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/gfs2_ondisk.h>
  20 #include <asm/semaphore.h>
  21 #include <asm/uaccess.h>
  22
  23 #include "gfs2.h"
  24 #include "lm_interface.h"
  25 #include "incore.h"
  26 #include "glock.h"
  27 #include "glops.h"
  28 #include "inode.h"
  29 #include "lm.h"
  30 #include "lops.h"
  31 #include "meta_io.h"
  32 #include "quota.h"
  33 #include "super.h"
  34 #include "util.h"
  35
  36 /*  Must be kept in sync with the beginning of struct gfs2_glock  */
  37 struct glock_plug {
  38         struct list_head gl_list;
  39         unsigned long gl_flags;
  40 };
  41
  42 struct greedy {
  43         struct gfs2_holder gr_gh;
  44         struct work_struct gr_work;
  45 };
  46
  47 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  48
  49 /**
  50  * relaxed_state_ok - is a requested lock compatible with the current lock mode?
  51  * @actual: the current state of the lock
  52  * @requested: the lock state that was requested by the caller
  53  * @flags: the modifier flags passed in by the caller
  54  *
  55  * Returns: 1 if the locks are compatible, 0 otherwise
  56  */
  57
  58 static inline int relaxed_state_ok(unsigned int actual, unsigned requested,
  59                                    int flags)
  60 {
  61         if (actual == requested)
  62                 return 1;
  63
  64         if (flags & GL_EXACT)
  65                 return 0;
  66
  67         if (actual == LM_ST_EXCLUSIVE && requested == LM_ST_SHARED)
  68                 return 1;
  69
  70         if (actual != LM_ST_UNLOCKED && (flags & LM_FLAG_ANY))
  71                 return 1;
  72
  73         return 0;
  74 }
  75
  76 /**
  77  * gl_hash() - Turn glock number into hash bucket number
  78  * @lock: The glock number
  79  *
  80  * Returns: The number of the corresponding hash bucket
  81  */
  82
  83 static unsigned int gl_hash(struct lm_lockname *name)
  84 {
  85         unsigned int h;
  86
  87         h = jhash(&name->ln_number, sizeof(uint64_t), 0);
  88         h = jhash(&name->ln_type, sizeof(unsigned int), h);
  89         h &= GFS2_GL_HASH_MASK;
  90
  91         return h;
  92 }
  93
  94 /**
  95  * glock_free() - Perform a few checks and then release struct gfs2_glock
  96  * @gl: The glock to release
  97  *
  98  * Also calls lock module to release its internal structure for this glock.
  99  *
 100  */
 101
 102 static void glock_free(struct gfs2_glock *gl)
 103 {
 104         struct gfs2_sbd *sdp = gl->gl_sbd;
 105         struct inode *aspace = gl->gl_aspace;
 106
 107         gfs2_lm_put_lock(sdp, gl->gl_lock);
 108
 109         if (aspace)
 110                 gfs2_aspace_put(aspace);
 111
 112         kmem_cache_free(gfs2_glock_cachep, gl);
 113 }
 114
 115 /**
 116  * gfs2_glock_hold() - increment reference count on glock
 117  * @gl: The glock to hold
 118  *
 119  */
 120
 121 void gfs2_glock_hold(struct gfs2_glock *gl)
 122 {
 123         kref_get(&gl->gl_ref);
 124 }
 125
 126 /* All work is done after the return from kref_put() so we
 127    can release the write_lock before the free. */
 128
 129 static void kill_glock(struct kref *kref)
 130 {
 131         struct gfs2_glock *gl = container_of(kref, struct gfs2_glock, gl_ref);
 132         struct gfs2_sbd *sdp = gl->gl_sbd;
 133
 134         gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
 135         gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
 136         gfs2_assert(sdp, list_empty(&gl->gl_holders));
 137         gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
 138         gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
 139         gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
 140 }
 141
 142 /**
 143  * gfs2_glock_put() - Decrement reference count on glock
 144  * @gl: The glock to put
 145  *
 146  */
 147
 148 int gfs2_glock_put(struct gfs2_glock *gl)
 149 {
 150         struct gfs2_sbd *sdp = gl->gl_sbd;
 151         struct gfs2_gl_hash_bucket *bucket = gl->gl_bucket;
 152         int rv = 0;
 153
 154         mutex_lock(&sdp->sd_invalidate_inodes_mutex);
 155
 156         write_lock(&bucket->hb_lock);
 157         if (kref_put(&gl->gl_ref, kill_glock)) {
 158                 list_del_init(&gl->gl_list);
 159                 write_unlock(&bucket->hb_lock);
 160                 glock_free(gl);
 161                 rv = 1;
 162                 goto out;
 163         }
 164         write_unlock(&bucket->hb_lock);
 165  out:
 166         mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
 167         return rv;
 168 }
 169
 170 /**
 171  * queue_empty - check to see if a glock's queue is empty
 172  * @gl: the glock
 173  * @head: the head of the queue to check
 174  *
 175  * This function protects the list in the event that a process already
 176  * has a holder on the list and is adding a second holder for itself.
 177  * The glmutex lock is what generally prevents processes from working
 178  * on the same glock at once, but the special case of adding a second
 179  * holder for yourself ("recursive" locking) doesn't involve locking
 180  * glmutex, making the spin lock necessary.
 181  *
 182  * Returns: 1 if the queue is empty
 183  */
 184
 185 static inline int queue_empty(struct gfs2_glock *gl, struct list_head *head)
 186 {
 187         int empty;
 188         spin_lock(&gl->gl_spin);
 189         empty = list_empty(head);
 190         spin_unlock(&gl->gl_spin);
 191         return empty;
 192 }
 193
 194 /**
 195  * search_bucket() - Find struct gfs2_glock by lock number
 196  * @bucket: the bucket to search
 197  * @name: The lock name
 198  *
 199  * Returns: NULL, or the struct gfs2_glock with the requested number
 200  */
 201
 202 static struct gfs2_glock *search_bucket(struct gfs2_gl_hash_bucket *bucket,
 203                                         struct lm_lockname *name)
 204 {
 205         struct gfs2_glock *gl;
 206
 207         list_for_each_entry(gl, &bucket->hb_list, gl_list) {
 208                 if (test_bit(GLF_PLUG, &gl->gl_flags))
 209                         continue;
 210                 if (!lm_name_equal(&gl->gl_name, name))
 211                         continue;
 212
 213                 kref_get(&gl->gl_ref);
 214
 215                 return gl;
 216         }
 217
 218         return NULL;
 219 }
 220
 221 /**
 222  * gfs2_glock_find() - Find glock by lock number
 223  * @sdp: The GFS2 superblock
 224  * @name: The lock name
 225  *
 226  * Returns: NULL, or the struct gfs2_glock with the requested number
 227  */
 228
 229 struct gfs2_glock *gfs2_glock_find(struct gfs2_sbd *sdp,
 230                                    struct lm_lockname *name)
 231 {
 232         struct gfs2_gl_hash_bucket *bucket = &sdp->sd_gl_hash[gl_hash(name)];
 233         struct gfs2_glock *gl;
 234
 235         read_lock(&bucket->hb_lock);
 236         gl = search_bucket(bucket, name);
 237         read_unlock(&bucket->hb_lock);
 238
 239         return gl;
 240 }
 241
 242 /**
 243  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 244  * @sdp: The GFS2 superblock
 245  * @number: the lock number
 246  * @glops: The glock_operations to use
 247  * @create: If 0, don't create the glock if it doesn't exist
 248  * @glp: the glock is returned here
 249  *
 250  * This does not lock a glock, just finds/creates structures for one.
 251  *
 252  * Returns: errno
 253  */
 254
 255 int gfs2_glock_get(struct gfs2_sbd *sdp, uint64_t number,
 256                    struct gfs2_glock_operations *glops, int create,
 257                    struct gfs2_glock **glp)
 258 {
 259         struct lm_lockname name;
 260         struct gfs2_glock *gl, *tmp;
 261         struct gfs2_gl_hash_bucket *bucket;
 262         int error;
 263
 264         name.ln_number = number;
 265         name.ln_type = glops->go_type;
 266         bucket = &sdp->sd_gl_hash[gl_hash(&name)];
 267
 268         read_lock(&bucket->hb_lock);
 269         gl = search_bucket(bucket, &name);
 270         read_unlock(&bucket->hb_lock);
 271
 272         if (gl || !create) {
 273                 *glp = gl;
 274                 return 0;
 275         }
 276
 277         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
 278         if (!gl)
 279                 return -ENOMEM;
 280
 281         memset(gl, 0, sizeof(struct gfs2_glock));
 282
 283         INIT_LIST_HEAD(&gl->gl_list);
 284         gl->gl_name = name;
 285         kref_init(&gl->gl_ref);
 286
 287         spin_lock_init(&gl->gl_spin);
 288
 289         gl->gl_state = LM_ST_UNLOCKED;
 290         INIT_LIST_HEAD(&gl->gl_holders);
 291         INIT_LIST_HEAD(&gl->gl_waiters1);
 292         INIT_LIST_HEAD(&gl->gl_waiters2);
 293         INIT_LIST_HEAD(&gl->gl_waiters3);
 294
 295         gl->gl_ops = glops;
 296
 297         gl->gl_bucket = bucket;
 298         INIT_LIST_HEAD(&gl->gl_reclaim);
 299
 300         gl->gl_sbd = sdp;
 301
 302         lops_init_le(&gl->gl_le, &gfs2_glock_lops);
 303         INIT_LIST_HEAD(&gl->gl_ail_list);
 304
 305         /* If this glock protects actual on-disk data or metadata blocks,
 306            create a VFS inode to manage the pages/buffers holding them. */
 307         if (glops == &gfs2_inode_glops ||
 308             glops == &gfs2_rgrp_glops ||
 309             glops == &gfs2_meta_glops) {
 310                 gl->gl_aspace = gfs2_aspace_get(sdp);
 311                 if (!gl->gl_aspace) {
 312                         error = -ENOMEM;
 313                         goto fail;
 314                 }
 315         }
 316
 317         error = gfs2_lm_get_lock(sdp, &name, &gl->gl_lock);
 318         if (error)
 319                 goto fail_aspace;
 320
 321         write_lock(&bucket->hb_lock);
 322         tmp = search_bucket(bucket, &name);
 323         if (tmp) {
 324                 write_unlock(&bucket->hb_lock);
 325                 glock_free(gl);
 326                 gl = tmp;
 327         } else {
 328                 list_add_tail(&gl->gl_list, &bucket->hb_list);
 329                 write_unlock(&bucket->hb_lock);
 330         }
 331
 332         *glp = gl;
 333
 334         return 0;
 335
 336  fail_aspace:
 337         if (gl->gl_aspace)
 338                 gfs2_aspace_put(gl->gl_aspace);
 339
 340  fail:
 341         kmem_cache_free(gfs2_glock_cachep, gl);
 342
 343         return error;
 344 }
 345
 346 /**
 347  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 348  * @gl: the glock
 349  * @state: the state we're requesting
 350  * @flags: the modifier flags
 351  * @gh: the holder structure
 352  *
 353  */
 354
 355 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, int flags,
 356                       struct gfs2_holder *gh)
 357 {
 358         INIT_LIST_HEAD(&gh->gh_list);
 359         gh->gh_gl = gl;
 360         gh->gh_owner = (flags & GL_NEVER_RECURSE) ? NULL : current;
 361         gh->gh_state = state;
 362         gh->gh_flags = flags;
 363         gh->gh_error = 0;
 364         gh->gh_iflags = 0;
 365         init_completion(&gh->gh_wait);
 366
 367         if (gh->gh_state == LM_ST_EXCLUSIVE)
 368                 gh->gh_flags |= GL_LOCAL_EXCL;
 369
 370         gfs2_glock_hold(gl);
 371 }
 372
 373 /**
 374  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 375  * @state: the state we're requesting
 376  * @flags: the modifier flags
 377  * @gh: the holder structure
 378  *
 379  * Don't mess with the glock.
