_gh = NULL;
 811         } else {
 812                 spin_unlock(&gl->gl_spin);
 813
 814                 new_gh = gfs2_holder_get(gl, state,
 815                                          LM_FLAG_TRY | GL_NEVER_RECURSE,
 816                                          GFP_KERNEL | __GFP_NOFAIL),
 817                 set_bit(HIF_DEMOTE, &new_gh->gh_iflags);
 818                 set_bit(HIF_DEALLOC, &new_gh->gh_iflags);
 819
 820                 goto restart;
 821         }
 822
 823  out:
 824         spin_unlock(&gl->gl_spin);
 825
 826         if (new_gh)
 827                 gfs2_holder_put(new_gh);
 828 }
 829
 830 /**
 831  * state_change - record that the glock is now in a different state
 832  * @gl: the glock
 833  * @new_state the new state
 834  *
 835  */
 836
 837 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 838 {
 839         int held1, held2;
 840
 841         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 842         held2 = (new_state != LM_ST_UNLOCKED);
 843
 844         if (held1 != held2) {
 845                 if (held2)
 846                         gfs2_glock_hold(gl);
 847                 else
 848                         gfs2_glock_put(gl);
 849         }
 850
 851         gl->gl_state = new_state;
 852 }
 853
 854 /**
 855  * xmote_bh - Called after the lock module is done acquiring a lock
 856  * @gl: The glock in question
 857  * @ret: the int returned from the lock module
 858  *
 859  */
 860
 861 static void xmote_bh(struct gfs2_glock *gl, unsigned int ret)
 862 {
 863         struct gfs2_sbd *sdp = gl->gl_sbd;
 864         struct gfs2_glock_operations *glops = gl->gl_ops;
 865         struct gfs2_holder *gh = gl->gl_req_gh;
 866         int prev_state = gl->gl_state;
 867         int op_done = 1;
 868
 869         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 870         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 871         gfs2_assert_warn(sdp, !(ret & LM_OUT_ASYNC));
 872
 873         state_change(gl, ret & LM_OUT_ST_MASK);
 874
 875         if (prev_state != LM_ST_UNLOCKED && !(ret & LM_OUT_CACHEABLE)) {
 876                 if (glops->go_inval)
 877                         glops->go_inval(gl, DIO_METADATA | DIO_DATA);
 878         } else if (gl->gl_state == LM_ST_DEFERRED) {
 879                 /* We might not want to do this here.
 880                    Look at moving to the inode glops. */
 881                 if (glops->go_inval)
 882                         glops->go_inval(gl, DIO_DATA);
 883         }
 884
 885         /*  Deal with each possible exit condition  */
 886
 887         if (!gh)
 888                 gl->gl_stamp = jiffies;
 889
 890         else if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
 891                 spin_lock(&gl->gl_spin);
 892                 list_del_init(&gh->gh_list);
 893                 gh->gh_error = -EIO;
 894                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 895                         do_unrecurse(gh);
 896                 spin_unlock(&gl->gl_spin);
 897
 898         } else if (test_bit(HIF_DEMOTE, &gh->gh_iflags)) {
 899                 spin_lock(&gl->gl_spin);
 900                 list_del_init(&gh->gh_list);
 901                 if (gl->gl_state == gh->gh_state ||
 902                     gl->gl_state == LM_ST_UNLOCKED)
 903                         gh->gh_error = 0;
 904                 else {
 905                         if (gfs2_assert_warn(sdp, gh->gh_flags &
 906                                         (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) == -1)
 907                                 fs_warn(sdp, "ret = 0x%.8X\n", ret);
 908                         gh->gh_error = GLR_TRYFAILED;
 909                 }
 910                 spin_unlock(&gl->gl_spin);
 911
 912                 if (ret & LM_OUT_CANCELED)
 913                         handle_callback(gl, LM_ST_UNLOCKED); /* Lame */
 914
 915         } else if (ret & LM_OUT_CANCELED) {
 916                 spin_lock(&gl->gl_spin);
 917                 list_del_init(&gh->gh_list);
 918                 gh->gh_error = GLR_CANCELED;
 919                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 920                         do_unrecurse(gh);
 921                 spin_unlock(&gl->gl_spin);
 922
 923         } else if (relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 924                 spin_lock(&gl->gl_spin);
 925                 list_move_tail(&gh->gh_list, &gl->gl_holders);
 926                 gh->gh_error = 0;
 927                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 928                 spin_unlock(&gl->gl_spin);
 929
 930                 set_bit(HIF_FIRST, &gh->gh_iflags);
 931
 932                 op_done = 0;
 933
 934         } else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 935                 spin_lock(&gl->gl_spin);
 936                 list_del_init(&gh->gh_list);
 937                 gh->gh_error = GLR_TRYFAILED;
 938                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 939                         do_unrecurse(gh);
 940                 spin_unlock(&gl->gl_spin);
 941
 942         } else {
 943                 if (gfs2_assert_withdraw(sdp, 0) == -1)
 944                         fs_err(sdp, "ret = 0x%.8X\n", ret);
 945         }
 946
 947         if (glops->go_xmote_bh)
 948                 glops->go_xmote_bh(gl);
 949
 950         if (op_done) {
 951                 spin_lock(&gl->gl_spin);
 952                 gl->gl_req_gh = NULL;
 953                 gl->gl_req_bh = NULL;
 954                 clear_bit(GLF_LOCK, &gl->gl_flags);
 955                 run_queue(gl);
 956                 spin_unlock(&gl->gl_spin);
 957         }
 958
 959         gfs2_glock_put(gl);
 960
 961         if (gh) {
 962                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 963                         gfs2_holder_put(gh);
 964                 else
 965                         complete(&gh->gh_wait);
 966         }
 967 }
 968
 969 /**
 970  * gfs2_glock_xmote_th - Call into the lock module to acquire or change a glock
 971  * @gl: The glock in question
 972  * @state: the requested state
 973  * @flags: modifier flags to the lock call
 974  *
 975  */
 976
 977 void gfs2_glock_xmote_th(struct gfs2_glock *gl, unsigned int state, int flags)
 978 {
 979         struct gfs2_sbd *sdp = gl->gl_sbd;
 980         struct gfs2_glock_operations *glops = gl->gl_ops;
 981         int lck_flags = flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB |
 982                                  LM_FLAG_NOEXP | LM_FLAG_ANY |
 983                                  LM_FLAG_PRIORITY);
 984         unsigned int lck_ret;
 985
 986         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 987         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 988         gfs2_assert_warn(sdp, state != LM_ST_UNLOCKED);
 989         gfs2_assert_warn(sdp, state != gl->gl_state);
 990
 991         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 992                 if (glops->go_sync)
 993                         glops->go_sync(gl,
 994                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
 995         }
 996
 997         gfs2_glock_hold(gl);
 998         gl->gl_req_bh = xmote_bh;
 999
1000         lck_ret = gfs2_lm_lock(sdp, gl->gl_lock, gl->gl_state, state,
1001                                lck_flags);
1002
1003         if (gfs2_assert_withdraw(sdp, !(lck_ret & LM_OUT_ERROR)))
1004                 return;
1005
1006         if (lck_ret & LM_OUT_ASYNC)
1007                 gfs2_assert_warn(sdp, lck_ret == LM_OUT_ASYNC);
1008         else
1009                 xmote_bh(gl, lck_ret);
