fs/gfs2/glock.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 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 version 2.
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/sched.h>
  13 #include <linux/slab.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/buffer_head.h>
  16 #include <linux/delay.h>
  17 #include <linux/sort.h>
  18 #include <linux/jhash.h>
  19 #include <linux/kallsyms.h>
  20 #include <linux/gfs2_ondisk.h>
  21 #include <linux/list.h>
  22 #include <linux/wait.h>
  23 #include <linux/module.h>
  24 #include <asm/uaccess.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/debugfs.h>
  27 #include <linux/kthread.h>
  28 #include <linux/freezer.h>
  29 #include <linux/workqueue.h>
  30 #include <linux/jiffies.h>
  31 #include <linux/rcupdate.h>
  32 #include <linux/rculist_bl.h>
  33 #include <linux/bit_spinlock.h>
  34 #include <linux/percpu.h>
  35 #include <linux/list_sort.h>
  36 #include <linux/lockref.h>
  37
  38 #include "gfs2.h"
  39 #include "incore.h"
  40 #include "glock.h"
  41 #include "glops.h"
  42 #include "inode.h"
  43 #include "lops.h"
  44 #include "meta_io.h"
  45 #include "quota.h"
  46 #include "super.h"
  47 #include "util.h"
  48 #include "bmap.h"
  49 #define CREATE_TRACE_POINTS
  50 #include "trace_gfs2.h"
  51
  52 struct gfs2_glock_iter {
  53         int hash;                       /* hash bucket index           */
  54         unsigned nhash;                 /* Index within current bucket */
  55         struct gfs2_sbd *sdp;           /* incore superblock           */
  56         struct gfs2_glock *gl;          /* current glock struct        */
  57         loff_t last_pos;                /* last position               */
  58 };
  59
  60 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  61
  62 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  63
  64 static struct dentry *gfs2_root;
  65 static struct workqueue_struct *glock_workqueue;
  66 struct workqueue_struct *gfs2_delete_workqueue;
  67 static LIST_HEAD(lru_list);
  68 static atomic_t lru_count = ATOMIC_INIT(0);
  69 static DEFINE_SPINLOCK(lru_lock);
  70
  71 #define GFS2_GL_HASH_SHIFT      15
  72 #define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
  73 #define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
  74
  75 static struct hlist_bl_head gl_hash_table[GFS2_GL_HASH_SIZE];
  76 static struct dentry *gfs2_root;
  77
  78 /**
  79  * gl_hash() - Turn glock number into hash bucket number
  80  * @lock: The glock number
  81  *
  82  * Returns: The number of the corresponding hash bucket
  83  */
  84
  85 static unsigned int gl_hash(const struct gfs2_sbd *sdp,
  86                             const struct lm_lockname *name)
  87 {
  88         unsigned int h;
  89
  90         h = jhash(&name->ln_number, sizeof(u64), 0);
  91         h = jhash(&name->ln_type, sizeof(unsigned int), h);
  92         h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
  93         h &= GFS2_GL_HASH_MASK;
  94
  95         return h;
  96 }
  97
  98 static inline void spin_lock_bucket(unsigned int hash)
  99 {
 100         hlist_bl_lock(&gl_hash_table[hash]);
 101 }
 102
 103 static inline void spin_unlock_bucket(unsigned int hash)
 104 {
 105         hlist_bl_unlock(&gl_hash_table[hash]);
 106 }
 107
 108 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 109 {
 110         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 111
 112         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
 113                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 114         } else {
 115                 kfree(gl->gl_lksb.sb_lvbptr);
 116                 kmem_cache_free(gfs2_glock_cachep, gl);
 117         }
 118 }
 119
 120 void gfs2_glock_free(struct gfs2_glock *gl)
 121 {
 122         struct gfs2_sbd *sdp = gl->gl_sbd;
 123
 124         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 125         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 126                 wake_up(&sdp->sd_glock_wait);
 127 }
 128
 129 /**
 130  * gfs2_glock_hold() - increment reference count on glock
 131  * @gl: The glock to hold
 132  *
 133  */
 134
 135 static void gfs2_glock_hold(struct gfs2_glock *gl)
 136 {
 137         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 138         lockref_get(&gl->gl_lockref);
 139 }
 140
 141 /**
 142  * demote_ok - Check to see if it's ok to unlock a glock
 143  * @gl: the glock
 144  *
 145  * Returns: 1 if it's ok
 146  */
 147
 148 static int demote_ok(const struct gfs2_glock *gl)
 149 {
 150         const struct gfs2_glock_operations *glops = gl->gl_ops;
 151
 152         if (gl->gl_state == LM_ST_UNLOCKED)
 153                 return 0;
 154         if (!list_empty(&gl->gl_holders))
 155                 return 0;
 156         if (glops->go_demote_ok)
 157                 return glops->go_demote_ok(gl);
 158         return 1;
 159 }
 160
 161
 162 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 163 {
 164         spin_lock(&lru_lock);
 165
 166         if (!list_empty(&gl->gl_lru))
 167                 list_del_init(&gl->gl_lru);
 168         else
 169                 atomic_inc(&lru_count);
 170
 171         list_add_tail(&gl->gl_lru, &lru_list);
 172         set_bit(GLF_LRU, &gl->gl_flags);
 173         spin_unlock(&lru_lock);
 174 }
 175
 176 static void __gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 177 {
 178         if (!list_empty(&gl->gl_lru)) {
 179                 list_del_init(&gl->gl_lru);
 180                 atomic_dec(&lru_count);
 181                 clear_bit(GLF_LRU, &gl->gl_flags);
 182         }
 183 }
 184
 185 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 186 {
 187         spin_lock(&lru_lock);
 188         __gfs2_glock_remove_from_lru(gl);
 189         spin_unlock(&lru_lock);
 190 }
 191
 192 /**
 193  * gfs2_glock_put() - Decrement reference count on glock
 194  * @gl: The glock to put
 195  *
 196  */
 197
 198 void gfs2_glock_put(struct gfs2_glock *gl)
 199 {
 200         struct gfs2_sbd *sdp = gl->gl_sbd;
 201         struct address_space *mapping = gfs2_glock2aspace(gl);
 202
 203         if (lockref_put_or_lock(&gl->gl_lockref))
 204                 return;
 205
 206         lockref_mark_dead(&gl->gl_lockref);
 207
 208         spin_lock(&lru_lock);
 209         __gfs2_glock_remove_from_lru(gl);
 210         spin_unlock(&lru_lock);
 211         spin_unlock(&gl->gl_lockref.lock);