 380  *
 381  */
 382
 383 void gfs2_holder_reinit(unsigned int state, int flags, struct gfs2_holder *gh)
 384 {
 385         gh->gh_state = state;
 386         gh->gh_flags = flags;
 387         if (gh->gh_state == LM_ST_EXCLUSIVE)
 388                 gh->gh_flags |= GL_LOCAL_EXCL;
 389
 390         gh->gh_iflags &= 1 << HIF_ALLOCED;
 391 }
 392
 393 /**
 394  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 395  * @gh: the holder structure
 396  *
 397  */
 398
 399 void gfs2_holder_uninit(struct gfs2_holder *gh)
 400 {
 401         gfs2_glock_put(gh->gh_gl);
 402         gh->gh_gl = NULL;
 403 }
 404
 405 /**
 406  * gfs2_holder_get - get a struct gfs2_holder structure
 407  * @gl: the glock
 408  * @state: the state we're requesting
 409  * @flags: the modifier flags
 410  * @gfp_flags: __GFP_NOFAIL
 411  *
 412  * Figure out how big an impact this function has.  Either:
 413  * 1) Replace it with a cache of structures hanging off the struct gfs2_sbd
 414  * 2) Leave it like it is
 415  *
 416  * Returns: the holder structure, NULL on ENOMEM
 417  */
 418
 419 struct gfs2_holder *gfs2_holder_get(struct gfs2_glock *gl, unsigned int state,
 420                                     int flags, gfp_t gfp_flags)
 421 {
 422         struct gfs2_holder *gh;
 423
 424         gh = kmalloc(sizeof(struct gfs2_holder), gfp_flags);
 425         if (!gh)
 426                 return NULL;
 427
 428         gfs2_holder_init(gl, state, flags, gh);
 429         set_bit(HIF_ALLOCED, &gh->gh_iflags);
 430
 431         return gh;
 432 }
 433
 434 /**
 435  * gfs2_holder_put - get rid of a struct gfs2_holder structure
 436  * @gh: the holder structure
 437  *
 438  */
 439
 440 void gfs2_holder_put(struct gfs2_holder *gh)
 441 {
 442         gfs2_holder_uninit(gh);
 443         kfree(gh);
 444 }
 445
 446 /**
 447  * handle_recurse - put other holder structures (marked recursive)
 448  *                  into the holders list
 449  * @gh: the holder structure
 450  *
 451  */
 452
 453 static void handle_recurse(struct gfs2_holder *gh)
 454 {
 455         struct gfs2_glock *gl = gh->gh_gl;
 456         struct gfs2_sbd *sdp = gl->gl_sbd;
 457         struct gfs2_holder *tmp_gh, *safe;
 458         int found = 0;
 459
 460         if (gfs2_assert_warn(sdp, gh->gh_owner))
 461                 return;
 462
 463         list_for_each_entry_safe(tmp_gh, safe, &gl->gl_waiters3, gh_list) {
 464                 if (tmp_gh->gh_owner != gh->gh_owner)
 465                         continue;
 466
 467                 gfs2_assert_warn(sdp,
 468                                  test_bit(HIF_RECURSE, &tmp_gh->gh_iflags));
 469
 470                 list_move_tail(&tmp_gh->gh_list, &gl->gl_holders);
 471                 tmp_gh->gh_error = 0;
 472                 set_bit(HIF_HOLDER, &tmp_gh->gh_iflags);
 473
 474                 complete(&tmp_gh->gh_wait);
 475
 476                 found = 1;
 477         }
 478
 479         gfs2_assert_warn(sdp, found);
 480 }
 481
 482 /**
 483  * do_unrecurse - a recursive holder was just dropped of the waiters3 list
 484  * @gh: the holder
 485  *
 486  * If there is only one other recursive holder, clear its HIF_RECURSE bit.
 487  * If there is more than one, leave them alone.
 488  *
 489  */
 490
 491 static void do_unrecurse(struct gfs2_holder *gh)
 492 {
 493         struct gfs2_glock *gl = gh->gh_gl;
 494         struct gfs2_sbd *sdp = gl->gl_sbd;
 495         struct gfs2_holder *tmp_gh, *last_gh = NULL;
 496         int found = 0;
 497
 498         if (gfs2_assert_warn(sdp, gh->gh_owner))
 499                 return;
 500
 501         list_for_each_entry(tmp_gh, &gl->gl_waiters3, gh_list) {
 502                 if (tmp_gh->gh_owner != gh->gh_owner)
 503                         continue;
 504
 505                 gfs2_assert_warn(sdp,
 506                                  test_bit(HIF_RECURSE, &tmp_gh->gh_iflags));
 507
 508                 if (found)
 509                         return;
 510
 511                 found = 1;
 512                 last_gh = tmp_gh;
 513         }
 514
 515         if (!gfs2_assert_warn(sdp, found))
 516                 clear_bit(HIF_RECURSE, &last_gh->gh_iflags);
 517 }
 518
 519 /**
 520  * rq_mutex - process a mutex request in the queue
 521  * @gh: the glock holder
 522  *
 523  * Returns: 1 if the queue is blocked
 524  */
 525
 526 static int rq_mutex(struct gfs2_holder *gh)
 527 {
 528         struct gfs2_glock *gl = gh->gh_gl;
 529
 530         list_del_init(&gh->gh_list);
 531         /*  gh->gh_error never examined.  */
 532         set_bit(GLF_LOCK, &gl->gl_flags);
 533         complete(&gh->gh_wait);
 534
 535         return 1;
 536 }
 537
 538 /**
 539  * rq_promote - process a promote request in the queue
 540  * @gh: the glock holder
 541  *
 542  * Acquire a new inter-node lock, or change a lock state to more restrictive.
 543  *
 544  * Returns: 1 if the queue is blocked
 545  */
 546
 547 static int rq_promote(struct gfs2_holder *gh)
 548 {
 549         struct gfs2_glock *gl = gh->gh_gl;
 550         struct gfs2_sbd *sdp = gl->gl_sbd;
 551         struct gfs2_glock_operations *glops = gl->gl_ops;
 552         int recurse;
 553
 554         if (!relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 555                 if (list_empty(&gl->gl_holders)) {
 556                         gl->gl_req_gh = gh;
 557                         set_bit(GLF_LOCK, &gl->gl_flags);
 558                         spin_unlock(&gl->gl_spin);
 559
 560                         if (atomic_read(&sdp->sd_reclaim_count) >
 561                             gfs2_tune_get(sdp, gt_reclaim_limit) &&
 562                             !(gh->gh_flags & LM_FLAG_PRIORITY)) {
 563                                 gfs2_reclaim_glock(sdp);
 564                                 gfs2_reclaim_glock(sdp);
 565                         }
 566
 567                         glops->go_xmote_th(gl, gh->gh_state,
 568                                            gh->gh_flags);
 569
 570                         spin_lock(&gl->gl_spin);
 571                 }
 572                 return 1;
 573         }
 574
 575         if (list_empty(&gl->gl_holders)) {
 576                 set_bit(HIF_FIRST, &gh->gh_iflags);
 577                 set_bit(GLF_LOCK, &gl->gl_flags);
 578                 recurse = 0;
 579         } else {
 580                 struct gfs2_holder *next_gh;
 581                 if (gh->gh_flags & GL_LOCAL_EXCL)
 582                         return 1;
 583                 next_gh = list_entry(gl->gl_holders.next, struct gfs2_holder,
 584                                      gh_list);
 585                 if (next_gh->gh_flags & GL_LOCAL_EXCL)
 586                          return 1;
 587                 recurse = test_bit(HIF_RECURSE, &gh->gh_iflags);
 588         }
 589
 590         list_move_tail(&gh->gh_list, &gl->gl_holders);
 591         gh->gh_error = 0;
 592         set_bit(HIF_HOLDER, &gh->gh_iflags);
 593
 594         if (recurse)
 595                 handle_recurse(gh);
 596
 597         complete(&gh->gh_wait);
 598
 599         return 0;
 600 }
 601
 602 /**
 603  * rq_demote - process a demote request in the queue
 604  * @gh: the glock holder
 605  *
 606  * Returns: 1 if the queue is blocked
 607  */
 608
 609 static int rq_demote(struct gfs2_holder *gh)
 610 {
 611         struct gfs2_glock *gl = gh->gh_gl;
 612         struct gfs2_glock_operations *glops = gl->gl_ops;
 613
 614         if (!list_empty(&gl->gl_holders))
 615                 return 1;
 616
 617         if (gl->gl_state == gh->gh_state || gl->gl_state == LM_ST_UNLOCKED) {
 618                 list_del_init(&gh->gh_list);
 619                 gh->gh_error = 0;
 620                 spin_unlock(&gl->gl_spin);
 621                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 622                         gfs2_holder_put(gh);
 623                 else
 624                         complete(&gh->gh_wait);
 625                 spin_lock(&gl->gl_spin);
 626         } else {
 627                 gl->gl_req_gh = gh;