1010 }
1011
1012 /**
1013  * drop_bh - Called after a lock module unlock completes
1014  * @gl: the glock
1015  * @ret: the return status
1016  *
1017  * Doesn't wake up the process waiting on the struct gfs2_holder (if any)
1018  * Doesn't drop the reference on the glock the top half took out
1019  *
1020  */
1021
1022 static void drop_bh(struct gfs2_glock *gl, unsigned int ret)
1023 {
1024         struct gfs2_sbd *sdp = gl->gl_sbd;
1025         struct gfs2_glock_operations *glops = gl->gl_ops;
1026         struct gfs2_holder *gh = gl->gl_req_gh;
1027
1028         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1029
1030         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1031         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1032         gfs2_assert_warn(sdp, !ret);
1033
1034         state_change(gl, LM_ST_UNLOCKED);
1035
1036         if (glops->go_inval)
1037                 glops->go_inval(gl, DIO_METADATA | DIO_DATA);
1038
1039         if (gh) {
1040                 spin_lock(&gl->gl_spin);
1041                 list_del_init(&gh->gh_list);
1042                 gh->gh_error = 0;
1043                 spin_unlock(&gl->gl_spin);
1044         }
1045
1046         if (glops->go_drop_bh)
1047                 glops->go_drop_bh(gl);
1048
1049         spin_lock(&gl->gl_spin);
1050         gl->gl_req_gh = NULL;
1051         gl->gl_req_bh = NULL;
1052         clear_bit(GLF_LOCK, &gl->gl_flags);
1053         run_queue(gl);
1054         spin_unlock(&gl->gl_spin);
1055
1056         gfs2_glock_put(gl);
1057
1058         if (gh) {
1059                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
1060                         gfs2_holder_put(gh);
1061                 else
1062                         complete(&gh->gh_wait);
1063         }
1064 }
1065
1066 /**
1067  * gfs2_glock_drop_th - call into the lock module to unlock a lock
1068  * @gl: the glock
1069  *
1070  */
1071
1072 void gfs2_glock_drop_th(struct gfs2_glock *gl)
1073 {
1074         struct gfs2_sbd *sdp = gl->gl_sbd;
1075         struct gfs2_glock_operations *glops = gl->gl_ops;
1076         unsigned int ret;
1077
1078         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1079         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1080         gfs2_assert_warn(sdp, gl->gl_state != LM_ST_UNLOCKED);
1081
1082         if (gl->gl_state == LM_ST_EXCLUSIVE) {
1083                 if (glops->go_sync)
1084                         glops->go_sync(gl,
1085                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
1086         }
1087
1088         gfs2_glock_hold(gl);
1089         gl->gl_req_bh = drop_bh;
1090
1091         ret = gfs2_lm_unlock(sdp, gl->gl_lock, gl->gl_state);
1092
1093         if (gfs2_assert_withdraw(sdp, !(ret & LM_OUT_ERROR)))
1094                 return;
1095
1096         if (!ret)
1097                 drop_bh(gl, ret);
1098         else
1099                 gfs2_assert_warn(sdp, ret == LM_OUT_ASYNC);
1100 }
1101
1102 /**
1103  * do_cancels - cancel requests for locks stuck waiting on an expire flag
1104  * @gh: the LM_FLAG_PRIORITY holder waiting to acquire the lock
1105  *
1106  * Don't cancel GL_NOCANCEL requests.
1107  */
1108
1109 static void do_cancels(struct gfs2_holder *gh)
1110 {
1111         struct gfs2_glock *gl = gh->gh_gl;
1112
1113         spin_lock(&gl->gl_spin);
1114
1115         while (gl->gl_req_gh != gh &&
1116                !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1117                !list_empty(&gh->gh_list)) {
1118                 if (gl->gl_req_bh &&
1119                     !(gl->gl_req_gh &&
1120                       (gl->gl_req_gh->gh_flags & GL_NOCANCEL))) {
1121                         spin_unlock(&gl->gl_spin);
1122                         gfs2_lm_cancel(gl->gl_sbd, gl->gl_lock);
1123                         msleep(100);
1124                         spin_lock(&gl->gl_spin);
1125                 } else {
1126                         spin_unlock(&gl->gl_spin);
1127                         msleep(100);
1128                         spin_lock(&gl->gl_spin);
1129                 }
1130         }
1131
1132         spin_unlock(&gl->gl_spin);
1133 }
1134
1135 /**
1136  * glock_wait_internal - wait on a glock acquisition
1137  * @gh: the glock holder
1138  *
1139  * Returns: 0 on success
1140  */
1141
1142 static int glock_wait_internal(struct gfs2_holder *gh)
1143 {
1144         struct gfs2_glock *gl = gh->gh_gl;
1145         struct gfs2_sbd *sdp = gl->gl_sbd;
1146         struct gfs2_glock_operations *glops = gl->gl_ops;
1147
1148         if (test_bit(HIF_ABORTED, &gh->gh_iflags))
1149                 return -EIO;
1150
1151         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1152                 spin_lock(&gl->gl_spin);
1153                 if (gl->gl_req_gh != gh &&
1154                     !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1155                     !list_empty(&gh->gh_list)) {
1156                         list_del_init(&gh->gh_list);
1157                         gh->gh_error = GLR_TRYFAILED;
1158                         if (test_bit(HIF_RECURSE, &gh->gh_iflags))
1159                                 do_unrecurse(gh);
1160                         run_queue(gl);
1161                         spin_unlock(&gl->gl_spin);
1162                         return gh->gh_error;
1163                 }
1164                 spin_unlock(&gl->gl_spin);
1165         }
1166
1167         if (gh->gh_flags & LM_FLAG_PRIORITY)
1168                 do_cancels(gh);
1169
1170         wait_for_completion(&gh->gh_wait);
1171
1172         if (gh->gh_error)
1173                 return gh->gh_error;
1174
1175         gfs2_assert_withdraw(sdp, test_bit(HIF_HOLDER, &gh->gh_iflags));
1176         gfs2_assert_withdraw(sdp, relaxed_state_ok(gl->gl_state,
1177                                                    gh->gh_state,
1178                                                    gh->gh_flags));
1179
1180         if (test_bit(HIF_FIRST, &gh->gh_iflags)) {
1181                 gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1182
1183                 if (glops->go_lock) {
1184                         gh->gh_error = glops->go_lock(gh);
1185                         if (gh->gh_error) {
1186                                 spin_lock(&gl->gl_spin);
1187                                 list_del_init(&gh->gh_list);
1188                                 if (test_and_clear_bit(HIF_RECURSE,
1189                                                        &gh->gh_iflags))
1190                                         do_unrecurse(gh);
1191                                 spin_unlock(&gl->gl_spin);
1192                         }
1193                 }
1194
1195                 spin_lock(&gl->gl_spin);
1196                 gl->gl_req_gh = NULL;
1197                 gl->gl_req_bh = NULL;
1198                 clear_bit(GLF_LOCK, &gl->gl_flags);
1199                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
1200                         handle_recurse(gh);
1201                 run_queue(gl);
1202                 spin_unlock(&gl->gl_spin);
1203         }
1204
1205         return gh->gh_error;
1206 }
1207
1208 static inline struct gfs2_holder *
1209 find_holder_by_owner(struct list_head *head, struct task_struct *owner)
1210 {
1211         struct gfs2_holder *gh;
1212
1213         list_for_each_entry(gh, head, gh_list) {
1214                 if (gh->gh_owner == owner)
1215                         return gh;
1216         }
1217
1218         return NULL;
1219 }
1220
1221 /**
1222  * recurse_check -
1223  *
1224  * Make sure the new holder is compatible with the pre-existing one.