 212         spin_lock_bucket(gl->gl_hash);
 213         hlist_bl_del_rcu(&gl->gl_list);
 214         spin_unlock_bucket(gl->gl_hash);
 215         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 216         GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
 217         trace_gfs2_glock_put(gl);
 218         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 219 }
 220
 221 /**
 222  * search_bucket() - Find struct gfs2_glock by lock number
 223  * @bucket: the bucket to search
 224  * @name: The lock name
 225  *
 226  * Returns: NULL, or the struct gfs2_glock with the requested number
 227  */
 228
 229 static struct gfs2_glock *search_bucket(unsigned int hash,
 230                                         const struct gfs2_sbd *sdp,
 231                                         const struct lm_lockname *name)
 232 {
 233         struct gfs2_glock *gl;
 234         struct hlist_bl_node *h;
 235
 236         hlist_bl_for_each_entry_rcu(gl, h, &gl_hash_table[hash], gl_list) {
 237                 if (!lm_name_equal(&gl->gl_name, name))
 238                         continue;
 239                 if (gl->gl_sbd != sdp)
 240                         continue;
 241                 if (lockref_get_not_dead(&gl->gl_lockref))
 242                         return gl;
 243         }
 244
 245         return NULL;
 246 }
 247
 248 /**
 249  * may_grant - check if its ok to grant a new lock
 250  * @gl: The glock
 251  * @gh: The lock request which we wish to grant
 252  *
 253  * Returns: true if its ok to grant the lock
 254  */
 255
 256 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 257 {
 258         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 259         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 260              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 261                 return 0;
 262         if (gl->gl_state == gh->gh_state)
 263                 return 1;
 264         if (gh->gh_flags & GL_EXACT)
 265                 return 0;
 266         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 267                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 268                         return 1;
 269                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 270                         return 1;
 271         }
 272         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 273                 return 1;
 274         return 0;
 275 }
 276
 277 static void gfs2_holder_wake(struct gfs2_holder *gh)
 278 {
 279         clear_bit(HIF_WAIT, &gh->gh_iflags);
 280         smp_mb__after_atomic();
 281         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 282 }
 283
 284 /**
 285  * do_error - Something unexpected has happened during a lock request
 286  *
 287  */
 288
 289 static inline void do_error(struct gfs2_glock *gl, const int ret)
 290 {
 291         struct gfs2_holder *gh, *tmp;
 292
 293         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 294                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 295                         continue;
 296                 if (ret & LM_OUT_ERROR)
 297                         gh->gh_error = -EIO;
 298                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 299                         gh->gh_error = GLR_TRYFAILED;
 300                 else
 301                         continue;
 302                 list_del_init(&gh->gh_list);
 303                 trace_gfs2_glock_queue(gh, 0);
 304                 gfs2_holder_wake(gh);
 305         }
 306 }
 307
 308 /**
 309  * do_promote - promote as many requests as possible on the current queue
 310  * @gl: The glock
 311  *
 312  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 313  *          if a type specific operation is underway.
 314  */
 315
 316 static int do_promote(struct gfs2_glock *gl)
 317 __releases(&gl->gl_spin)
 318 __acquires(&gl->gl_spin)
 319 {
 320         const struct gfs2_glock_operations *glops = gl->gl_ops;
 321         struct gfs2_holder *gh, *tmp;
 322         int ret;
 323
 324 restart:
 325         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 326                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 327                         continue;
 328                 if (may_grant(gl, gh)) {
 329                         if (gh->gh_list.prev == &gl->gl_holders &&
 330                             glops->go_lock) {
 331                                 spin_unlock(&gl->gl_spin);
 332                                 /* FIXME: eliminate this eventually */
 333                                 ret = glops->go_lock(gh);
 334                                 spin_lock(&gl->gl_spin);
 335                                 if (ret) {
 336                                         if (ret == 1)
 337                                                 return 2;
 338                                         gh->gh_error = ret;
 339                                         list_del_init(&gh->gh_list);
 340                                         trace_gfs2_glock_queue(gh, 0);
 341                                         gfs2_holder_wake(gh);
 342                                         goto restart;
 343                                 }
 344                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 345                                 trace_gfs2_promote(gh, 1);
 346                                 gfs2_holder_wake(gh);
 347                                 goto restart;
 348                         }
 349                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 350                         trace_gfs2_promote(gh, 0);
 351                         gfs2_holder_wake(gh);
 352                         continue;
 353                 }
 354                 if (gh->gh_list.prev == &gl->gl_holders)
 355                         return 1;
 356                 do_error(gl, 0);
 357                 break;
 358         }
 359         return 0;
 360 }
 361
 362 /**
 363  * find_first_waiter - find the first gh that's waiting for the glock
 364  * @gl: the glock
 365  */
 366
 367 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 368 {
 369         struct gfs2_holder *gh;
 370
 371         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 372                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 373                         return gh;
 374         }
 375         return NULL;
 376 }