 628                 set_bit(GLF_LOCK, &gl->gl_flags);
 629                 spin_unlock(&gl->gl_spin);
 630
 631                 if (gh->gh_state == LM_ST_UNLOCKED ||
 632                     gl->gl_state != LM_ST_EXCLUSIVE)
 633                         glops->go_drop_th(gl);
 634                 else
 635                         glops->go_xmote_th(gl, gh->gh_state, gh->gh_flags);
 636
 637                 spin_lock(&gl->gl_spin);
 638         }
 639
 640         return 0;
 641 }
 642
 643 /**
 644  * rq_greedy - process a queued request to drop greedy status
 645  * @gh: the glock holder
 646  *
 647  * Returns: 1 if the queue is blocked
 648  */
 649
 650 static int rq_greedy(struct gfs2_holder *gh)
 651 {
 652         struct gfs2_glock *gl = gh->gh_gl;
 653
 654         list_del_init(&gh->gh_list);
 655         /*  gh->gh_error never examined.  */
 656         clear_bit(GLF_GREEDY, &gl->gl_flags);
 657         spin_unlock(&gl->gl_spin);
 658
 659         gfs2_holder_uninit(gh);
 660         kfree(container_of(gh, struct greedy, gr_gh));
 661
 662         spin_lock(&gl->gl_spin);
 663
 664         return 0;
 665 }
 666
 667 /**
 668  * run_queue - process holder structures on a glock
 669  * @gl: the glock
 670  *
 671  */
 672
 673 static void run_queue(struct gfs2_glock *gl)
 674 {
 675         struct gfs2_holder *gh;
 676         int blocked = 1;
 677
 678         for (;;) {
 679                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 680                         break;
 681
 682                 if (!list_empty(&gl->gl_waiters1)) {
 683                         gh = list_entry(gl->gl_waiters1.next,
 684                                         struct gfs2_holder, gh_list);
 685
 686                         if (test_bit(HIF_MUTEX, &gh->gh_iflags))
 687                                 blocked = rq_mutex(gh);
 688                         else
 689                                 gfs2_assert_warn(gl->gl_sbd, 0);
 690
 691                 } else if (!list_empty(&gl->gl_waiters2) &&
 692                            !test_bit(GLF_SKIP_WAITERS2, &gl->gl_flags)) {
 693                         gh = list_entry(gl->gl_waiters2.next,
 694                                         struct gfs2_holder, gh_list);
 695
 696                         if (test_bit(HIF_DEMOTE, &gh->gh_iflags))
 697                                 blocked = rq_demote(gh);
 698                         else if (test_bit(HIF_GREEDY, &gh->gh_iflags))
 699                                 blocked = rq_greedy(gh);
 700                         else
 701                                 gfs2_assert_warn(gl->gl_sbd, 0);
 702
 703                 } else if (!list_empty(&gl->gl_waiters3)) {
 704                         gh = list_entry(gl->gl_waiters3.next,
 705                                         struct gfs2_holder, gh_list);
 706
 707                         if (test_bit(HIF_PROMOTE, &gh->gh_iflags))
 708                                 blocked = rq_promote(gh);
 709                         else
 710                                 gfs2_assert_warn(gl->gl_sbd, 0);
 711
 712                 } else
 713                         break;
 714
 715                 if (blocked)
 716                         break;
 717         }
 718 }
 719
 720 /**
 721  * gfs2_glmutex_lock - acquire a local lock on a glock
 722  * @gl: the glock
 723  *
 724  * Gives caller exclusive access to manipulate a glock structure.
 725  */
 726
 727 void gfs2_glmutex_lock(struct gfs2_glock *gl)
 728 {
 729         struct gfs2_holder gh;
 730
 731         gfs2_holder_init(gl, 0, 0, &gh);
 732         set_bit(HIF_MUTEX, &gh.gh_iflags);
 733
 734         spin_lock(&gl->gl_spin);
 735         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 736                 list_add_tail(&gh.gh_list, &gl->gl_waiters1);
 737         else
 738                 complete(&gh.gh_wait);
 739         spin_unlock(&gl->gl_spin);
 740
 741         wait_for_completion(&gh.gh_wait);
 742         gfs2_holder_uninit(&gh);
 743 }
 744
 745 /**
 746  * gfs2_glmutex_trylock - try to acquire a local lock on a glock
 747  * @gl: the glock
 748  *
 749  * Returns: 1 if the glock is acquired
 750  */
 751
 752 int gfs2_glmutex_trylock(struct gfs2_glock *gl)
 753 {
 754         int acquired = 1;
 755
 756         spin_lock(&gl->gl_spin);
 757         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 758                 acquired = 0;
 759         spin_unlock(&gl->gl_spin);
 760
 761         return acquired;
 762 }
 763
 764 /**
 765  * gfs2_glmutex_unlock - release a local lock on a glock
 766  * @gl: the glock
 767  *
 768  */
 769
 770 void gfs2_glmutex_unlock(struct gfs2_glock *gl)
 771 {
 772         spin_lock(&gl->gl_spin);
 773         clear_bit(GLF_LOCK, &gl->gl_flags);
 774         run_queue(gl);
 775         spin_unlock(&gl->gl_spin);
 776 }
 777
 778 /**
 779  * handle_callback - add a demote request to a lock's queue
 780  * @gl: the glock
 781  * @state: the state the caller wants us to change to
 782  *
 783  */
 784
 785 static void handle_callback(struct gfs2_glock *gl, unsigned int state)
 786 {
 787         struct gfs2_holder *gh, *new_gh = NULL;
 788
 789  restart:
 790         spin_lock(&gl->gl_spin);
 791
 792         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
 793                 if (test_bit(HIF_DEMOTE, &gh->gh_iflags) &&
 794                     gl->gl_req_gh != gh) {
 795                         if (gh->gh_state != state)
 796                                 gh->gh_state = LM_ST_UNLOCKED;
 797                         goto out;
 798                 }
 799         }
 800
 801         if (new_gh) {
 802                 list_add_tail(&new_gh->gh_list, &gl->gl_waiters2);
 803                 new_gh = NULL;
 804         } else {
 805                 spin_unlock(&gl->gl_spin);
 806
 807                 new_gh = gfs2_holder_get(gl, state,
 808                                          LM_FLAG_TRY | GL_NEVER_RECURSE,
 809                                          GFP_KERNEL | __GFP_NOFAIL),
 810                 set_bit(HIF_DEMOTE, &new_gh->gh_iflags);
 811                 set_bit(HIF_DEALLOC, &new_gh->gh_iflags);
 812
 813                 goto restart;
 814         }
 815
 816  out:
 817         spin_unlock(&gl->gl_spin);
 818
 819         if (new_gh)
 820                 gfs2_holder_put(new_gh);
 821 }
 822
 823 /**
 824  * state_change - record that the glock is now in a different state
 825  * @gl: the glock
 826  * @new_state the new state
 827  *
 828  */
 829
 830 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 831 {
 832         int held1, held2;
 833
 834         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 835         held2 = (new_state != LM_ST_UNLOCKED);
 836
 837         if (held1 != held2) {
 838                 if (held2)
 839                         gfs2_glock_hold(gl);
 840                 else
 841                         gfs2_glock_put(gl);
 842         }
 843
 844         gl->gl_state = new_state;
 845 }
 846
 847 /**
 848  * xmote_bh - Called after the lock module is done acquiring a lock
 849  * @gl: The glock in question
 850  * @ret: the int returned from the lock module
 851  *
 852  */
 853
 854 static void xmote_bh(struct gfs2_glock *gl, unsigned int ret)
 855 {
 856         struct gfs2_sbd *sdp = gl->gl_sbd;
 857         struct gfs2_glock_operations *glops = gl->gl_ops;
 858         struct gfs2_holder *gh = gl->gl_req_gh;
 859         int prev_state = gl->gl_state;
 860         int op_done = 1;
 861
 862         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 863         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 864         gfs2_assert_warn(sdp, !(ret & LM_OUT_ASYNC));
 865
 866         state_change(gl, ret & LM_OUT_ST_MASK);
 867
 868         if (prev_state != LM_ST_UNLOCKED && !(ret & LM_OUT_CACHEABLE)) {
 869                 if (glops->go_inval)
 870                         glops->go_inval(gl, DIO_METADATA | DIO_DATA);
 871         } else if (gl->gl_state == LM_ST_DEFERRED) {
 872                 /* We might not want to do this here.