1225  *
1226  */
1227
1228 static int recurse_check(struct gfs2_holder *existing, struct gfs2_holder *new,
1229                          unsigned int state)
1230 {
1231         struct gfs2_sbd *sdp = existing->gh_gl->gl_sbd;
1232
1233         if (gfs2_assert_warn(sdp, (new->gh_flags & LM_FLAG_ANY) ||
1234                                   !(existing->gh_flags & LM_FLAG_ANY)))
1235                 goto fail;
1236
1237         if (gfs2_assert_warn(sdp, (existing->gh_flags & GL_LOCAL_EXCL) ||
1238                                   !(new->gh_flags & GL_LOCAL_EXCL)))
1239                 goto fail;
1240
1241         if (gfs2_assert_warn(sdp, relaxed_state_ok(state, new->gh_state,
1242                                                    new->gh_flags)))
1243                 goto fail;
1244
1245         return 0;
1246
1247  fail:
1248         print_symbol(KERN_WARNING "GFS2: Existing holder from %s\n",
1249                      existing->gh_ip);
1250         print_symbol(KERN_WARNING "GFS2: New holder from %s\n", new->gh_ip);
1251         set_bit(HIF_ABORTED, &new->gh_iflags);
1252         return -EINVAL;
1253 }
1254
1255 /**
1256  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1257  * @gh: the holder structure to add
1258  *
1259  */
1260
1261 static void add_to_queue(struct gfs2_holder *gh)
1262 {
1263         struct gfs2_glock *gl = gh->gh_gl;
1264         struct gfs2_holder *existing;
1265
1266         if (!gh->gh_owner)
1267                 goto out;
1268
1269         existing = find_holder_by_owner(&gl->gl_holders, gh->gh_owner);
1270         if (existing) {
1271                 if (recurse_check(existing, gh, gl->gl_state))
1272                         return;
1273
1274                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1275                 set_bit(HIF_HOLDER, &gh->gh_iflags);
1276
1277                 gh->gh_error = 0;
1278                 complete(&gh->gh_wait);
1279
1280                 return;
1281         }
1282
1283         existing = find_holder_by_owner(&gl->gl_waiters3, gh->gh_owner);
1284         if (existing) {
1285                 if (recurse_check(existing, gh, existing->gh_state))
1286                         return;
1287
1288                 set_bit(HIF_RECURSE, &gh->gh_iflags);
1289                 set_bit(HIF_RECURSE, &existing->gh_iflags);
1290
1291                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1292
1293                 return;
1294         }
1295
1296  out:
1297         if (gh->gh_flags & LM_FLAG_PRIORITY)
1298                 list_add(&gh->gh_list, &gl->gl_waiters3);
1299         else
1300                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1301 }
1302
1303 /**
1304  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1305  * @gh: the holder structure
1306  *
1307  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1308  *
1309  * Returns: 0, GLR_TRYFAILED, or errno on failure
1310  */
1311
1312 int gfs2_glock_nq(struct gfs2_holder *gh)
1313 {
1314         struct gfs2_glock *gl = gh->gh_gl;
1315         struct gfs2_sbd *sdp = gl->gl_sbd;
1316         int error = 0;
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_glock_operations *glops = gl->gl_ops;
1407
1408         if (gh->gh_flags & GL_SYNC)
1409                 set_bit(GLF_SYNC, &gl->gl_flags);
1410
1411         if (gh->gh_flags & GL_NOCACHE)
1412                 handle_callback(gl, LM_ST_UNLOCKED);
1413
1414         gfs2_glmutex_lock(gl);
1415
1416         spin_lock(&gl->gl_spin);
1417         list_del_init(&gh->gh_list);
1418
1419         if (list_empty(&gl->gl_holders)) {
1420                 spin_unlock(&gl->gl_spin);
1421
1422                 if (glops->go_unlock)
1423                         glops->go_unlock(gh);
1424
1425                 if (test_bit(GLF_SYNC, &gl->gl_flags)) {
1426                         if (glops->go_sync)
1427                                 glops->go_sync(gl, DIO_METADATA | DIO_DATA);
1428                 }
1429
1430                 gl->gl_stamp = jiffies;
1431
1432                 spin_lock(&gl->gl_spin);
1433         }
1434
1435         clear_bit(GLF_LOCK, &gl->gl_flags);
1436         run_queue(gl);
1437         spin_unlock(&gl->gl_spin);
1438 }
1439
1440 /**
1441  * gfs2_glock_prefetch - Try to prefetch a glock
1442  * @gl: the glock
1443  * @state: the state to prefetch in
1444  * @flags: flags passed to go_xmote_th()
1445  *
1446  */
1447
1448 void gfs2_glock_prefetch(struct gfs2_glock *gl, unsigned int state, int flags)
1449 {
1450         struct gfs2_glock_operations *glops = gl->gl_ops;
1451
1452         spin_lock(&gl->gl_spin);
1453
1454         if (test_bit(GLF_LOCK, &gl->gl_flags) ||
1455             !list_empty(&gl->gl_holders) ||
1456             !list_empty(&gl->gl_waiters1) ||
1457             !list_empty(&gl->gl_waiters2) ||
1458             !list_empty(&gl->gl_waiters3) ||
1459             relaxed_state_ok(gl->gl_state, state, flags)) {
1460                 spin_unlock(&gl->gl_spin);
1461                 return;
1462         }
1463
1464         set_bit(GLF_PREFETCH, &gl->gl_flags);
1465         set_bit(GLF_LOCK, &gl->gl_flags);
1466         spin_unlock(&gl->gl_spin);
1467
1468         glops->go_xmote_th(gl, state, flags);
1469 }
1470
1471 /**
1472  * gfs2_glock_force_drop - Force a glock to be uncached
1473  * @gl: the glock
1474  *
1475  */
1476
1477 void gfs2_glock_force_drop(struct gfs2_glock *gl)
1478 {
1479         struct gfs2_holder gh;
1480
1481         gfs2_holder_init(gl, LM_ST_UNLOCKED, GL_NEVER_RECURSE, &gh);
1482         set_bit(HIF_DEMOTE, &gh.gh_iflags);
1483
1484         spin_lock(&gl->gl_spin);
1485         list_add_tail(&gh.gh_list, &gl->gl_waiters2);
1486         run_queue(gl);
1487         spin_unlock(&gl->gl_spin);
1488
1489         wait_for_completion(&gh.gh_wait);
1490         gfs2_holder_uninit(&gh);
1491 }
1492
1493 static void greedy_work(void *data)
1494 {
1495         struct greedy *gr = (struct greedy *)data;
1496         struct gfs2_holder *gh = &gr->gr_gh;
1497         struct gfs2_glock *gl = gh->gh_gl;
1498         struct gfs2_glock_operations *glops = gl->gl_ops;
1499