 377
 378 /**
 379  * state_change - record that the glock is now in a different state
 380  * @gl: the glock
 381  * @new_state the new state
 382  *
 383  */
 384
 385 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 386 {
 387         int held1, held2;
 388
 389         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 390         held2 = (new_state != LM_ST_UNLOCKED);
 391
 392         if (held1 != held2) {
 393                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 394                 if (held2)
 395                         gl->gl_lockref.count++;
 396                 else
 397                         gl->gl_lockref.count--;
 398         }
 399         if (held1 && held2 && list_empty(&gl->gl_holders))
 400                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 401
 402         if (new_state != gl->gl_target)
 403                 /* shorten our minimum hold time */
 404                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 405                                        GL_GLOCK_MIN_HOLD);
 406         gl->gl_state = new_state;
 407         gl->gl_tchange = jiffies;
 408 }
 409
 410 static void gfs2_demote_wake(struct gfs2_glock *gl)
 411 {
 412         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 413         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 414         smp_mb__after_atomic();
 415         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 416 }
 417
 418 /**
 419  * finish_xmote - The DLM has replied to one of our lock requests
 420  * @gl: The glock
 421  * @ret: The status from the DLM
 422  *
 423  */
 424
 425 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 426 {
 427         const struct gfs2_glock_operations *glops = gl->gl_ops;
 428         struct gfs2_holder *gh;
 429         unsigned state = ret & LM_OUT_ST_MASK;
 430         int rv;
 431
 432         spin_lock(&gl->gl_spin);
 433         trace_gfs2_glock_state_change(gl, state);
 434         state_change(gl, state);
 435         gh = find_first_waiter(gl);
 436
 437         /* Demote to UN request arrived during demote to SH or DF */
 438         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 439             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 440                 gl->gl_target = LM_ST_UNLOCKED;
 441
 442         /* Check for state != intended state */
 443         if (unlikely(state != gl->gl_target)) {
 444                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 445                         /* move to back of queue and try next entry */
 446                         if (ret & LM_OUT_CANCELED) {
 447                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 448                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 449                                 gh = find_first_waiter(gl);
 450                                 gl->gl_target = gh->gh_state;
 451                                 goto retry;
 452                         }
 453                         /* Some error or failed "try lock" - report it */
 454                         if ((ret & LM_OUT_ERROR) ||
 455                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 456                                 gl->gl_target = gl->gl_state;
 457                                 do_error(gl, ret);
 458                                 goto out;
 459                         }
 460                 }
 461                 switch(state) {
 462                 /* Unlocked due to conversion deadlock, try again */
 463                 case LM_ST_UNLOCKED:
 464 retry:
 465                         do_xmote(gl, gh, gl->gl_target);
 466                         break;
 467                 /* Conversion fails, unlock and try again */
 468                 case LM_ST_SHARED:
 469                 case LM_ST_DEFERRED:
 470                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 471                         break;
 472                 default: /* Everything else */
 473                         pr_err("wanted %u got %u\n", gl->gl_target, state);
 474                         GLOCK_BUG_ON(gl, 1);
 475                 }
 476                 spin_unlock(&gl->gl_spin);
 477                 return;
 478         }
 479
 480         /* Fast path - we got what we asked for */
 481         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 482                 gfs2_demote_wake(gl);
 483         if (state != LM_ST_UNLOCKED) {
 484                 if (glops->go_xmote_bh) {
 485                         spin_unlock(&gl->gl_spin);
 486                         rv = glops->go_xmote_bh(gl, gh);
 487                         spin_lock(&gl->gl_spin);
 488                         if (rv) {
 489                                 do_error(gl, rv);
 490                                 goto out;
 491                         }
 492                 }
 493                 rv = do_promote(gl);
 494                 if (rv == 2)
 495                         goto out_locked;
 496         }
 497 out:
 498         clear_bit(GLF_LOCK, &gl->gl_flags);
 499 out_locked:
 500         spin_unlock(&gl->gl_spin);
 501 }
 502
 503 /**
 504  * do_xmote - Calls the DLM to change the state of a lock
 505  * @gl: The lock state
 506  * @gh: The holder (only for promotes)
 507  * @target: The target lock state
 508  *
 509  */
 510
 511 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 512 __releases(&gl->gl_spin)
 513 __acquires(&gl->gl_spin)
 514 {
 515         const struct gfs2_glock_operations *glops = gl->gl_ops;
 516         struct gfs2_sbd *sdp = gl->gl_sbd;
 517         unsigned int lck_flags = gh ? gh->gh_flags : 0;
 518         int ret;
 519
 520         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 521                       LM_FLAG_PRIORITY);
 522         GLOCK_BUG_ON(gl, gl->gl_state == target);
 523         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 524         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 525             glops->go_inval) {
 526                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 527                 do_error(gl, 0); /* Fail queued try locks */
 528         }
 529         gl->gl_req = target;
 530         set_bit(GLF_BLOCKING, &gl->gl_flags);
 531         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 532             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 533             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 534                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 535         spin_unlock(&gl->gl_spin);