 873                    Look at moving to the inode glops. */
 874                 if (glops->go_inval)
 875                         glops->go_inval(gl, DIO_DATA);
 876         }
 877
 878         /*  Deal with each possible exit condition  */
 879
 880         if (!gh)
 881                 gl->gl_stamp = jiffies;
 882
 883         else if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
 884                 spin_lock(&gl->gl_spin);
 885                 list_del_init(&gh->gh_list);
 886                 gh->gh_error = -EIO;
 887                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 888                         do_unrecurse(gh);
 889                 spin_unlock(&gl->gl_spin);
 890
 891         } else if (test_bit(HIF_DEMOTE, &gh->gh_iflags)) {
 892                 spin_lock(&gl->gl_spin);
 893                 list_del_init(&gh->gh_list);
 894                 if (gl->gl_state == gh->gh_state ||
 895                     gl->gl_state == LM_ST_UNLOCKED)
 896                         gh->gh_error = 0;
 897                 else {
 898                         if (gfs2_assert_warn(sdp, gh->gh_flags &
 899                                         (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) == -1)
 900                                 fs_warn(sdp, "ret = 0x%.8X\n", ret);
 901                         gh->gh_error = GLR_TRYFAILED;
 902                 }
 903                 spin_unlock(&gl->gl_spin);
 904
 905                 if (ret & LM_OUT_CANCELED)
 906                         handle_callback(gl, LM_ST_UNLOCKED); /* Lame */
 907
 908         } else if (ret & LM_OUT_CANCELED) {
 909                 spin_lock(&gl->gl_spin);
 910                 list_del_init(&gh->gh_list);
 911                 gh->gh_error = GLR_CANCELED;
 912                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 913                         do_unrecurse(gh);
 914                 spin_unlock(&gl->gl_spin);
 915
 916         } else if (relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 917                 spin_lock(&gl->gl_spin);
 918                 list_move_tail(&gh->gh_list, &gl->gl_holders);
 919                 gh->gh_error = 0;
 920                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 921                 spin_unlock(&gl->gl_spin);
 922
 923                 set_bit(HIF_FIRST, &gh->gh_iflags);
 924
 925                 op_done = 0;
 926
 927         } else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 928                 spin_lock(&gl->gl_spin);
 929                 list_del_init(&gh->gh_list);
 930                 gh->gh_error = GLR_TRYFAILED;
 931                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
 932                         do_unrecurse(gh);
 933                 spin_unlock(&gl->gl_spin);
 934
 935         } else {
 936                 if (gfs2_assert_withdraw(sdp, 0) == -1)
 937                         fs_err(sdp, "ret = 0x%.8X\n", ret);
 938         }
 939
 940         if (glops->go_xmote_bh)
 941                 glops->go_xmote_bh(gl);
 942
 943         if (op_done) {
 944                 spin_lock(&gl->gl_spin);
 945                 gl->gl_req_gh = NULL;
 946                 gl->gl_req_bh = NULL;
 947                 clear_bit(GLF_LOCK, &gl->gl_flags);
 948                 run_queue(gl);
 949                 spin_unlock(&gl->gl_spin);
 950         }
 951
 952         gfs2_glock_put(gl);
 953
 954         if (gh) {
 955                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 956                         gfs2_holder_put(gh);
 957                 else
 958                         complete(&gh->gh_wait);
 959         }
 960 }
 961
 962 /**
 963  * gfs2_glock_xmote_th - Call into the lock module to acquire or change a glock
 964  * @gl: The glock in question
 965  * @state: the requested state
 966  * @flags: modifier flags to the lock call
 967  *
 968  */
 969
 970 void gfs2_glock_xmote_th(struct gfs2_glock *gl, unsigned int state, int flags)
 971 {
 972         struct gfs2_sbd *sdp = gl->gl_sbd;
 973         struct gfs2_glock_operations *glops = gl->gl_ops;
 974         int lck_flags = flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB |
 975                                  LM_FLAG_NOEXP | LM_FLAG_ANY |
 976                                  LM_FLAG_PRIORITY);
 977         unsigned int lck_ret;
 978
 979         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 980         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 981         gfs2_assert_warn(sdp, state != LM_ST_UNLOCKED);
 982         gfs2_assert_warn(sdp, state != gl->gl_state);
 983
 984         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 985                 if (glops->go_sync)
 986                         glops->go_sync(gl,
 987                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
 988         }
 989
 990         gfs2_glock_hold(gl);
 991         gl->gl_req_bh = xmote_bh;
 992
 993         lck_ret = gfs2_lm_lock(sdp, gl->gl_lock, gl->gl_state, state,
 994                                lck_flags);
 995
 996         if (gfs2_assert_withdraw(sdp, !(lck_ret & LM_OUT_ERROR)))
 997                 return;
 998
 999         if (lck_ret & LM_OUT_ASYNC)
1000                 gfs2_assert_warn(sdp, lck_ret == LM_OUT_ASYNC);
1001         else
1002                 xmote_bh(gl, lck_ret);
1003 }
1004
1005 /**
1006  * drop_bh - Called after a lock module unlock completes
1007  * @gl: the glock
1008  * @ret: the return status
1009  *
1010  * Doesn't wake up the process waiting on the struct gfs2_holder (if any)
1011  * Doesn't drop the reference on the glock the top half took out
1012  *
1013  */
1014
1015 static void drop_bh(struct gfs2_glock *gl, unsigned int ret)
1016 {
1017         struct gfs2_sbd *sdp = gl->gl_sbd;
1018         struct gfs2_glock_operations *glops = gl->gl_ops;
1019         struct gfs2_holder *gh = gl->gl_req_gh;
1020
1021         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1022
1023         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1024         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1025         gfs2_assert_warn(sdp, !ret);
1026
1027         state_change(gl, LM_ST_UNLOCKED);
1028
1029         if (glops->go_inval)
1030                 glops->go_inval(gl, DIO_METADATA | DIO_DATA);
1031
1032         if (gh) {
1033                 spin_lock(&gl->gl_spin);
1034                 list_del_init(&gh->gh_list);
1035                 gh->gh_error = 0;
1036                 spin_unlock(&gl->gl_spin);
1037         }
1038
1039         if (glops->go_drop_bh)
1040                 glops->go_drop_bh(gl);
1041
1042         spin_lock(&gl->gl_spin);
1043         gl->gl_req_gh = NULL;
1044         gl->gl_req_bh = NULL;
1045         clear_bit(GLF_LOCK, &gl->gl_flags);
1046         run_queue(gl);
1047         spin_unlock(&gl->gl_spin);
1048
1049         gfs2_glock_put(gl);
1050
1051         if (gh) {
1052                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
1053                         gfs2_holder_put(gh);
1054                 else
1055                         complete(&gh->gh_wait);
1056         }
1057 }
1058
1059 /**
1060  * gfs2_glock_drop_th - call into the lock module to unlock a lock
1061  * @gl: the glock
1062  *
1063  */
1064
1065 void gfs2_glock_drop_th(struct gfs2_glock *gl)
1066 {
1067         struct gfs2_sbd *sdp = gl->gl_sbd;
1068         struct gfs2_glock_operations *glops = gl->gl_ops;
1069         unsigned int ret;
1070
1071         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1072         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1073         gfs2_assert_warn(sdp, gl->gl_state != LM_ST_UNLOCKED);
1074
1075         if (gl->gl_state == LM_ST_EXCLUSIVE) {
1076                 if (glops->go_sync)
1077                         glops->go_sync(gl,
1078                                        DIO_METADATA | DIO_DATA | DIO_RELEASE);
1079         }
1080
1081         gfs2_glock_hold(gl);
1082         gl->gl_req_bh = drop_bh;
1083
1084         ret = gfs2_lm_unlock(sdp, gl->gl_lock, gl->gl_state);
1085
1086         if (gfs2_assert_withdraw(sdp, !(ret & LM_OUT_ERROR)))
1087                 return;
1088
1089         if (!ret)
1090                 drop_bh(gl, ret);
1091         else
1092                 gfs2_assert_warn(sdp, ret == LM_OUT_ASYNC);
1093 }
1094
1095 /**
1096  * do_cancels - cancel requests for locks stuck waiting on an expire flag
1097  * @gh: the LM_FLAG_PRIORITY holder waiting to acquire the lock
1098  *
1099  * Don't cancel GL_NOCANCEL requests.
1100  */
1101
1102 static void do_cancels(struct gfs2_holder *gh)
1103 {
1104         struct gfs2_glock *gl = gh->gh_gl;
1105
1106         spin_lock(&gl->gl_spin);
1107
1108         while (gl->gl_req_gh != gh &&
1109                !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1110                !list_empty(&gh->gh_list)) {
1111                 if (gl->gl_req_bh &&
1112                     !(gl->gl_req_gh &&
1113                       (gl->gl_req_gh->gh_flags & GL_NOCANCEL))) {
1114                         spin_unlock(&gl->gl_spin);
1115                         gfs2_lm_cancel(gl->gl_sbd, gl->gl_lock);
1116                         msleep(100);
1117                         spin_lock(&gl->gl_spin);
1118                 } else {
1119                         spin_unlock(&gl->gl_spin);
1120                         msleep(100);
1121                         spin_lock(&gl->gl_spin);
1122                 }
1123         }
1124
1125         spin_unlock(&gl->gl_spin);
1126 }
1127
1128 /**
1129  * glock_wait_internal - wait on a glock acquisition
1130  * @gh: the glock holder
1131  *
1132  * Returns: 0 on success
1133  */
1134
1135 static int glock_wait_internal(struct gfs2_holder *gh)
1136 {
1137         struct gfs2_glock *gl = gh->gh_gl;
1138         struct gfs2_sbd *sdp = gl->gl_sbd;
1139         struct gfs2_glock_operations *glops = gl->gl_ops;
1140
1141         if (test_bit(HIF_ABORTED, &gh->gh_iflags))
1142                 return -EIO;
1143
1144         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1145                 spin_lock(&gl->gl_spin);
1146                 if (gl->gl_req_gh != gh &&
1147                     !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1148                     !list_empty(&gh->gh_list)) {
1149                         list_del_init(&gh->gh_list);
1150                         gh->gh_error = GLR_TRYFAILED;
1151                         if (test_bit(HIF_RECURSE, &gh->gh_iflags))
1152                                 do_unrecurse(gh);
1153                         run_queue(gl);
1154                         spin_unlock(&gl->gl_spin);
1155                         return gh->gh_error;
1156                 }
1157                 spin_unlock(&gl->gl_spin);
1158         }
1159
1160         if (gh->gh_flags & LM_FLAG_PRIORITY)
1161                 do_cancels(gh);
1162
1163         wait_for_completion(&gh->gh_wait);
1164
1165         if (gh->gh_error)
1166                 return gh->gh_error;
1167
1168         gfs2_assert_withdraw(sdp, test_bit(HIF_HOLDER, &gh->gh_iflags));
1169         gfs2_assert_withdraw(sdp, relaxed_state_ok(gl->gl_state,
1170                                                    gh->gh_state,
1171                                                    gh->gh_flags));
1172
1173         if (test_bit(HIF_FIRST, &gh->gh_iflags)) {
1174                 gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1175
1176                 if (glops->go_lock) {
1177                         gh->gh_error = glops->go_lock(gh);
1178                         if (gh->gh_error) {
1179                                 spin_lock(&gl->gl_spin);
1180                                 list_del_init(&gh->gh_list);
1181                                 if (test_and_clear_bit(HIF_RECURSE,
1182                                                        &gh->gh_iflags))
1183                                         do_unrecurse(gh);
1184                                 spin_unlock(&gl->gl_spin);
1185                         }
1186                 }
1187
1188                 spin_lock(&gl->gl_spin);
1189                 gl->gl_req_gh = NULL;
1190                 gl->gl_req_bh = NULL;
1191                 clear_bit(GLF_LOCK, &gl->gl_flags);
1192                 if (test_bit(HIF_RECURSE, &gh->gh_iflags))
1193                         handle_recurse(gh);
1194                 run_queue(gl);
1195                 spin_unlock(&gl->gl_spin);
1196         }
1197
1198         return gh->gh_error;
1199 }
1200
1201 static inline struct gfs2_holder *
1202 find_holder_by_owner(struct list_head *head, struct task_struct *owner)
1203 {
1204         struct gfs2_holder *gh;
1205
1206         list_for_each_entry(gh, head, gh_list) {
1207                 if (gh->gh_owner == owner)
1208                         return gh;
1209         }
1210
1211         return NULL;
1212 }
1213
1214 /**
1215  * recurse_check -
1216  *
1217  * Make sure the new holder is compatible with the pre-existing one.