1500         clear_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1501
1502         if (glops->go_greedy)
1503                 glops->go_greedy(gl);
1504
1505         spin_lock(&gl->gl_spin);
1506
1507         if (list_empty(&gl->gl_waiters2)) {
1508                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1509                 spin_unlock(&gl->gl_spin);
1510                 gfs2_holder_uninit(gh);
1511                 kfree(gr);
1512         } else {
1513                 gfs2_glock_hold(gl);
1514                 list_add_tail(&gh->gh_list, &gl->gl_waiters2);
1515                 run_queue(gl);
1516                 spin_unlock(&gl->gl_spin);
1517                 gfs2_glock_put(gl);
1518         }
1519 }
1520
1521 /**
1522  * gfs2_glock_be_greedy -
1523  * @gl:
1524  * @time:
1525  *
1526  * Returns: 0 if go_greedy will be called, 1 otherwise
1527  */
1528
1529 int gfs2_glock_be_greedy(struct gfs2_glock *gl, unsigned int time)
1530 {
1531         struct greedy *gr;
1532         struct gfs2_holder *gh;
1533
1534         if (!time ||
1535             gl->gl_sbd->sd_args.ar_localcaching ||
1536             test_and_set_bit(GLF_GREEDY, &gl->gl_flags))
1537                 return 1;
1538
1539         gr = kmalloc(sizeof(struct greedy), GFP_KERNEL);
1540         if (!gr) {
1541                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1542                 return 1;
1543         }
1544         gh = &gr->gr_gh;
1545
1546         gfs2_holder_init(gl, 0, GL_NEVER_RECURSE, gh);
1547         set_bit(HIF_GREEDY, &gh->gh_iflags);
1548         INIT_WORK(&gr->gr_work, greedy_work, gr);
1549
1550         set_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1551         schedule_delayed_work(&gr->gr_work, time);
1552
1553         return 0;
1554 }
1555
1556 /**
1557  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1558  * @gh: the holder structure
1559  *
1560  */
1561
1562 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1563 {
1564         gfs2_glock_dq(gh);
1565         gfs2_holder_uninit(gh);
1566 }
1567
1568 /**
1569  * gfs2_glock_nq_num - acquire a glock based on lock number
1570  * @sdp: the filesystem
1571  * @number: the lock number
1572  * @glops: the glock operations for the type of glock
1573  * @state: the state to acquire the glock in
1574  * @flags: modifier flags for the aquisition
1575  * @gh: the struct gfs2_holder
1576  *
1577  * Returns: errno
1578  */
1579
1580 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, uint64_t number,
1581                       struct gfs2_glock_operations *glops, unsigned int state,
1582                       int flags, struct gfs2_holder *gh)
1583 {
1584         struct gfs2_glock *gl;
1585         int error;
1586
1587         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1588         if (!error) {
1589                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1590                 gfs2_glock_put(gl);
1591         }
1592
1593         return error;
1594 }
1595
1596 /**
1597  * glock_compare - Compare two struct gfs2_glock structures for sorting
1598  * @arg_a: the first structure
1599  * @arg_b: the second structure
1600  *
1601  */
1602
1603 static int glock_compare(const void *arg_a, const void *arg_b)
1604 {
1605         struct gfs2_holder *gh_a = *(struct gfs2_holder **)arg_a;
1606         struct gfs2_holder *gh_b = *(struct gfs2_holder **)arg_b;
1607         struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1608         struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1609         int ret = 0;
1610
1611         if (a->ln_number > b->ln_number)
1612                 ret = 1;
1613         else if (a->ln_number < b->ln_number)
1614                 ret = -1;
1615         else {
1616                 if (gh_a->gh_state == LM_ST_SHARED &&
1617                     gh_b->gh_state == LM_ST_EXCLUSIVE)
1618                         ret = 1;
1619                 else if (!(gh_a->gh_flags & GL_LOCAL_EXCL) &&
1620                          (gh_b->gh_flags & GL_LOCAL_EXCL))
1621                         ret = 1;
1622         }
1623
1624         return ret;
1625 }
1626
1627 /**
1628  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1629  * @num_gh: the number of structures
1630  * @ghs: an array of struct gfs2_holder structures
1631  *
1632  * Returns: 0 on success (all glocks acquired),
1633  *          errno on failure (no glocks acquired)
1634  */
1635
1636 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1637                      struct gfs2_holder **p)
1638 {
1639         unsigned int x;
1640         int error = 0;
1641
1642         for (x = 0; x < num_gh; x++)
1643                 p[x] = &ghs[x];
1644
1645         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1646
1647         for (x = 0; x < num_gh; x++) {
1648                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1649
1650                 error = gfs2_glock_nq(p[x]);
1651                 if (error) {
1652                         while (x--)
1653                                 gfs2_glock_dq(p[x]);
1654                         break;
1655                 }
1656         }
1657
1658         return error;
1659 }
1660
1661 /**
1662  * gfs2_glock_nq_m - acquire multiple glocks
1663  * @num_gh: the number of structures
1664  * @ghs: an array of struct gfs2_holder structures
1665  *
1666  * Figure out how big an impact this function has.  Either:
1667  * 1) Replace this code with code that calls gfs2_glock_prefetch()
1668  * 2) Forget async stuff and just call nq_m_sync()
1669  * 3) Leave it like it is
1670  *
1671  * Returns: 0 on success (all glocks acquired),
1672  *          errno on failure (no glocks acquired)
1673  */
1674
1675 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1676 {
1677         int *e;
1678         unsigned int x;
1679         int borked = 0, serious = 0;
1680         int error = 0;
1681
1682         if (!num_gh)
1683                 return 0;
1684
1685         if (num_gh == 1) {
1686                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1687                 return gfs2_glock_nq(ghs);
1688         }
1689