 536         if (glops->go_sync)
 537                 glops->go_sync(gl);
 538         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 539                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 540         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 541
 542         gfs2_glock_hold(gl);
 543         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 544                 /* lock_dlm */
 545                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 546                 if (ret) {
 547                         pr_err("lm_lock ret %d\n", ret);
 548                         GLOCK_BUG_ON(gl, 1);
 549                 }
 550         } else { /* lock_nolock */
 551                 finish_xmote(gl, target);
 552                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
 553                         gfs2_glock_put(gl);
 554         }
 555
 556         spin_lock(&gl->gl_spin);
 557 }
 558
 559 /**
 560  * find_first_holder - find the first "holder" gh
 561  * @gl: the glock
 562  */
 563
 564 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 565 {
 566         struct gfs2_holder *gh;
 567
 568         if (!list_empty(&gl->gl_holders)) {
 569                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 570                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 571                         return gh;
 572         }
 573         return NULL;
 574 }
 575
 576 /**
 577  * run_queue - do all outstanding tasks related to a glock
 578  * @gl: The glock in question
 579  * @nonblock: True if we must not block in run_queue
 580  *
 581  */
 582
 583 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 584 __releases(&gl->gl_spin)
 585 __acquires(&gl->gl_spin)
 586 {
 587         struct gfs2_holder *gh = NULL;
 588         int ret;
 589
 590         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 591                 return;
 592
 593         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 594
 595         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 596             gl->gl_demote_state != gl->gl_state) {
 597                 if (find_first_holder(gl))
 598                         goto out_unlock;
 599                 if (nonblock)
 600                         goto out_sched;
 601                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 602                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 603                 gl->gl_target = gl->gl_demote_state;
 604         } else {
 605                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 606                         gfs2_demote_wake(gl);
 607                 ret = do_promote(gl);
 608                 if (ret == 0)
 609                         goto out_unlock;
 610                 if (ret == 2)
 611                         goto out;
 612                 gh = find_first_waiter(gl);
 613                 gl->gl_target = gh->gh_state;
 614                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 615                         do_error(gl, 0); /* Fail queued try locks */
 616         }
 617         do_xmote(gl, gh, gl->gl_target);
 618 out:
 619         return;
 620
 621 out_sched:
 622         clear_bit(GLF_LOCK, &gl->gl_flags);
 623         smp_mb__after_atomic();
 624         gl->gl_lockref.count++;
 625         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
 626                 gl->gl_lockref.count--;
 627         return;
 628
 629 out_unlock:
 630         clear_bit(GLF_LOCK, &gl->gl_flags);
 631         smp_mb__after_atomic();
 632         return;
 633 }
 634
 635 static void delete_work_func(struct work_struct *work)
 636 {
 637         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 638         struct gfs2_sbd *sdp = gl->gl_sbd;
 639         struct gfs2_inode *ip;
 640         struct inode *inode;
 641         u64 no_addr = gl->gl_name.ln_number;
 642
 643         ip = gl->gl_object;
 644         /* Note: Unsafe to dereference ip as we don't hold right refs/locks */
 645
 646         if (ip)
 647                 inode = gfs2_ilookup(sdp->sd_vfs, no_addr, 1);
 648         else
 649                 inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 650         if (inode && !IS_ERR(inode)) {
 651                 d_prune_aliases(inode);
 652                 iput(inode);
 653         }
 654         gfs2_glock_put(gl);
 655 }
 656
 657 static void glock_work_func(struct work_struct *work)
 658 {
 659         unsigned long delay = 0;
 660         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 661         int drop_ref = 0;
 662
 663         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 664                 finish_xmote(gl, gl->gl_reply);
 665                 drop_ref = 1;
 666         }
 667         spin_lock(&gl->gl_spin);
 668         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 669             gl->gl_state != LM_ST_UNLOCKED &&
 670             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 671                 unsigned long holdtime, now = jiffies;
 672
 673                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 674                 if (time_before(now, holdtime))
 675                         delay = holdtime - now;
 676
 677                 if (!delay) {
 678                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 679                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 680                 }
 681         }
 682         run_queue(gl, 0);
 683         spin_unlock(&gl->gl_spin);
 684         if (!delay)
 685                 gfs2_glock_put(gl);
 686         else {
 687                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 688                         delay = 0;
 689                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
 690                         gfs2_glock_put(gl);
 691         }
 692         if (drop_ref)
 693                 gfs2_glock_put(gl);
 694 }
 695
 696 /**
 697  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 698  * @sdp: The GFS2 superblock
 699  * @number: the lock number
 700  * @glops: The glock_operations to use
 701  * @create: If 0, don't create the glock if it doesn't exist
 702  * @glp: the glock is returned here
 703  *
 704  * This does not lock a glock, just finds/creates structures for one.