1218  *
1219  */
1220
1221 static int recurse_check(struct gfs2_holder *existing, struct gfs2_holder *new,
1222                          unsigned int state)
1223 {
1224         struct gfs2_sbd *sdp = existing->gh_gl->gl_sbd;
1225
1226         if (gfs2_assert_warn(sdp, (new->gh_flags & LM_FLAG_ANY) ||
1227                                   !(existing->gh_flags & LM_FLAG_ANY)))
1228                 goto fail;
1229
1230         if (gfs2_assert_warn(sdp, (existing->gh_flags & GL_LOCAL_EXCL) ||
1231                                   !(new->gh_flags & GL_LOCAL_EXCL)))
1232                 goto fail;
1233
1234         if (gfs2_assert_warn(sdp, relaxed_state_ok(state, new->gh_state,
1235                                                    new->gh_flags)))
1236                 goto fail;
1237
1238         return 0;
1239
1240  fail:
1241         set_bit(HIF_ABORTED, &new->gh_iflags);
1242         return -EINVAL;
1243 }
1244
1245 /**
1246  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1247  * @gh: the holder structure to add
1248  *
1249  */
1250
1251 static void add_to_queue(struct gfs2_holder *gh)
1252 {
1253         struct gfs2_glock *gl = gh->gh_gl;
1254         struct gfs2_holder *existing;
1255
1256         if (!gh->gh_owner)
1257                 goto out;
1258
1259         existing = find_holder_by_owner(&gl->gl_holders, gh->gh_owner);
1260         if (existing) {
1261                 if (recurse_check(existing, gh, gl->gl_state))
1262                         return;
1263
1264                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1265                 set_bit(HIF_HOLDER, &gh->gh_iflags);
1266
1267                 gh->gh_error = 0;
1268                 complete(&gh->gh_wait);
1269
1270                 return;
1271         }
1272
1273         existing = find_holder_by_owner(&gl->gl_waiters3, gh->gh_owner);
1274         if (existing) {
1275                 if (recurse_check(existing, gh, existing->gh_state))
1276                         return;
1277
1278                 set_bit(HIF_RECURSE, &gh->gh_iflags);
1279                 set_bit(HIF_RECURSE, &existing->gh_iflags);
1280
1281                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1282
1283                 return;
1284         }
1285
1286  out:
1287         if (gh->gh_flags & LM_FLAG_PRIORITY)
1288                 list_add(&gh->gh_list, &gl->gl_waiters3);
1289         else
1290                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1291 }
1292
1293 /**
1294  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1295  * @gh: the holder structure
1296  *
1297  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1298  *
1299  * Returns: 0, GLR_TRYFAILED, or errno on failure
1300  */
1301
1302 int gfs2_glock_nq(struct gfs2_holder *gh)
1303 {
1304         struct gfs2_glock *gl = gh->gh_gl;
1305         struct gfs2_sbd *sdp = gl->gl_sbd;
1306         int error = 0;
1307
1308  restart:
1309         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
1310                 set_bit(HIF_ABORTED, &gh->gh_iflags);
1311                 return -EIO;
1312         }
1313
1314         set_bit(HIF_PROMOTE, &gh->gh_iflags);
1315
1316         spin_lock(&gl->gl_spin);
1317         add_to_queue(gh);
1318         run_queue(gl);
1319         spin_unlock(&gl->gl_spin);
1320
1321         if (!(gh->gh_flags & GL_ASYNC)) {
1322                 error = glock_wait_internal(gh);
1323                 if (error == GLR_CANCELED) {
1324                         msleep(1000);
1325                         goto restart;
1326                 }
1327         }
1328
1329         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1330
1331         return error;
1332 }
1333
1334 /**
1335  * gfs2_glock_poll - poll to see if an async request has been completed
1336  * @gh: the holder
1337  *
1338  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1339  */
1340
1341 int gfs2_glock_poll(struct gfs2_holder *gh)
1342 {
1343         struct gfs2_glock *gl = gh->gh_gl;
1344         int ready = 0;
1345
1346         spin_lock(&gl->gl_spin);
1347
1348         if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1349                 ready = 1;
1350         else if (list_empty(&gh->gh_list)) {
1351                 if (gh->gh_error == GLR_CANCELED) {
1352                         spin_unlock(&gl->gl_spin);
1353                         msleep(1000);
1354                         if (gfs2_glock_nq(gh))
1355                                 return 1;
1356                         return 0;
1357                 } else
1358                         ready = 1;
1359         }
1360
1361         spin_unlock(&gl->gl_spin);
1362
1363         return ready;
1364 }
1365
1366 /**
1367  * gfs2_glock_wait - wait for a lock acquisition that ended in a GLR_ASYNC
1368  * @gh: the holder structure
1369  *
1370  * Returns: 0, GLR_TRYFAILED, or errno on failure
1371  */
1372
1373 int gfs2_glock_wait(struct gfs2_holder *gh)
1374 {
1375         int error;
1376
1377         error = glock_wait_internal(gh);
1378         if (error == GLR_CANCELED) {
1379                 msleep(1000);
1380                 gh->gh_flags &= ~GL_ASYNC;
1381                 error = gfs2_glock_nq(gh);
1382         }
1383
1384         return error;
1385 }
1386
1387 /**
1388  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1389  * @gh: the glock holder
1390  *
1391  */
1392
1393 void gfs2_glock_dq(struct gfs2_holder *gh)
1394 {
1395         struct gfs2_glock *gl = gh->gh_gl;
1396         struct gfs2_glock_operations *glops = gl->gl_ops;
1397
1398         if (gh->gh_flags & GL_SYNC)
1399                 set_bit(GLF_SYNC, &gl->gl_flags);
1400
1401         if (gh->gh_flags & GL_NOCACHE)
1402                 handle_callback(gl, LM_ST_UNLOCKED);
1403
1404         gfs2_glmutex_lock(gl);
1405
1406         spin_lock(&gl->gl_spin);
1407         list_del_init(&gh->gh_list);
1408
1409         if (list_empty(&gl->gl_holders)) {
1410                 spin_unlock(&gl->gl_spin);
1411
1412                 if (glops->go_unlock)
1413                         glops->go_unlock(gh);
1414
1415                 if (test_bit(GLF_SYNC, &gl->gl_flags)) {
1416                         if (glops->go_sync)
1417                                 glops->go_sync(gl, DIO_METADATA | DIO_DATA);
1418                 }
1419
1420                 gl->gl_stamp = jiffies;
1421
1422                 spin_lock(&gl->gl_spin);
1423         }
1424
1425         clear_bit(GLF_LOCK, &gl->gl_flags);
1426         run_queue(gl);
1427         spin_unlock(&gl->gl_spin);
1428 }
1429
1430 /**
1431  * gfs2_glock_prefetch - Try to prefetch a glock
1432  * @gl: the glock
1433  * @state: the state to prefetch in
1434  * @flags: flags passed to go_xmote_th()
1435  *
1436  */
1437
1438 void gfs2_glock_prefetch(struct gfs2_glock *gl, unsigned int state, int flags)
1439 {
1440         struct gfs2_glock_operations *glops = gl->gl_ops;
1441
1442         spin_lock(&gl->gl_spin);
1443
1444         if (test_bit(GLF_LOCK, &gl->gl_flags) ||
1445             !list_empty(&gl->gl_holders) ||
1446             !list_empty(&gl->gl_waiters1) ||
1447             !list_empty(&gl->gl_waiters2) ||
1448             !list_empty(&gl->gl_waiters3) ||
1449             relaxed_state_ok(gl->gl_state, state, flags)) {
1450                 spin_unlock(&gl->gl_spin);
1451                 return;
1452         }
1453
1454         set_bit(GLF_PREFETCH, &gl->gl_flags);
1455         set_bit(GLF_LOCK, &gl->gl_flags);
1456         spin_unlock(&gl->gl_spin);
1457
1458         glops->go_xmote_th(gl, state, flags);
1459 }
1460
1461 /**
1462  * gfs2_glock_force_drop - Force a glock to be uncached
1463  * @gl: the glock
1464  *
1465  */
1466
1467 void gfs2_glock_force_drop(struct gfs2_glock *gl)
1468 {
1469         struct gfs2_holder gh;
1470
1471         gfs2_holder_init(gl, LM_ST_UNLOCKED, GL_NEVER_RECURSE, &gh);
1472         set_bit(HIF_DEMOTE, &gh.gh_iflags);
1473
1474         spin_lock(&gl->gl_spin);
1475         list_add_tail(&gh.gh_list, &gl->gl_waiters2);
1476         run_queue(gl);
1477         spin_unlock(&gl->gl_spin);
1478
1479         wait_for_completion(&gh.gh_wait);
1480         gfs2_holder_uninit(&gh);
1481 }
1482
1483 static void greedy_work(void *data)
1484 {
1485         struct greedy *gr = (struct greedy *)data;
1486         struct gfs2_holder *gh = &gr->gr_gh;
1487         struct gfs2_glock *gl = gh->gh_gl;
1488         struct gfs2_glock_operations *glops = gl->gl_ops;
1489
1490         clear_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1491
1492         if (glops->go_greedy)
1493                 glops->go_greedy(gl);
1494
1495         spin_lock(&gl->gl_spin);
1496
1497         if (list_empty(&gl->gl_waiters2)) {
1498                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1499                 spin_unlock(&gl->gl_spin);
1500                 gfs2_holder_uninit(gh);
1501                 kfree(gr);
1502         } else {
1503                 gfs2_glock_hold(gl);
1504                 list_add_tail(&gh->gh_list, &gl->gl_waiters2);
1505                 run_queue(gl);
1506                 spin_unlock(&gl->gl_spin);
1507                 gfs2_glock_put(gl);
1508         }
1509 }
1510
1511 /**
1512  * gfs2_glock_be_greedy -
1513  * @gl:
1514  * @time:
1515  *
1516  * Returns: 0 if go_greedy will be called, 1 otherwise
1517  */
1518
1519 int gfs2_glock_be_greedy(struct gfs2_glock *gl, unsigned int time)
1520 {
1521         struct greedy *gr;
1522         struct gfs2_holder *gh;
1523
1524         if (!time ||
1525             gl->gl_sbd->sd_args.ar_localcaching ||
1526             test_and_set_bit(GLF_GREEDY, &gl->gl_flags))
1527                 return 1;
1528
1529         gr = kmalloc(sizeof(struct greedy), GFP_KERNEL);
1530         if (!gr) {
1531                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1532                 return 1;
1533         }
1534         gh = &gr->gr_gh;
1535
1536         gfs2_holder_init(gl, 0, GL_NEVER_RECURSE, gh);
1537         set_bit(HIF_GREEDY, &gh->gh_iflags);
1538         INIT_WORK(&gr->gr_work, greedy_work, gr);
1539
1540         set_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1541         schedule_delayed_work(&gr->gr_work, time);
1542
1543         return 0;
1544 }
1545
1546 /**
1547  * gfs2_glock_nq_init - intialize a holder and enqueue it on a glock
1548  * @gl: the glock
1549  * @state: the state we're requesting
1550  * @flags: the modifier flags
1551  * @gh: the holder structure
1552  *
1553  * Returns: 0, GLR_*, or errno
1554  */
1555
1556 int gfs2_glock_nq_init(struct gfs2_glock *gl, unsigned int state, int flags,
1557                        struct gfs2_holder *gh)
1558 {
1559         int error;
