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