 705  *
 706  * Returns: errno
 707  */
 708
 709 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 710                    const struct gfs2_glock_operations *glops, int create,
 711                    struct gfs2_glock **glp)
 712 {
 713         struct super_block *s = sdp->sd_vfs;
 714         struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
 715         struct gfs2_glock *gl, *tmp;
 716         unsigned int hash = gl_hash(sdp, &name);
 717         struct address_space *mapping;
 718         struct kmem_cache *cachep;
 719
 720         rcu_read_lock();
 721         gl = search_bucket(hash, sdp, &name);
 722         rcu_read_unlock();
 723
 724         *glp = gl;
 725         if (gl)
 726                 return 0;
 727         if (!create)
 728                 return -ENOENT;
 729
 730         if (glops->go_flags & GLOF_ASPACE)
 731                 cachep = gfs2_glock_aspace_cachep;
 732         else
 733                 cachep = gfs2_glock_cachep;
 734         gl = kmem_cache_alloc(cachep, GFP_KERNEL);
 735         if (!gl)
 736                 return -ENOMEM;
 737
 738         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 739
 740         if (glops->go_flags & GLOF_LVB) {
 741                 gl->gl_lksb.sb_lvbptr = kzalloc(GFS2_MIN_LVB_SIZE, GFP_KERNEL);
 742                 if (!gl->gl_lksb.sb_lvbptr) {
 743                         kmem_cache_free(cachep, gl);
 744                         return -ENOMEM;
 745                 }
 746         }
 747
 748         atomic_inc(&sdp->sd_glock_disposal);
 749         gl->gl_sbd = sdp;
 750         gl->gl_flags = 0;
 751         gl->gl_name = name;
 752         gl->gl_lockref.count = 1;
 753         gl->gl_state = LM_ST_UNLOCKED;
 754         gl->gl_target = LM_ST_UNLOCKED;
 755         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 756         gl->gl_hash = hash;
 757         gl->gl_ops = glops;
 758         gl->gl_dstamp = ktime_set(0, 0);
 759         preempt_disable();
 760         /* We use the global stats to estimate the initial per-glock stats */
 761         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 762         preempt_enable();
 763         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 764         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 765         gl->gl_tchange = jiffies;
 766         gl->gl_object = NULL;
 767         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 768         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 769         INIT_WORK(&gl->gl_delete, delete_work_func);
 770
 771         mapping = gfs2_glock2aspace(gl);
 772         if (mapping) {
 773                 mapping->a_ops = &gfs2_meta_aops;
 774                 mapping->host = s->s_bdev->bd_inode;
 775                 mapping->flags = 0;
 776                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 777                 mapping->private_data = NULL;
 778                 mapping->backing_dev_info = s->s_bdi;
 779                 mapping->writeback_index = 0;
 780         }
 781
 782         spin_lock_bucket(hash);
 783         tmp = search_bucket(hash, sdp, &name);
 784         if (tmp) {
 785                 spin_unlock_bucket(hash);
 786                 kfree(gl->gl_lksb.sb_lvbptr);
 787                 kmem_cache_free(cachep, gl);
 788                 atomic_dec(&sdp->sd_glock_disposal);
 789                 gl = tmp;
 790         } else {
 791                 hlist_bl_add_head_rcu(&gl->gl_list, &gl_hash_table[hash]);
 792                 spin_unlock_bucket(hash);
 793         }
 794
 795         *glp = gl;
 796
 797         return 0;
 798 }
 799
 800 /**
 801  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 802  * @gl: the glock
 803  * @state: the state we're requesting
 804  * @flags: the modifier flags
 805  * @gh: the holder structure
 806  *
 807  */
 808
 809 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
 810                       struct gfs2_holder *gh)
 811 {
 812         INIT_LIST_HEAD(&gh->gh_list);
 813         gh->gh_gl = gl;
 814         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 815         gh->gh_owner_pid = get_pid(task_pid(current));
 816         gh->gh_state = state;
 817         gh->gh_flags = flags;
 818         gh->gh_error = 0;
 819         gh->gh_iflags = 0;
 820         gfs2_glock_hold(gl);
 821 }
 822
 823 /**
 824  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 825  * @state: the state we're requesting
 826  * @flags: the modifier flags
 827  * @gh: the holder structure
 828  *
 829  * Don't mess with the glock.
 830  *
 831  */
 832
 833 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
 834 {
 835         gh->gh_state = state;
 836         gh->gh_flags = flags;
 837         gh->gh_iflags = 0;
 838         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 839         if (gh->gh_owner_pid)
 840                 put_pid(gh->gh_owner_pid);
 841         gh->gh_owner_pid = get_pid(task_pid(current));
 842 }
 843
 844 /**
 845  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 846  * @gh: the holder structure
 847  *
 848  */
 849
 850 void gfs2_holder_uninit(struct gfs2_holder *gh)
 851 {
 852         put_pid(gh->gh_owner_pid);
 853         gfs2_glock_put(gh->gh_gl);
 854         gh->gh_gl = NULL;
 855         gh->gh_ip = 0;
 856 }
 857
 858 /**
 859  * gfs2_glock_holder_wait
 860  * @word: unused
 861  *
 862  * This function and gfs2_glock_demote_wait both show up in the WCHAN
 863  * field. Thus I've separated these otherwise identical functions in
 864  * order to be more informative to the user.
 865  */
 866
 867 static int gfs2_glock_holder_wait(void *word)
 868 {
 869         schedule();
 870         return 0;
 871 }
 872
 873 static int gfs2_glock_demote_wait(void *word)
 874 {
 875         schedule();
 876         return 0;
 877 }
 878
 879 /**
 880  * gfs2_glock_wait - wait on a glock acquisition
 881  * @gh: the glock holder
 882  *
 883  * Returns: 0 on success
 884  */
 885
 886 int gfs2_glock_wait(struct gfs2_holder *gh)
 887 {
 888         unsigned long time1 = jiffies;
 889
 890         might_sleep();
 891         wait_on_bit(&gh->gh_iflags, HIF_WAIT, gfs2_glock_holder_wait, TASK_UNINTERRUPTIBLE);
 892         if (time_after(jiffies, time1 + HZ)) /* have we waited > a second? */
 893                 /* Lengthen the minimum hold time. */
 894                 gh->gh_gl->gl_hold_time = min(gh->gh_gl->gl_hold_time +
 895                                               GL_GLOCK_HOLD_INCR,
 896                                               GL_GLOCK_MAX_HOLD);
 897         return gh->gh_error;
 898 }
 899
 900 /**
 901  * handle_callback - process a demote request
 902  * @gl: the glock
 903  * @state: the state the caller wants us to change to
 904  *
 905  * There are only two requests that we are going to see in actual
 906  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 907  */
 908
 909 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 910                             unsigned long delay, bool remote)
 911 {
 912         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 913
 914         set_bit(bit, &gl->gl_flags);
 915         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 916                 gl->gl_demote_state = state;
 917                 gl->gl_demote_time = jiffies;
 918         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 919                         gl->gl_demote_state != state) {
 920                 gl->gl_demote_state = LM_ST_UNLOCKED;
 921         }
 922         if (gl->gl_ops->go_callback)
 923                 gl->gl_ops->go_callback(gl, remote);
 924         trace_gfs2_demote_rq(gl, remote);
 925 }
 926
 927 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 928 {
 929         struct va_format vaf;
 930         va_list args;
 931
 932         va_start(args, fmt);
 933
 934         if (seq) {
 935                 seq_vprintf(seq, fmt, args);
 936         } else {
 937                 vaf.fmt = fmt;
 938                 vaf.va = &args;
 939
 940                 pr_err("%pV", &vaf);
 941         }
 942
 943         va_end(args);
 944 }
 945
 946 /**
 947  * add_to_queue - Add a holder to the wait queue (but look for recursion)
 948  * @gh: the holder structure to add
 949  *
 950  * Eventually we should move the recursive locking trap to a
 951  * debugging option or something like that. This is the fast
 952  * path and needs to have the minimum number of distractions.