1560
1561         gfs2_holder_init(gl, state, flags, gh);
1562
1563         error = gfs2_glock_nq(gh);
1564         if (error)
1565                 gfs2_holder_uninit(gh);
1566
1567         return error;
1568 }
1569
1570 /**
1571  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1572  * @gh: the holder structure
1573  *
1574  */
1575
1576 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1577 {
1578         gfs2_glock_dq(gh);
1579         gfs2_holder_uninit(gh);
1580 }
1581
1582 /**
1583  * gfs2_glock_nq_num - acquire a glock based on lock number
1584  * @sdp: the filesystem
1585  * @number: the lock number
1586  * @glops: the glock operations for the type of glock
1587  * @state: the state to acquire the glock in
1588  * @flags: modifier flags for the aquisition
1589  * @gh: the struct gfs2_holder
1590  *
1591  * Returns: errno
1592  */
1593
1594 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, uint64_t number,
1595                       struct gfs2_glock_operations *glops, unsigned int state,
1596                       int flags, struct gfs2_holder *gh)
1597 {
1598         struct gfs2_glock *gl;
1599         int error;
1600
1601         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1602         if (!error) {
1603                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1604                 gfs2_glock_put(gl);
1605         }
1606
1607         return error;
1608 }
1609
1610 /**
1611  * glock_compare - Compare two struct gfs2_glock structures for sorting
1612  * @arg_a: the first structure
1613  * @arg_b: the second structure
1614  *
1615  */
1616
1617 static int glock_compare(const void *arg_a, const void *arg_b)
1618 {
1619         struct gfs2_holder *gh_a = *(struct gfs2_holder **)arg_a;
1620         struct gfs2_holder *gh_b = *(struct gfs2_holder **)arg_b;
1621         struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1622         struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1623         int ret = 0;
1624
1625         if (a->ln_number > b->ln_number)
1626                 ret = 1;
1627         else if (a->ln_number < b->ln_number)
1628                 ret = -1;
1629         else {
1630                 if (gh_a->gh_state == LM_ST_SHARED &&
1631                     gh_b->gh_state == LM_ST_EXCLUSIVE)
1632                         ret = 1;
1633                 else if (!(gh_a->gh_flags & GL_LOCAL_EXCL) &&
1634                          (gh_b->gh_flags & GL_LOCAL_EXCL))
1635                         ret = 1;
1636         }
1637
1638         return ret;
1639 }
1640
1641 /**
1642  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1643  * @num_gh: the number of structures
1644  * @ghs: an array of struct gfs2_holder structures
1645  *
1646  * Returns: 0 on success (all glocks acquired),
1647  *          errno on failure (no glocks acquired)
1648  */
1649
1650 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1651                      struct gfs2_holder **p)
1652 {
1653         unsigned int x;
1654         int error = 0;
1655
1656         for (x = 0; x < num_gh; x++)
1657                 p[x] = &ghs[x];
1658
1659         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1660
1661         for (x = 0; x < num_gh; x++) {
1662                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1663
1664                 error = gfs2_glock_nq(p[x]);
1665                 if (error) {
1666                         while (x--)
1667                                 gfs2_glock_dq(p[x]);
1668                         break;
1669                 }
1670         }
1671
1672         return error;
1673 }
1674
1675 /**
1676  * gfs2_glock_nq_m - acquire multiple glocks
1677  * @num_gh: the number of structures
1678  * @ghs: an array of struct gfs2_holder structures
1679  *
1680  * Figure out how big an impact this function has.  Either:
1681  * 1) Replace this code with code that calls gfs2_glock_prefetch()
1682  * 2) Forget async stuff and just call nq_m_sync()
1683  * 3) Leave it like it is
1684  *
1685  * Returns: 0 on success (all glocks acquired),
1686  *          errno on failure (no glocks acquired)
1687  */
1688
1689 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1690 {
1691         int *e;
1692         unsigned int x;
1693         int borked = 0, serious = 0;
1694         int error = 0;
1695
1696         if (!num_gh)
1697                 return 0;
1698
1699         if (num_gh == 1) {
1700                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1701                 return gfs2_glock_nq(ghs);
1702         }
1703
1704         e = kcalloc(num_gh, sizeof(struct gfs2_holder *), GFP_KERNEL);
1705         if (!e)
1706                 return -ENOMEM;
1707
1708         for (x = 0; x < num_gh; x++) {
1709                 ghs[x].gh_flags |= LM_FLAG_TRY | GL_ASYNC;
1710                 error = gfs2_glock_nq(&ghs[x]);
1711                 if (error) {
1712                         borked = 1;
1713                         serious = error;
1714                         num_gh = x;
1715                         break;
1716                 }
1717         }
1718
1719         for (x = 0; x < num_gh; x++) {
1720                 error = e[x] = glock_wait_internal(&ghs[x]);
1721                 if (error) {
1722                         borked = 1;
1723                         if (error != GLR_TRYFAILED && error != GLR_CANCELED)
1724                                 serious = error;
1725                 }
1726         }
1727
1728         if (!borked) {
1729                 kfree(e);
1730                 return 0;
1731         }
1732
1733         for (x = 0; x < num_gh; x++)
1734                 if (!e[x])
1735                         gfs2_glock_dq(&ghs[x]);
1736
1737         if (serious)
1738                 error = serious;
1739         else {
1740                 for (x = 0; x < num_gh; x++)
1741                         gfs2_holder_reinit(ghs[x].gh_state, ghs[x].gh_flags,
1742                                           &ghs[x]);
1743                 error = nq_m_sync(num_gh, ghs, (struct gfs2_holder **)e);
1744         }
1745
1746         kfree(e);
1747
1748         return error;
1749 }
1750
1751 /**
1752  * gfs2_glock_dq_m - release multiple glocks
1753  * @num_gh: the number of structures
1754  * @ghs: an array of struct gfs2_holder structures
1755  *
1756  */
1757
1758 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1759 {
1760         unsigned int x;
1761
1762         for (x = 0; x < num_gh; x++)
1763                 gfs2_glock_dq(&ghs[x]);
1764 }
1765
1766 /**
1767  * gfs2_glock_dq_uninit_m - release multiple glocks
1768  * @num_gh: the number of structures
1769  * @ghs: an array of struct gfs2_holder structures
1770  *
1771  */
1772
1773 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1774 {
1775         unsigned int x;
1776
1777         for (x = 0; x < num_gh; x++)
1778                 gfs2_glock_dq_uninit(&ghs[x]);
1779 }
1780
1781 /**
1782  * gfs2_glock_prefetch_num - prefetch a glock based on lock number
1783  * @sdp: the filesystem
1784  * @number: the lock number
1785  * @glops: the glock operations for the type of glock
1786  * @state: the state to acquire the glock in
1787  * @flags: modifier flags for the aquisition
1788  *
1789  * Returns: errno
1790  */
1791
1792 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, uint64_t number,
1793                              struct gfs2_glock_operations *glops,
1794                              unsigned int state, int flags)
1795 {
1796         struct gfs2_glock *gl;
1797         int error;
1798
1799         if (atomic_read(&sdp->sd_reclaim_count) <
1800             gfs2_tune_get(sdp, gt_reclaim_limit)) {
1801                 error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1802                 if (!error) {
1803                         gfs2_glock_prefetch(gl, state, flags);
1804                         gfs2_glock_put(gl);
1805                 }
1806         }
1807 }
1808
1809 /**
1810  * gfs2_lvb_hold - attach a LVB from a glock
1811  * @gl: The glock in question
1812  *
1813  */
1814
1815 int gfs2_lvb_hold(struct gfs2_glock *gl)
1816 {
1817         int error;
1818
1819         gfs2_glmutex_lock(gl);
1820
1821         if (!atomic_read(&gl->gl_lvb_count)) {
1822                 error = gfs2_lm_hold_lvb(gl->gl_sbd, gl->gl_lock, &gl->gl_lvb);
1823                 if (error) {
1824                         gfs2_glmutex_unlock(gl);
1825                         return error;
1826                 }
1827                 gfs2_glock_hold(gl);
1828         }
1829         atomic_inc(&gl->gl_lvb_count);
1830
1831         gfs2_glmutex_unlock(gl);
1832
1833         return 0;
1834 }
1835
1836 /**
1837  * gfs2_lvb_unhold - detach a LVB from a glock
1838  * @gl: The glock in question
1839  *
1840  */
1841
1842 void gfs2_lvb_unhold(struct gfs2_glock *gl)
1843 {
1844         gfs2_glock_hold(gl);
1845         gfs2_glmutex_lock(gl);
1846
1847         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count) > 0);
1848         if (atomic_dec_and_test(&gl->gl_lvb_count)) {
1849                 gfs2_lm_unhold_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1850                 gl->gl_lvb = NULL;
1851                 gfs2_glock_put(gl);
1852         }
1853
1854         gfs2_glmutex_unlock(gl);
1855         gfs2_glock_put(gl);
1856 }
1857
1858 void gfs2_lvb_sync(struct gfs2_glock *gl)
1859 {
1860         gfs2_glmutex_lock(gl);
1861
1862         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count));
1863         if (!gfs2_assert_warn(gl->gl_sbd, gfs2_glock_is_held_excl(gl)))
1864                 gfs2_lm_sync_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1865
1866         gfs2_glmutex_unlock(gl);
1867 }
1868
1869 static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
1870                         unsigned int state)
1871 {
1872         struct gfs2_glock *gl;
1873
1874         gl = gfs2_glock_find(sdp, name);
1875         if (!gl)
1876                 return;
1877
1878         if (gl->gl_ops->go_callback)
1879                 gl->gl_ops->go_callback(gl, state);
1880         handle_callback(gl, state);
1881
1882         spin_lock(&gl->gl_spin);
1883         run_queue(gl);
1884         spin_unlock(&gl->gl_spin);
1885
1886         gfs2_glock_put(gl);
1887 }
1888
1889 /**
1890  * gfs2_glock_cb - Callback used by locking module
1891  * @fsdata: Pointer to the superblock
1892  * @type: Type of callback
1893  * @data: Type dependent data pointer
1894  *
1895  * Called by the locking module when it wants to tell us something.
1896  * Either we need to drop a lock, one of our ASYNC requests completed, or
1897  * a journal from another client needs to be recovered.