1690         e = kcalloc(num_gh, sizeof(struct gfs2_holder *), GFP_KERNEL);
1691         if (!e)
1692                 return -ENOMEM;
1693
1694         for (x = 0; x < num_gh; x++) {
1695                 ghs[x].gh_flags |= LM_FLAG_TRY | GL_ASYNC;
1696                 error = gfs2_glock_nq(&ghs[x]);
1697                 if (error) {
1698                         borked = 1;
1699                         serious = error;
1700                         num_gh = x;
1701                         break;
1702                 }
1703         }
1704
1705         for (x = 0; x < num_gh; x++) {
1706                 error = e[x] = glock_wait_internal(&ghs[x]);
1707                 if (error) {
1708                         borked = 1;
1709                         if (error != GLR_TRYFAILED && error != GLR_CANCELED)
1710                                 serious = error;
1711                 }
1712         }
1713
1714         if (!borked) {
1715                 kfree(e);
1716                 return 0;
1717         }
1718
1719         for (x = 0; x < num_gh; x++)
1720                 if (!e[x])
1721                         gfs2_glock_dq(&ghs[x]);
1722
1723         if (serious)
1724                 error = serious;
1725         else {
1726                 for (x = 0; x < num_gh; x++)
1727                         gfs2_holder_reinit(ghs[x].gh_state, ghs[x].gh_flags,
1728                                           &ghs[x]);
1729                 error = nq_m_sync(num_gh, ghs, (struct gfs2_holder **)e);
1730         }
1731
1732         kfree(e);
1733
1734         return error;
1735 }
1736
1737 /**
1738  * gfs2_glock_dq_m - release multiple glocks
1739  * @num_gh: the number of structures
1740  * @ghs: an array of struct gfs2_holder structures
1741  *
1742  */
1743
1744 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1745 {
1746         unsigned int x;
1747
1748         for (x = 0; x < num_gh; x++)
1749                 gfs2_glock_dq(&ghs[x]);
1750 }
1751
1752 /**
1753  * gfs2_glock_dq_uninit_m - release multiple glocks
1754  * @num_gh: the number of structures
1755  * @ghs: an array of struct gfs2_holder structures
1756  *
1757  */
1758
1759 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1760 {
1761         unsigned int x;
1762
1763         for (x = 0; x < num_gh; x++)
1764                 gfs2_glock_dq_uninit(&ghs[x]);
1765 }
1766
1767 /**
1768  * gfs2_glock_prefetch_num - prefetch a glock based on lock number
1769  * @sdp: the filesystem
1770  * @number: the lock number
1771  * @glops: the glock operations for the type of glock
1772  * @state: the state to acquire the glock in
1773  * @flags: modifier flags for the aquisition
1774  *
1775  * Returns: errno
1776  */
1777
1778 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, uint64_t number,
1779                              struct gfs2_glock_operations *glops,
1780                              unsigned int state, int flags)
1781 {
1782         struct gfs2_glock *gl;
1783         int error;
1784
1785         if (atomic_read(&sdp->sd_reclaim_count) <
1786             gfs2_tune_get(sdp, gt_reclaim_limit)) {
1787                 error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1788                 if (!error) {
1789                         gfs2_glock_prefetch(gl, state, flags);
1790                         gfs2_glock_put(gl);
1791                 }
1792         }
1793 }
1794
1795 /**
1796  * gfs2_lvb_hold - attach a LVB from a glock
1797  * @gl: The glock in question
1798  *
1799  */
1800
1801 int gfs2_lvb_hold(struct gfs2_glock *gl)
1802 {
1803         int error;
1804
1805         gfs2_glmutex_lock(gl);
1806
1807         if (!atomic_read(&gl->gl_lvb_count)) {
1808                 error = gfs2_lm_hold_lvb(gl->gl_sbd, gl->gl_lock, &gl->gl_lvb);
1809                 if (error) {
1810                         gfs2_glmutex_unlock(gl);
1811                         return error;
1812                 }
1813                 gfs2_glock_hold(gl);
1814         }
1815         atomic_inc(&gl->gl_lvb_count);
1816
1817         gfs2_glmutex_unlock(gl);
1818
1819         return 0;
1820 }
1821
1822 /**
1823  * gfs2_lvb_unhold - detach a LVB from a glock
1824  * @gl: The glock in question
1825  *
1826  */
1827
1828 void gfs2_lvb_unhold(struct gfs2_glock *gl)
1829 {
1830         gfs2_glock_hold(gl);
1831         gfs2_glmutex_lock(gl);
1832
1833         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count) > 0);
1834         if (atomic_dec_and_test(&gl->gl_lvb_count)) {
1835                 gfs2_lm_unhold_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1836                 gl->gl_lvb = NULL;
1837                 gfs2_glock_put(gl);
1838         }
1839
1840         gfs2_glmutex_unlock(gl);
1841         gfs2_glock_put(gl);
1842 }
1843
1844 void gfs2_lvb_sync(struct gfs2_glock *gl)
1845 {
1846         gfs2_glmutex_lock(gl);
1847
1848         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count));
1849         if (!gfs2_assert_warn(gl->gl_sbd, gfs2_glock_is_held_excl(gl)))
1850                 gfs2_lm_sync_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1851
1852         gfs2_glmutex_unlock(gl);
1853 }
1854
1855 static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
1856                         unsigned int state)
1857 {
1858         struct gfs2_glock *gl;
1859
1860         gl = gfs2_glock_find(sdp, name);
1861         if (!gl)
1862                 return;
1863
1864         if (gl->gl_ops->go_callback)
1865                 gl->gl_ops->go_callback(gl, state);
1866         handle_callback(gl, state);
1867
1868         spin_lock(&gl->gl_spin);
1869         run_queue(gl);
1870         spin_unlock(&gl->gl_spin);
1871
1872         gfs2_glock_put(gl);
1873 }
1874
1875 /**
1876  * gfs2_glock_cb - Callback used by locking module
1877  * @fsdata: Pointer to the superblock
1878  * @type: Type of callback
1879  * @data: Type dependent data pointer
1880  *
1881  * Called by the locking module when it wants to tell us something.
1882  * Either we need to drop a lock, one of our ASYNC requests completed, or
1883  * a journal from another client needs to be recovered.