 953  *
 954  */
 955
 956 static inline void add_to_queue(struct gfs2_holder *gh)
 957 __releases(&gl->gl_spin)
 958 __acquires(&gl->gl_spin)
 959 {
 960         struct gfs2_glock *gl = gh->gh_gl;
 961         struct gfs2_sbd *sdp = gl->gl_sbd;
 962         struct list_head *insert_pt = NULL;
 963         struct gfs2_holder *gh2;
 964         int try_futile = 0;
 965
 966         BUG_ON(gh->gh_owner_pid == NULL);
 967         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
 968                 BUG();
 969
 970         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 971                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 972                         try_futile = !may_grant(gl, gh);
 973                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 974                         goto fail;
 975         }
 976
 977         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
 978                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
 979                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
 980                         goto trap_recursive;
 981                 if (try_futile &&
 982                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 983 fail:
 984                         gh->gh_error = GLR_TRYFAILED;
 985                         gfs2_holder_wake(gh);
 986                         return;
 987                 }
 988                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
 989                         continue;
 990                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
 991                         insert_pt = &gh2->gh_list;
 992         }
 993         set_bit(GLF_QUEUED, &gl->gl_flags);
 994         trace_gfs2_glock_queue(gh, 1);
 995         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
 996         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
 997         if (likely(insert_pt == NULL)) {
 998                 list_add_tail(&gh->gh_list, &gl->gl_holders);
 999                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1000                         goto do_cancel;
1001                 return;
1002         }
1003         list_add_tail(&gh->gh_list, insert_pt);
1004 do_cancel:
1005         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1006         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1007                 spin_unlock(&gl->gl_spin);
1008                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1009                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1010                 spin_lock(&gl->gl_spin);
1011         }
1012         return;
1013
1014 trap_recursive:
1015         pr_err("original: %pSR\n", (void *)gh2->gh_ip);
1016         pr_err("pid: %d\n", pid_nr(gh2->gh_owner_pid));
1017         pr_err("lock type: %d req lock state : %d\n",
1018                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1019         pr_err("new: %pSR\n", (void *)gh->gh_ip);
1020         pr_err("pid: %d\n", pid_nr(gh->gh_owner_pid));
1021         pr_err("lock type: %d req lock state : %d\n",
1022                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1023         gfs2_dump_glock(NULL, gl);
1024         BUG();
1025 }
1026
1027 /**
1028  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1029  * @gh: the holder structure
1030  *
1031  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1032  *
1033  * Returns: 0, GLR_TRYFAILED, or errno on failure
1034  */
1035
1036 int gfs2_glock_nq(struct gfs2_holder *gh)
1037 {
1038         struct gfs2_glock *gl = gh->gh_gl;
1039         struct gfs2_sbd *sdp = gl->gl_sbd;
1040         int error = 0;
1041
1042         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1043                 return -EIO;
1044
1045         if (test_bit(GLF_LRU, &gl->gl_flags))
1046                 gfs2_glock_remove_from_lru(gl);
1047
1048         spin_lock(&gl->gl_spin);
1049         add_to_queue(gh);
1050         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1051                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1052                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1053                 gl->gl_lockref.count++;
1054                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1055                         gl->gl_lockref.count--;
1056         }
1057         run_queue(gl, 1);
1058         spin_unlock(&gl->gl_spin);
1059
1060         if (!(gh->gh_flags & GL_ASYNC))
1061                 error = gfs2_glock_wait(gh);
1062
1063         return error;
1064 }
1065
1066 /**
1067  * gfs2_glock_poll - poll to see if an async request has been completed
1068  * @gh: the holder
1069  *
1070  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1071  */
1072
1073 int gfs2_glock_poll(struct gfs2_holder *gh)
1074 {
1075         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1076 }
1077
1078 /**
1079  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1080  * @gh: the glock holder
1081  *
1082  */
1083
1084 void gfs2_glock_dq(struct gfs2_holder *gh)
1085 {
1086         struct gfs2_glock *gl = gh->gh_gl;
1087         const struct gfs2_glock_operations *glops = gl->gl_ops;
1088         unsigned delay = 0;
1089         int fast_path = 0;
1090
1091         spin_lock(&gl->gl_spin);
1092         if (gh->gh_flags & GL_NOCACHE)
1093                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1094
1095         list_del_init(&gh->gh_list);
1096         if (find_first_holder(gl) == NULL) {
1097                 if (glops->go_unlock) {
1098                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1099                         spin_unlock(&gl->gl_spin);
1100                         glops->go_unlock(gh);
1101                         spin_lock(&gl->gl_spin);
1102                         clear_bit(GLF_LOCK, &gl->gl_flags);
1103                 }
1104                 if (list_empty(&gl->gl_holders) &&
1105                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1106                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1107                         fast_path = 1;
1108         }
1109         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1110                 gfs2_glock_add_to_lru(gl);
1111
1112         trace_gfs2_glock_queue(gh, 0);
1113         spin_unlock(&gl->gl_spin);
1114         if (likely(fast_path))
1115                 return;
1116
1117         gfs2_glock_hold(gl);
1118         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1119             !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1120             gl->gl_name.ln_type == LM_TYPE_INODE)
1121                 delay = gl->gl_hold_time;
1122         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1123                 gfs2_glock_put(gl);
1124 }
1125
1126 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1127 {
1128         struct gfs2_glock *gl = gh->gh_gl;