1898  */
1899
1900 void gfs2_glock_cb(lm_fsdata_t *fsdata, unsigned int type, void *data)
1901 {
1902         struct gfs2_sbd *sdp = (struct gfs2_sbd *)fsdata;
1903
1904         switch (type) {
1905         case LM_CB_NEED_E:
1906                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_UNLOCKED);
1907                 return;
1908
1909         case LM_CB_NEED_D:
1910                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_DEFERRED);
1911                 return;
1912
1913         case LM_CB_NEED_S:
1914                 blocking_cb(sdp, (struct lm_lockname *)data, LM_ST_SHARED);
1915                 return;
1916
1917         case LM_CB_ASYNC: {
1918                 struct lm_async_cb *async = (struct lm_async_cb *)data;
1919                 struct gfs2_glock *gl;
1920
1921                 gl = gfs2_glock_find(sdp, &async->lc_name);
1922                 if (gfs2_assert_warn(sdp, gl))
1923                         return;
1924                 if (!gfs2_assert_warn(sdp, gl->gl_req_bh))
1925                         gl->gl_req_bh(gl, async->lc_ret);
1926                 gfs2_glock_put(gl);
1927
1928                 return;
1929         }
1930
1931         case LM_CB_NEED_RECOVERY:
1932                 gfs2_jdesc_make_dirty(sdp, *(unsigned int *)data);
1933                 if (sdp->sd_recoverd_process)
1934                         wake_up_process(sdp->sd_recoverd_process);
1935                 return;
1936
1937         case LM_CB_DROPLOCKS:
1938                 gfs2_gl_hash_clear(sdp, NO_WAIT);
1939                 gfs2_quota_scan(sdp);
1940                 return;
1941
1942         default:
1943                 gfs2_assert_warn(sdp, 0);
1944                 return;
1945         }
1946 }
1947
1948 /**
1949  * gfs2_try_toss_inode - try to remove a particular inode struct from cache
1950  * sdp: the filesystem
1951  * inum: the inode number
1952  *
1953  */
1954
1955 void gfs2_try_toss_inode(struct gfs2_sbd *sdp, struct gfs2_inum *inum)
1956 {
1957         struct gfs2_glock *gl;
1958         struct gfs2_inode *ip;
1959         int error;
1960
1961         error = gfs2_glock_get(sdp, inum->no_addr, &gfs2_inode_glops,
1962                                NO_CREATE, &gl);
1963         if (error || !gl)
1964                 return;
1965
1966         if (!gfs2_glmutex_trylock(gl))
1967                 goto out;
1968
1969         ip = gl->gl_object;
1970         if (!ip)
1971                 goto out_unlock;
1972
1973         if (atomic_read(&ip->i_count))
1974                 goto out_unlock;
1975
1976         gfs2_inode_destroy(ip);
1977
1978  out_unlock:
1979         gfs2_glmutex_unlock(gl);
1980
1981  out:
1982         gfs2_glock_put(gl);
1983 }
1984
1985 /**
1986  * gfs2_iopen_go_callback - Try to kick the inode/vnode associated with an
1987  *                          iopen glock from memory
1988  * @io_gl: the iopen glock
1989  * @state: the state into which the glock should be put
1990  *
1991  */
1992
1993 void gfs2_iopen_go_callback(struct gfs2_glock *io_gl, unsigned int state)
1994 {
1995         struct gfs2_glock *i_gl;
1996
1997         if (state != LM_ST_UNLOCKED)
1998                 return;
1999
2000         spin_lock(&io_gl->gl_spin);
2001         i_gl = io_gl->gl_object;
2002         if (i_gl) {
2003                 gfs2_glock_hold(i_gl);
2004                 spin_unlock(&io_gl->gl_spin);
2005         } else {
2006                 spin_unlock(&io_gl->gl_spin);
2007                 return;
2008         }
2009
2010         if (gfs2_glmutex_trylock(i_gl)) {
2011                 struct gfs2_inode *ip = i_gl->gl_object;
2012                 if (ip) {
2013                         gfs2_try_toss_vnode(ip);
2014                         gfs2_glmutex_unlock(i_gl);
2015                         gfs2_glock_schedule_for_reclaim(i_gl);
2016                         goto out;
2017                 }
2018                 gfs2_glmutex_unlock(i_gl);
2019         }
2020
2021  out:
2022         gfs2_glock_put(i_gl);
2023 }
2024
2025 /**
2026  * demote_ok - Check to see if it's ok to unlock a glock
2027  * @gl: the glock
2028  *
2029  * Returns: 1 if it's ok
2030  */
2031
2032 static int demote_ok(struct gfs2_glock *gl)
2033 {
2034         struct gfs2_sbd *sdp = gl->gl_sbd;
2035         struct gfs2_glock_operations *glops = gl->gl_ops;
2036         int demote = 1;
2037
2038         if (test_bit(GLF_STICKY, &gl->gl_flags))
2039                 demote = 0;
2040         else if (test_bit(GLF_PREFETCH, &gl->gl_flags))
2041                 demote = time_after_eq(jiffies,
2042                                     gl->gl_stamp +
2043                                     gfs2_tune_get(sdp, gt_prefetch_secs) * HZ);
2044         else if (glops->go_demote_ok)
2045                 demote = glops->go_demote_ok(gl);
2046
2047         return demote;
2048 }
2049
2050 /**
2051  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
2052  * @gl: the glock
2053  *
2054  */
2055
2056 void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
2057 {
2058         struct gfs2_sbd *sdp = gl->gl_sbd;
2059
2060         spin_lock(&sdp->sd_reclaim_lock);
2061         if (list_empty(&gl->gl_reclaim)) {
2062                 gfs2_glock_hold(gl);
2063                 list_add(&gl->gl_reclaim, &sdp->sd_reclaim_list);
2064                 atomic_inc(&sdp->sd_reclaim_count);
2065         }
2066         spin_unlock(&sdp->sd_reclaim_lock);
2067
2068         wake_up(&sdp->sd_reclaim_wq);
2069 }
2070
2071 /**
2072  * gfs2_reclaim_glock - process the next glock on the filesystem's reclaim list
2073  * @sdp: the filesystem
2074  *
2075  * Called from gfs2_glockd() glock reclaim daemon, or when promoting a
2076  * different glock and we notice that there are a lot of glocks in the
2077  * reclaim list.
2078  *
2079  */
2080
2081 void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
2082 {
2083         struct gfs2_glock *gl;
2084
2085         spin_lock(&sdp->sd_reclaim_lock);
2086         if (list_empty(&sdp->sd_reclaim_list)) {
2087                 spin_unlock(&sdp->sd_reclaim_lock);
2088                 return;
2089         }
2090         gl = list_entry(sdp->sd_reclaim_list.next,
2091                         struct gfs2_glock, gl_reclaim);
2092         list_del_init(&gl->gl_reclaim);
2093         spin_unlock(&sdp->sd_reclaim_lock);
2094
2095         atomic_dec(&sdp->sd_reclaim_count);
2096         atomic_inc(&sdp->sd_reclaimed);
2097
2098         if (gfs2_glmutex_trylock(gl)) {
2099                 if (gl->gl_ops == &gfs2_inode_glops) {
2100                         struct gfs2_inode *ip = gl->gl_object;
2101                         if (ip && !atomic_read(&ip->i_count))
2102                                 gfs2_inode_destroy(ip);
2103                 }
2104                 if (queue_empty(gl, &gl->gl_holders) &&
2105                     gl->gl_state != LM_ST_UNLOCKED &&
2106                     demote_ok(gl))
2107                         handle_callback(gl, LM_ST_UNLOCKED);
2108                 gfs2_glmutex_unlock(gl);
2109         }
2110
2111         gfs2_glock_put(gl);
2112 }
2113
2114 /**
2115  * examine_bucket - Call a function for glock in a hash bucket
2116  * @examiner: the function
2117  * @sdp: the filesystem
2118  * @bucket: the bucket
2119  *
2120  * Returns: 1 if the bucket has entries
2121  */
2122
2123 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
2124                           struct gfs2_gl_hash_bucket *bucket)
2125 {
2126         struct glock_plug plug;
2127         struct list_head *tmp;
2128         struct gfs2_glock *gl;
2129         int entries;
2130
2131         /* Add "plug" to end of bucket list, work back up list from there */
2132         memset(&plug.gl_flags, 0, sizeof(unsigned long));
2133         set_bit(GLF_PLUG, &plug.gl_flags);
2134
2135         write_lock(&bucket->hb_lock);
2136         list_add(&plug.gl_list, &bucket->hb_list);
2137         write_unlock(&bucket->hb_lock);
2138
2139         for (;;) {
2140                 write_lock(&bucket->hb_lock);
2141
2142                 for (;;) {
2143                         tmp = plug.gl_list.next;
2144
2145                         if (tmp == &bucket->hb_list) {
2146                                 list_del(&plug.gl_list);
2147                                 entries = !list_empty(&bucket->hb_list);
2148                                 write_unlock(&bucket->hb_lock);
2149                                 return entries;
2150                         }
2151                         gl = list_entry(tmp, struct gfs2_glock, gl_list);
2152
2153                         /* Move plug up list */
2154                         list_move(&plug.gl_list, &gl->gl_list);
2155
2156                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2157                                 continue;
2158
2159                         /* examiner() must glock_put() */
2160                         gfs2_glock_hold(gl);
2161
2162                         break;
2163                 }
2164
2165                 write_unlock(&bucket->hb_lock);
2166
2167                 examiner(gl);
2168         }
2169 }
2170
2171 /**
2172  * scan_glock - look at a glock and see if we can reclaim it
2173  * @gl: the glock to look at
2174  *
2175  */
2176
2177 static void scan_glock(struct gfs2_glock *gl)
2178 {
2179         if (gfs2_glmutex_trylock(gl)) {
2180                 if (gl->gl_ops == &gfs2_inode_glops) {
2181                         struct gfs2_inode *ip = gl->gl_object;
2182                         if (ip && !atomic_read(&ip->i_count))
2183                                 goto out_schedule;
2184                 }
2185                 if (queue_empty(gl, &gl->gl_holders) &&
2186                     gl->gl_state != LM_ST_UNLOCKED &&
2187                     demote_ok(gl))
2188                         goto out_schedule;
2189
2190                 gfs2_glmutex_unlock(gl);
2191         }
2192
2193         gfs2_glock_put(gl);
2194
2195         return;
2196
2197  out_schedule:
2198         gfs2_glmutex_unlock(gl);
2199         gfs2_glock_schedule_for_reclaim(gl);
2200         gfs2_glock_put(gl);
2201 }
2202
2203 /**
2204  * gfs2_scand_internal - Look for glocks and inodes to toss from memory
2205  * @sdp: the filesystem
2206  *
2207  */
2208
2209 void gfs2_scand_internal(struct gfs2_sbd *sdp)
2210 {
2211         unsigned int x;
2212
2213         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2214                 examine_bucket(scan_glock, sdp, &sdp->sd_gl_hash[x]);
2215                 cond_resched();
2216         }
2217 }
2218
2219 /**
2220  * clear_glock - look at a glock and see if we can free it from glock cache
2221  * @gl: the glock to look at
2222  *
2223  */
2224
2225 static void clear_glock(struct gfs2_glock *gl)
2226 {
2227         struct gfs2_sbd *sdp = gl->gl_sbd;
2228         int released;
2229
2230         spin_lock(&sdp->sd_reclaim_lock);
2231         if (!list_empty(&gl->gl_reclaim)) {
2232                 list_del_init(&gl->gl_reclaim);
2233                 atomic_dec(&sdp->sd_reclaim_count);
2234                 released = gfs2_glock_put(gl);
2235                 gfs2_assert(sdp, !released);
2236         }
2237         spin_unlock(&sdp->sd_reclaim_lock);
2238
2239         if (gfs2_glmutex_trylock(gl)) {
2240                 if (gl->gl_ops == &gfs2_inode_glops) {
2241                         struct gfs2_inode *ip = gl->gl_object;
2242                         if (ip && !atomic_read(&ip->i_count))
2243                                 gfs2_inode_destroy(ip);
2244                 }
2245                 if (queue_empty(gl, &gl->gl_holders) &&
2246                     gl->gl_state != LM_ST_UNLOCKED)
2247                         handle_callback(gl, LM_ST_UNLOCKED);
2248
2249                 gfs2_glmutex_unlock(gl);
2250         }
2251
2252         gfs2_glock_put(gl);
2253 }
2254
2255 /**
2256  * gfs2_gl_hash_clear - Empty out the glock hash table
2257  * @sdp: the filesystem
2258  * @wait: wait until it's all gone
2259  *
2260  * Called when unmounting the filesystem, or when inter-node lock manager
2261  * requests DROPLOCKS because it is running out of capacity.