1884  */
1885
1886 void gfs2_glock_cb(lm_fsdata_t *fsdata, unsigned int type, void *data)
1887 {
1888         struct gfs2_sbd *sdp = (struct gfs2_sbd *)fsdata;
1889
1890         switch (type) {
1891         case LM_CB_NEED_E:
1892                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_UNLOCKED);
1893                 return;
1894
1895         case LM_CB_NEED_D:
1896                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_DEFERRED);
1897                 return;
1898
1899         case LM_CB_NEED_S:
1900                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_SHARED);
1901                 return;
1902
1903         case LM_CB_ASYNC: {
1904                 struct lm_async_cb *async = (struct lm_async_cb *)data;
1905                 struct gfs2_glock *gl;
1906
1907                 gl = gfs2_glock_find(sdp, &async->lc_name);
1908                 if (gfs2_assert_warn(sdp, gl))
1909                         return;
1910                 if (!gfs2_assert_warn(sdp, gl->gl_req_bh))
1911                         gl->gl_req_bh(gl, async->lc_ret);
1912                 gfs2_glock_put(gl);
1913
1914                 return;
1915         }
1916
1917         case LM_CB_NEED_RECOVERY:
1918                 gfs2_jdesc_make_dirty(sdp, *(unsigned int *)data);
1919                 if (sdp->sd_recoverd_process)
1920                         wake_up_process(sdp->sd_recoverd_process);
1921                 return;
1922
1923         case LM_CB_DROPLOCKS:
1924                 gfs2_gl_hash_clear(sdp, NO_WAIT);
1925                 gfs2_quota_scan(sdp);
1926                 return;
1927
1928         default:
1929                 gfs2_assert_warn(sdp, 0);
1930                 return;
1931         }
1932 }
1933
1934 /**
1935  * gfs2_try_toss_inode - try to remove a particular inode struct from cache
1936  * sdp: the filesystem
1937  * inum: the inode number
1938  *
1939  */
1940
1941 void gfs2_try_toss_inode(struct gfs2_sbd *sdp, struct gfs2_inum *inum)
1942 {
1943         struct gfs2_glock *gl;
1944         struct gfs2_inode *ip;
1945         int error;
1946
1947         error = gfs2_glock_get(sdp, inum->no_addr, &gfs2_inode_glops,
1948                                NO_CREATE, &gl);
1949         if (error || !gl)
1950                 return;
1951
1952         if (!gfs2_glmutex_trylock(gl))
1953                 goto out;
1954
1955         ip = gl->gl_object;
1956         if (!ip)
1957                 goto out_unlock;
1958
1959         if (atomic_read(&ip->i_count))
1960                 goto out_unlock;
1961
1962         gfs2_inode_destroy(ip);
1963
1964  out_unlock:
1965         gfs2_glmutex_unlock(gl);
1966
1967  out:
1968         gfs2_glock_put(gl);
1969 }
1970
1971 /**
1972  * gfs2_iopen_go_callback - Try to kick the inode/vnode associated with an
1973  *                          iopen glock from memory
1974  * @io_gl: the iopen glock
1975  * @state: the state into which the glock should be put
1976  *
1977  */
1978
1979 void gfs2_iopen_go_callback(struct gfs2_glock *io_gl, unsigned int state)
1980 {
1981         struct gfs2_glock *i_gl;
1982
1983         if (state != LM_ST_UNLOCKED)
1984                 return;
1985
1986         spin_lock(&io_gl->gl_spin);
1987         i_gl = io_gl->gl_object;
1988         if (i_gl) {
1989                 gfs2_glock_hold(i_gl);
1990                 spin_unlock(&io_gl->gl_spin);
1991         } else {
1992                 spin_unlock(&io_gl->gl_spin);
1993                 return;
1994         }
1995
1996         if (gfs2_glmutex_trylock(i_gl)) {
1997                 struct gfs2_inode *ip = i_gl->gl_object;
1998                 if (ip) {
1999                         gfs2_try_toss_vnode(ip);
2000                         gfs2_glmutex_unlock(i_gl);
2001                         gfs2_glock_schedule_for_reclaim(i_gl);
2002                         goto out;
2003                 }
2004                 gfs2_glmutex_unlock(i_gl);
2005         }
2006
2007  out:
2008         gfs2_glock_put(i_gl);
2009 }
2010
2011 /**
2012  * demote_ok - Check to see if it's ok to unlock a glock
2013  * @gl: the glock
2014  *
2015  * Returns: 1 if it's ok
2016  */
2017
2018 static int demote_ok(struct gfs2_glock *gl)
2019 {
2020         struct gfs2_sbd *sdp = gl->gl_sbd;
2021         struct gfs2_glock_operations *glops = gl->gl_ops;
2022         int demote = 1;
2023
2024         if (test_bit(GLF_STICKY, &gl->gl_flags))
2025                 demote = 0;
2026         else if (test_bit(GLF_PREFETCH, &gl->gl_flags))
2027                 demote = time_after_eq(jiffies,
2028                                     gl->gl_stamp +
2029                                     gfs2_tune_get(sdp, gt_prefetch_secs) * HZ);
2030         else if (glops->go_demote_ok)
2031                 demote = glops->go_demote_ok(gl);
2032
2033         return demote;
2034 }
2035
2036 /**
2037  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
2038  * @gl: the glock
2039  *
2040  */
2041
2042 void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
2043 {
2044         struct gfs2_sbd *sdp = gl->gl_sbd;
2045
2046         spin_lock(&sdp->sd_reclaim_lock);
2047         if (list_empty(&gl->gl_reclaim)) {
2048                 gfs2_glock_hold(gl);
2049                 list_add(&gl->gl_reclaim, &sdp->sd_reclaim_list);
2050                 atomic_inc(&sdp->sd_reclaim_count);
2051         }
2052         spin_unlock(&sdp->sd_reclaim_lock);
2053
2054         wake_up(&sdp->sd_reclaim_wq);
2055 }
2056
2057 /**
2058  * gfs2_reclaim_glock - process the next glock on the filesystem's reclaim list
2059  * @sdp: the filesystem
2060  *
2061  * Called from gfs2_glockd() glock reclaim daemon, or when promoting a
2062  * different glock and we notice that there are a lot of glocks in the
2063  * reclaim list.