1129         gfs2_glock_dq(gh);
1130         might_sleep();
1131         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, gfs2_glock_demote_wait, TASK_UNINTERRUPTIBLE);
1132 }
1133
1134 /**
1135  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1136  * @gh: the holder structure
1137  *
1138  */
1139
1140 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1141 {
1142         gfs2_glock_dq(gh);
1143         gfs2_holder_uninit(gh);
1144 }
1145
1146 /**
1147  * gfs2_glock_nq_num - acquire a glock based on lock number
1148  * @sdp: the filesystem
1149  * @number: the lock number
1150  * @glops: the glock operations for the type of glock
1151  * @state: the state to acquire the glock in
1152  * @flags: modifier flags for the acquisition
1153  * @gh: the struct gfs2_holder
1154  *
1155  * Returns: errno
1156  */
1157
1158 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1159                       const struct gfs2_glock_operations *glops,
1160                       unsigned int state, int flags, struct gfs2_holder *gh)
1161 {
1162         struct gfs2_glock *gl;
1163         int error;
1164
1165         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1166         if (!error) {
1167                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1168                 gfs2_glock_put(gl);
1169         }
1170
1171         return error;
1172 }
1173
1174 /**
1175  * glock_compare - Compare two struct gfs2_glock structures for sorting
1176  * @arg_a: the first structure
1177  * @arg_b: the second structure
1178  *
1179  */
1180
1181 static int glock_compare(const void *arg_a, const void *arg_b)
1182 {
1183         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1184         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1185         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1186         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1187
1188         if (a->ln_number > b->ln_number)
1189                 return 1;
1190         if (a->ln_number < b->ln_number)
1191                 return -1;
1192         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1193         return 0;
1194 }
1195
1196 /**
1197  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1198  * @num_gh: the number of structures
1199  * @ghs: an array of struct gfs2_holder structures
1200  *
1201  * Returns: 0 on success (all glocks acquired),
1202  *          errno on failure (no glocks acquired)
1203  */
1204
1205 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1206                      struct gfs2_holder **p)
1207 {
1208         unsigned int x;
1209         int error = 0;
1210
1211         for (x = 0; x < num_gh; x++)
1212                 p[x] = &ghs[x];
1213
1214         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1215
1216         for (x = 0; x < num_gh; x++) {
1217                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1218
1219                 error = gfs2_glock_nq(p[x]);
1220                 if (error) {
1221                         while (x--)
1222                                 gfs2_glock_dq(p[x]);
1223                         break;
1224                 }
1225         }
1226
1227         return error;
1228 }
1229
1230 /**
1231  * gfs2_glock_nq_m - acquire multiple glocks
1232  * @num_gh: the number of structures
1233  * @ghs: an array of struct gfs2_holder structures
1234  *
1235  *
1236  * Returns: 0 on success (all glocks acquired),
1237  *          errno on failure (no glocks acquired)
1238  */
1239
1240 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1241 {
1242         struct gfs2_holder *tmp[4];
1243         struct gfs2_holder **pph = tmp;
1244         int error = 0;
1245
1246         switch(num_gh) {
1247         case 0:
1248                 return 0;
1249         case 1:
1250                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1251                 return gfs2_glock_nq(ghs);
1252         default:
1253                 if (num_gh <= 4)
1254                         break;
1255                 pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
1256                 if (!pph)
1257                         return -ENOMEM;
1258         }
1259
1260         error = nq_m_sync(num_gh, ghs, pph);
1261
1262         if (pph != tmp)
1263                 kfree(pph);
1264
1265         return error;
1266 }
1267
1268 /**
1269  * gfs2_glock_dq_m - release multiple glocks
1270  * @num_gh: the number of structures
1271  * @ghs: an array of struct gfs2_holder structures
1272  *
1273  */
1274
1275 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1276 {
1277         while (num_gh--)
1278                 gfs2_glock_dq(&ghs[num_gh]);
1279 }
1280
1281 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1282 {
1283         unsigned long delay = 0;
1284         unsigned long holdtime;
1285         unsigned long now = jiffies;
1286
1287         gfs2_glock_hold(gl);
1288         holdtime = gl->gl_tchange + gl->gl_hold_time;
1289         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1290             gl->gl_name.ln_type == LM_TYPE_INODE) {
1291                 if (time_before(now, holdtime))
1292                         delay = holdtime - now;
1293                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1294                         delay = gl->gl_hold_time;
1295         }
1296
1297         spin_lock(&gl->gl_spin);
1298         handle_callback(gl, state, delay, true);
1299         spin_unlock(&gl->gl_spin);
1300         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1301                 gfs2_glock_put(gl);
1302 }
1303
1304 /**
1305  * gfs2_should_freeze - Figure out if glock should be frozen
1306  * @gl: The glock in question
1307  *
1308  * Glocks are not frozen if (a) the result of the dlm operation is
1309  * an error, (b) the locking operation was an unlock operation or
1310  * (c) if there is a "noexp" flagged request anywhere in the queue
1311  *
1312  * Returns: 1 if freezing should occur, 0 otherwise
1313  */
1314
1315 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1316 {
1317         const struct gfs2_holder *gh;
1318
1319         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1320                 return 0;
1321         if (gl->gl_target == LM_ST_UNLOCKED)
1322                 return 0;
1323
1324         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1325                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1326                         continue;
1327                 if (LM_FLAG_NOEXP & gh->gh_flags)
1328                         return 0;
1329         }
1330
1331         return 1;
1332 }
1333
1334 /**
1335  * gfs2_glock_complete - Callback used by locking
1336  * @gl: Pointer to the glock
1337  * @ret: The return value from the dlm
1338  *
1339  * The gl_reply field is under the gl_spin lock so that it is ok
1340  * to use a bitfield shared with other glock state fields.