2262  */
2263
2264 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
2265 {
2266         unsigned long t;
2267         unsigned int x;
2268         int cont;
2269
2270         t = jiffies;
2271
2272         for (;;) {
2273                 cont = 0;
2274
2275                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
2276                         if (examine_bucket(clear_glock, sdp,
2277                                            &sdp->sd_gl_hash[x]))
2278                                 cont = 1;
2279
2280                 if (!wait || !cont)
2281                         break;
2282
2283                 if (time_after_eq(jiffies,
2284                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
2285                         fs_warn(sdp, "Unmount seems to be stalled. "
2286                                      "Dumping lock state...\n");
2287                         gfs2_dump_lockstate(sdp);
2288                         t = jiffies;
2289                 }
2290
2291                 /* invalidate_inodes() requires that the sb inodes list
2292                    not change, but an async completion callback for an
2293                    unlock can occur which does glock_put() which
2294                    can call iput() which will change the sb inodes list.
2295                    invalidate_inodes_mutex prevents glock_put()'s during
2296                    an invalidate_inodes() */
2297
2298                 mutex_lock(&sdp->sd_invalidate_inodes_mutex);
2299                 invalidate_inodes(sdp->sd_vfs);
2300                 mutex_unlock(&sdp->sd_invalidate_inodes_mutex);
2301                 yield();
2302         }
2303 }
2304
2305 /*
2306  *  Diagnostic routines to help debug distributed deadlock
2307  */
2308
2309 /**
2310  * dump_holder - print information about a glock holder
2311  * @str: a string naming the type of holder
2312  * @gh: the glock holder
2313  *
2314  * Returns: 0 on success, -ENOBUFS when we run out of space
2315  */
2316
2317 static int dump_holder(char *str, struct gfs2_holder *gh)
2318 {
2319         unsigned int x;
2320         int error = -ENOBUFS;
2321
2322         printk(KERN_INFO "  %s\n", str);
2323         printk(KERN_INFO "    owner = %ld\n",
2324                    (gh->gh_owner) ? (long)gh->gh_owner->pid : -1);
2325         printk(KERN_INFO "    gh_state = %u\n", gh->gh_state);
2326         printk(KERN_INFO "    gh_flags =");
2327         for (x = 0; x < 32; x++)
2328                 if (gh->gh_flags & (1 << x))
2329                         printk(" %u", x);
2330         printk(" \n");
2331         printk(KERN_INFO "    error = %d\n", gh->gh_error);
2332         printk(KERN_INFO "    gh_iflags =");
2333         for (x = 0; x < 32; x++)
2334                 if (test_bit(x, &gh->gh_iflags))
2335                         printk(" %u", x);
2336         printk(" \n");
2337
2338         error = 0;
2339
2340         return error;
2341 }
2342
2343 /**
2344  * dump_inode - print information about an inode
2345  * @ip: the inode
2346  *
2347  * Returns: 0 on success, -ENOBUFS when we run out of space
2348  */
2349
2350 static int dump_inode(struct gfs2_inode *ip)
2351 {
2352         unsigned int x;
2353         int error = -ENOBUFS;
2354
2355         printk(KERN_INFO "  Inode:\n");
2356         printk(KERN_INFO "    num = %llu %llu\n",
2357                     ip->i_num.no_formal_ino, ip->i_num.no_addr);
2358         printk(KERN_INFO "    type = %u\n", IF2DT(ip->i_di.di_mode));
2359         printk(KERN_INFO "    i_count = %d\n", atomic_read(&ip->i_count));
2360         printk(KERN_INFO "    i_flags =");
2361         for (x = 0; x < 32; x++)
2362                 if (test_bit(x, &ip->i_flags))
2363                         printk(" %u", x);
2364         printk(" \n");
2365         printk(KERN_INFO "    vnode = %s\n", (ip->i_vnode) ? "yes" : "no");
2366
2367         error = 0;
2368
2369         return error;
2370 }
2371
2372 /**
2373  * dump_glock - print information about a glock
2374  * @gl: the glock
2375  * @count: where we are in the buffer
2376  *
2377  * Returns: 0 on success, -ENOBUFS when we run out of space
2378  */
2379
2380 static int dump_glock(struct gfs2_glock *gl)
2381 {
2382         struct gfs2_holder *gh;
2383         unsigned int x;
2384         int error = -ENOBUFS;
2385
2386         spin_lock(&gl->gl_spin);
2387
2388         printk(KERN_INFO "Glock (%u, %llu)\n",
2389                     gl->gl_name.ln_type,
2390                     gl->gl_name.ln_number);
2391         printk(KERN_INFO "  gl_flags =");
2392         for (x = 0; x < 32; x++)
2393                 if (test_bit(x, &gl->gl_flags))
2394                         printk(" %u", x);
2395         printk(" \n");
2396         printk(KERN_INFO "  gl_ref = %d\n", atomic_read(&gl->gl_ref.refcount));
2397         printk(KERN_INFO "  gl_state = %u\n", gl->gl_state);
2398         printk(KERN_INFO "  req_gh = %s\n", (gl->gl_req_gh) ? "yes" : "no");
2399         printk(KERN_INFO "  req_bh = %s\n", (gl->gl_req_bh) ? "yes" : "no");
2400         printk(KERN_INFO "  lvb_count = %d\n", atomic_read(&gl->gl_lvb_count));
2401         printk(KERN_INFO "  object = %s\n", (gl->gl_object) ? "yes" : "no");
2402         printk(KERN_INFO "  le = %s\n",
2403                    (list_empty(&gl->gl_le.le_list)) ? "no" : "yes");
2404         printk(KERN_INFO "  reclaim = %s\n",
2405                     (list_empty(&gl->gl_reclaim)) ? "no" : "yes");
2406         if (gl->gl_aspace)
2407                 printk(KERN_INFO "  aspace = %lu\n",
2408                             gl->gl_aspace->i_mapping->nrpages);
2409         else
2410                 printk(KERN_INFO "  aspace = no\n");
2411         printk(KERN_INFO "  ail = %d\n", atomic_read(&gl->gl_ail_count));
2412         if (gl->gl_req_gh) {
2413                 error = dump_holder("Request", gl->gl_req_gh);
2414                 if (error)
2415                         goto out;
2416         }
2417         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
2418                 error = dump_holder("Holder", gh);
2419                 if (error)
2420                         goto out;
2421         }
2422         list_for_each_entry(gh, &gl->gl_waiters1, gh_list) {
2423                 error = dump_holder("Waiter1", gh);
2424                 if (error)
2425                         goto out;
2426         }
2427         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
2428                 error = dump_holder("Waiter2", gh);
2429                 if (error)
2430                         goto out;
2431         }
2432         list_for_each_entry(gh, &gl->gl_waiters3, gh_list) {
2433                 error = dump_holder("Waiter3", gh);
2434                 if (error)
2435                         goto out;
2436         }
2437         if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object) {
2438                 if (!test_bit(GLF_LOCK, &gl->gl_flags) &&
2439                     list_empty(&gl->gl_holders)) {
2440                         error = dump_inode(gl->gl_object);
2441                         if (error)
2442                                 goto out;
2443                 } else {
2444                         error = -ENOBUFS;
2445                         printk(KERN_INFO "  Inode: busy\n");
2446                 }
2447         }
2448
2449         error = 0;
2450
2451  out:
2452         spin_unlock(&gl->gl_spin);
2453
2454         return error;
2455 }
2456
2457 /**
2458  * gfs2_dump_lockstate - print out the current lockstate
2459  * @sdp: the filesystem
2460  * @ub: the buffer to copy the information into
2461  *
2462  * If @ub is NULL, dump the lockstate to the console.
2463  *
2464  */
2465
2466 int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
2467 {
2468         struct gfs2_gl_hash_bucket *bucket;
2469         struct gfs2_glock *gl;
2470         unsigned int x;
2471         int error = 0;
2472
2473         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2474                 bucket = &sdp->sd_gl_hash[x];
2475
2476                 read_lock(&bucket->hb_lock);
2477
2478                 list_for_each_entry(gl, &bucket->hb_list, gl_list) {
2479                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2480                                 continue;
2481
2482                         error = dump_glock(gl);
2483                         if (error)
2484                                 break;
2485                 }
2486
2487                 read_unlock(&bucket->hb_lock);
2488
2489                 if (error)
2490                         break;
2491         }
2492
2493
2494         return error;
2495 }
2496