2064  *
2065  */
2066
2067 void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
2068 {
2069         struct gfs2_glock *gl;
2070
2071         spin_lock(&sdp->sd_reclaim_lock);
2072         if (list_empty(&sdp->sd_reclaim_list)) {
2073                 spin_unlock(&sdp->sd_reclaim_lock);
2074                 return;
2075         }
2076         gl = list_entry(sdp->sd_reclaim_list.next,
2077                         struct gfs2_glock, gl_reclaim);
2078         list_del_init(&gl->gl_reclaim);
2079         spin_unlock(&sdp->sd_reclaim_lock);
2080
2081         atomic_dec(&sdp->sd_reclaim_count);
2082         atomic_inc(&sdp->sd_reclaimed);
2083
2084         if (gfs2_glmutex_trylock(gl)) {
2085                 if (gl->gl_ops == &gfs2_inode_glops) {
2086                         struct gfs2_inode *ip = gl->gl_object;
2087                         if (ip && !atomic_read(&ip->i_count))
2088                                 gfs2_inode_destroy(ip);
2089                 }
2090                 if (queue_empty(gl, &gl->gl_holders) &&
2091                     gl->gl_state != LM_ST_UNLOCKED &&
2092                     demote_ok(gl))
2093                         handle_callback(gl, LM_ST_UNLOCKED);
2094                 gfs2_glmutex_unlock(gl);
2095         }
2096
2097         gfs2_glock_put(gl);
2098 }
2099
2100 /**
2101  * examine_bucket - Call a function for glock in a hash bucket
2102  * @examiner: the function
2103  * @sdp: the filesystem
2104  * @bucket: the bucket
2105  *
2106  * Returns: 1 if the bucket has entries
2107  */
2108
2109 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
2110                           struct gfs2_gl_hash_bucket *bucket)
2111 {
2112         struct glock_plug plug;
2113         struct list_head *tmp;
2114         struct gfs2_glock *gl;
2115         int entries;
2116
2117         /* Add "plug" to end of bucket list, work back up list from there */
2118         memset(&plug.gl_flags, 0, sizeof(unsigned long));
2119         set_bit(GLF_PLUG, &plug.gl_flags);
2120
2121         write_lock(&bucket->hb_lock);
2122         list_add(&plug.gl_list, &bucket->hb_list);
2123         write_unlock(&bucket->hb_lock);
2124
2125         for (;;) {
2126                 write_lock(&bucket->hb_lock);
2127
2128                 for (;;) {
2129                         tmp = plug.gl_list.next;
2130
2131                         if (tmp == &bucket->hb_list) {
2132                                 list_del(&plug.gl_list);
2133                                 entries = !list_empty(&bucket->hb_list);
2134                                 write_unlock(&bucket->hb_lock);
2135                                 return entries;
2136                         }
2137                         gl = list_entry(tmp, struct gfs2_glock, gl_list);
2138
2139                         /* Move plug up list */
2140                         list_move(&plug.gl_list, &gl->gl_list);
2141
2142                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2143                                 continue;
2144
2145                         /* examiner() must glock_put() */
2146                         gfs2_glock_hold(gl);
2147
2148                         break;
2149                 }
2150
2151                 write_unlock(&bucket->hb_lock);
2152
2153                 examiner(gl);
2154         }
2155 }
2156
2157 /**
2158  * scan_glock - look at a glock and see if we can reclaim it
2159  * @gl: the glock to look at
2160  *
2161  */
2162
2163 static void scan_glock(struct gfs2_glock *gl)
2164 {
2165         if (gfs2_glmutex_trylock(gl)) {
2166                 if (gl->gl_ops == &gfs2_inode_glops) {
2167                         struct gfs2_inode *ip = gl->gl_object;
2168                         if (ip && !atomic_read(&ip->i_count))
2169                                 goto out_schedule;
2170                 }
2171                 if (queue_empty(gl, &gl->gl_holders) &&
2172                     gl->gl_state != LM_ST_UNLOCKED &&
2173                     demote_ok(gl))
2174                         goto out_schedule;
2175
2176                 gfs2_glmutex_unlock(gl);
2177         }
2178
2179         gfs2_glock_put(gl);
2180
2181         return;
2182
2183  out_schedule:
2184         gfs2_glmutex_unlock(gl);
2185         gfs2_glock_schedule_for_reclaim(gl);
2186         gfs2_glock_put(gl);
2187 }
2188
2189 /**
2190  * gfs2_scand_internal - Look for glocks and inodes to toss from memory
2191  * @sdp: the filesystem
2192  *
2193  */
2194
2195 void gfs2_scand_internal(struct gfs2_sbd *sdp)
2196 {
2197         unsigned int x;
2198
2199         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2200                 examine_bucket(scan_glock, sdp, &sdp->sd_gl_hash[x]);
2201                 cond_resched();
2202         }
2203 }
2204
2205 /**
2206  * clear_glock - look at a glock and see if we can free it from glock cache
2207  * @gl: the glock to look at
2208  *
2209  */
2210
2211 static void clear_glock(struct gfs2_glock *gl)
2212 {
2213         struct gfs2_sbd *sdp = gl->gl_sbd;
2214         int released;
2215
2216         spin_lock(&sdp->sd_reclaim_lock);
2217         if (!list_empty(&gl->gl_reclaim)) {
2218                 list_del_init(&gl->gl_reclaim);
2219                 atomic_dec(&sdp->sd_reclaim_count);
2220                 released = gfs2_glock_put(gl);
2221                 gfs2_assert(sdp, !released);
2222         }
2223         spin_unlock(&sdp->sd_reclaim_lock);
2224
2225         if (gfs2_glmutex_trylock(gl)) {
2226                 if (gl->gl_ops == &gfs2_inode_glops) {
2227                         struct gfs2_inode *ip = gl->gl_object;
2228                         if (ip && !atomic_read(&ip->i_count))
2229                                 gfs2_inode_destroy(ip);
2230                 }
2231                 if (queue_empty(gl, &gl->gl_holders) &&
2232                     gl->gl_state != LM_ST_UNLOCKED)
2233                         handle_callback(gl, LM_ST_UNLOCKED);
2234
2235                 gfs2_glmutex_unlock(gl);
2236         }
2237
2238         gfs2_glock_put(gl);
2239 }
2240
2241 /**
2242  * gfs2_gl_hash_clear - Empty out the glock hash table
2243  * @sdp: the filesystem
2244  * @wait: wait until it's all gone
2245  *
2246  * Called when unmounting the filesystem, or when inter-node lock manager
2247  * requests DROPLOCKS because it is running out of capacity.
2248  */
2249
2250 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
2251 {
2252         unsigned long t;
2253         unsigned int x;
2254         int cont;
2255
2256         t = jiffies;
2257
2258         for (;;) {
2259                 cont = 0;
2260
2261                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
2262                         if (examine_bucket(clear_glock, sdp,
2263                                            &sdp->sd_gl_hash[x]))
2264                                 cont = 1;
2265
2266                 if (!wait || !cont)
2267                         break;
2268
2269                 if (time_after_eq(jiffies,
2270                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
2271                         fs_warn(sdp, "Unmount seems to be stalled. "
2272                                      "Dumping lock state...\n");
2273                         gfs2_dump_lockstate(sdp);
2274                         t = jiffies;
2275                 }
2276
2277                 /* invalidate_inodes() requires that the sb inodes list
2278                    not change, but an async completion callback for an
2279                    unlock can occur which does glock_put() which
2280                    can call iput() which will change the sb inodes list.
2281                    invalidate_inodes_mutex prevents glock_put()'s during
2282                    an invalidate_inodes() */
2283
2284                 mutex_lock(&sdp->sd_invalidate_inodes_mutex);
2285                 invalidate_inodes(sdp->sd_vfs);
2286                 mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
2287                 yield();
2288         }
2289 }
2290
2291 /*
2292  *  Diagnostic routines to help debug distributed deadlock
2293  */
2294
2295 /**
2296  * dump_holder - print information about a glock holder
2297  * @str: a string naming the type of holder
2298  * @gh: the glock holder
2299  *
2300  * Returns: 0 on success, -ENOBUFS when we run out of space
2301  */
2302
2303 static int dump_holder(char *str, struct gfs2_holder *gh)
2304 {
2305         unsigned int x;
2306         int error = -ENOBUFS;
2307
2308         printk(KERN_INFO "  %s\n", str);
2309         printk(KERN_INFO "    owner = %ld\n",
2310                    (gh->gh_owner) ? (long)gh->gh_owner->pid : -1);
2311         printk(KERN_INFO "    gh_state = %u\n", gh->gh_state);
2312         printk(KERN_INFO "    gh_flags =");
2313         for (x = 0; x < 32; x++)
2314                 if (gh->gh_flags & (1 << x))
2315                         printk(" %u", x);
2316         printk(" \n");
2317         printk(KERN_INFO "    error = %d\n", gh->gh_error);
2318         printk(KERN_INFO "    gh_iflags =");
2319         for (x = 0; x < 32; x++)
2320                 if (test_bit(x, &gh->gh_iflags))
2321                         printk(" %u", x);
2322         printk(" \n");
2323         print_symbol(KERN_INFO "    initialized at: %s\n", gh->gh_ip);
2324
2325         error = 0;
2326
2327         return error;
2328 }
2329
2330 /**
2331  * dump_inode - print information about an inode
2332  * @ip: the inode
2333  *
2334  * Returns: 0 on success, -ENOBUFS when we run out of space
2335  */
2336
2337 static int dump_inode(struct gfs2_inode *ip)
2338 {
2339         unsigned int x;
2340         int error = -ENOBUFS;
2341
2342         printk(KERN_INFO "  Inode:\n");
2343         printk(KERN_INFO "    num = %llu %llu\n",
2344                     ip->i_num.no_formal_ino, ip->i_num.no_addr);
2345         printk(KERN_INFO "    type = %u\n", IF2DT(ip->i_di.di_mode));
2346         printk(KERN_INFO "    i_count = %d\n", atomic_read(&ip->i_count));
2347         printk(KERN_INFO "    i_flags =");
2348         for (x = 0; x < 32; x++)
2349                 if (test_bit(x, &ip->i_flags))
2350                         printk(" %u", x);
2351         printk(" \n");
2352         printk(KERN_INFO "    vnode = %s\n", (ip->i_vnode) ? "yes" : "no");
2353
2354         error = 0;
2355
2356         return error;
2357 }
2358
2359 /**
2360  * dump_glock - print information about a glock
2361  * @gl: the glock
2362  * @count: where we are in the buffer
2363  *
2364  * Returns: 0 on success, -ENOBUFS when we run out of space
2365  */
2366
2367 static int dump_glock(struct gfs2_glock *gl)
2368 {
2369         struct gfs2_holder *gh;
2370         unsigned int x;
2371         int error = -ENOBUFS;
2372
2373         spin_lock(&gl->gl_spin);
2374
2375         printk(KERN_INFO "Glock (%u, %llu)\n",
2376                     gl->gl_name.ln_type,
2377                     gl->gl_name.ln_number);
2378         printk(KERN_INFO "  gl_flags =");
2379         for (x = 0; x < 32; x++)
2380                 if (test_bit(x, &gl->gl_flags))
2381                         printk(" %u", x);
2382         printk(" \n");
2383         printk(KERN_INFO "  gl_ref = %d\n", atomic_read(&gl->gl_ref.refcount));
2384         printk(KERN_INFO "  gl_state = %u\n", gl->gl_state);
2385         printk(KERN_INFO "  req_gh = %s\n", (gl->gl_req_gh) ? "yes" : "no");
2386         printk(KERN_INFO "  req_bh = %s\n", (gl->gl_req_bh) ? "yes" : "no");
2387         printk(KERN_INFO "  lvb_count = %d\n", atomic_read(&gl->gl_lvb_count));
2388         printk(KERN_INFO "  object = %s\n", (gl->gl_object) ? "yes" : "no");
2389         printk(KERN_INFO "  le = %s\n",
2390                    (list_empty(&gl->gl_le.le_list)) ? "no" : "yes");
2391         printk(KERN_INFO "  reclaim = %s\n",
2392                     (list_empty(&gl->gl_reclaim)) ? "no" : "yes");
2393         if (gl->gl_aspace)
2394                 printk(KERN_INFO "  aspace = %lu\n",
2395                             gl->gl_aspace->i_mapping->nrpages);
2396         else
2397                 printk(KERN_INFO "  aspace = no\n");
2398         printk(KERN_INFO "  ail = %d\n", atomic_read(&gl->gl_ail_count));
2399         if (gl->gl_req_gh) {
2400                 error = dump_holder("Request", gl->gl_req_gh);
2401                 if (error)
2402                         goto out;
2403         }
2404         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
2405                 error = dump_holder("Holder", gh);
2406                 if (error)
2407                         goto out;
2408         }
2409         list_for_each_entry(gh, &gl->gl_waiters1, gh_list) {
2410                 error = dump_holder("Waiter1", gh);
2411                 if (error)
2412                         goto out;
2413         }
2414         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
2415                 error = dump_holder("Waiter2", gh);
2416                 if (error)
2417                         goto out;
2418         }
2419         list_for_each_entry(gh, &gl->gl_waiters3, gh_list) {
2420                 error = dump_holder("Waiter3", gh);
2421                 if (error)
2422                         goto out;
2423         }
2424         if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object) {
2425                 if (!test_bit(GLF_LOCK, &gl->gl_flags) &&
2426                     list_empty(&gl->gl_holders)) {
2427                         error = dump_inode(gl->gl_object);
2428                         if (error)
2429                                 goto out;
2430                 } else {
2431                         error = -ENOBUFS;
2432                         printk(KERN_INFO "  Inode: busy\n");
2433                 }
2434         }
2435
2436         error = 0;
2437
2438  out:
2439         spin_unlock(&gl->gl_spin);
2440
2441         return error;
2442 }
2443
2444 /**
2445  * gfs2_dump_lockstate - print out the current lockstate
2446  * @sdp: the filesystem
2447  * @ub: the buffer to copy the information into
2448  *
2449  * If @ub is NULL, dump the lockstate to the console.
2450  *
2451  */
2452
2453 int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
2454 {
2455         struct gfs2_gl_hash_bucket *bucket;
2456         struct gfs2_glock *gl;
2457         unsigned int x;
2458         int error = 0;
2459
2460         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2461                 bucket = &sdp->sd_gl_hash[x];
2462
2463                 read_lock(&bucket->hb_lock);
2464
2465                 list_for_each_entry(gl, &bucket->hb_list, gl_list) {
2466                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2467                                 continue;
2468
2469                         error = dump_glock(gl);
2470                         if (error)
2471                                 break;
2472                 }
2473
2474                 read_unlock(&bucket->hb_lock);
2475
2476                 if (error)
2477                         break;
2478         }
2479
2480
2481         return error;
2482 }
2483