1341  */
1342
1343 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1344 {
1345         struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
1346
1347         spin_lock(&gl->gl_spin);
1348         gl->gl_reply = ret;
1349
1350         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1351                 if (gfs2_should_freeze(gl)) {
1352                         set_bit(GLF_FROZEN, &gl->gl_flags);
1353                         spin_unlock(&gl->gl_spin);
1354                         return;
1355                 }
1356         }
1357
1358         gl->gl_lockref.count++;
1359         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1360         spin_unlock(&gl->gl_spin);
1361
1362         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1363                 gfs2_glock_put(gl);
1364 }
1365
1366 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1367 {
1368         struct gfs2_glock *gla, *glb;
1369
1370         gla = list_entry(a, struct gfs2_glock, gl_lru);
1371         glb = list_entry(b, struct gfs2_glock, gl_lru);
1372
1373         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1374                 return 1;
1375         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1376                 return -1;
1377
1378         return 0;
1379 }
1380
1381 /**
1382  * gfs2_dispose_glock_lru - Demote a list of glocks
1383  * @list: The list to dispose of
1384  *
1385  * Disposing of glocks may involve disk accesses, so that here we sort
1386  * the glocks by number (i.e. disk location of the inodes) so that if
1387  * there are any such accesses, they'll be sent in order (mostly).
1388  *
1389  * Must be called under the lru_lock, but may drop and retake this
1390  * lock. While the lru_lock is dropped, entries may vanish from the
1391  * list, but no new entries will appear on the list (since it is
1392  * private)
1393  */
1394
1395 static void gfs2_dispose_glock_lru(struct list_head *list)
1396 __releases(&lru_lock)
1397 __acquires(&lru_lock)
1398 {
1399         struct gfs2_glock *gl;
1400
1401         list_sort(NULL, list, glock_cmp);
1402
1403         while(!list_empty(list)) {
1404                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1405                 list_del_init(&gl->gl_lru);
1406                 if (!spin_trylock(&gl->gl_spin)) {
1407                         list_add(&gl->gl_lru, &lru_list);
1408                         atomic_inc(&lru_count);
1409                         continue;
1410                 }
1411                 clear_bit(GLF_LRU, &gl->gl_flags);
1412                 spin_unlock(&lru_lock);
1413                 gl->gl_lockref.count++;
1414                 if (demote_ok(gl))
1415                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1416                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1417                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1418                         gl->gl_lockref.count--;
1419                 spin_unlock(&gl->gl_spin);
1420                 spin_lock(&lru_lock);
1421         }
1422 }
1423
1424 /**
1425  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1426  * @nr: The number of entries to scan
1427  *
1428  * This function selects the entries on the LRU which are able to
1429  * be demoted, and then kicks off the process by calling
1430  * gfs2_dispose_glock_lru() above.
1431  */
1432
1433 static long gfs2_scan_glock_lru(int nr)
1434 {
1435         struct gfs2_glock *gl;
1436         LIST_HEAD(skipped);
1437         LIST_HEAD(dispose);
1438         long freed = 0;
1439
1440         spin_lock(&lru_lock);
1441         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1442                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1443
1444                 /* Test for being demotable */
1445                 if (!test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1446                         list_move(&gl->gl_lru, &dispose);
1447                         atomic_dec(&lru_count);
1448                         freed++;
1449                         continue;
1450                 }
1451
1452                 list_move(&gl->gl_lru, &skipped);
1453         }
1454         list_splice(&skipped, &lru_list);
1455         if (!list_empty(&dispose))
1456                 gfs2_dispose_glock_lru(&dispose);
1457         spin_unlock(&lru_lock);
1458
1459         return freed;
1460 }
1461
1462 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1463                                             struct shrink_control *sc)
1464 {
1465         if (!(sc->gfp_mask & __GFP_FS))
1466                 return SHRINK_STOP;
1467         return gfs2_scan_glock_lru(sc->nr_to_scan);
1468 }
1469
1470 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1471                                              struct shrink_control *sc)
1472 {
1473         return vfs_pressure_ratio(atomic_read(&lru_count));
1474 }
1475
1476 static struct shrinker glock_shrinker = {
1477         .seeks = DEFAULT_SEEKS,
1478         .count_objects = gfs2_glock_shrink_count,
1479         .scan_objects = gfs2_glock_shrink_scan,
1480 };
1481
1482 /**
1483  * examine_bucket - Call a function for glock in a hash bucket
1484  * @examiner: the function
1485  * @sdp: the filesystem
1486  * @bucket: the bucket
1487  *
1488  */
1489
1490 static void examine_bucket(glock_examiner examiner, const struct gfs2_sbd *sdp,
1491                           unsigned int hash)
1492 {
1493         struct gfs2_glock *gl;
1494         struct hlist_bl_head *head = &gl_hash_table[hash];
1495         struct hlist_bl_node *pos;
1496
1497         rcu_read_lock();
1498         hlist_bl_for_each_entry_rcu(gl, pos, head, gl_list) {
1499                 if ((gl->gl_sbd == sdp) && lockref_get_not_dead(&gl->gl_lockref))
1500                         examiner(gl);
1501         }
1502         rcu_read_unlock();
1503         cond_resched();
1504 }
1505
1506 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1507 {
1508         unsigned x;
1509
1510         for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
1511                 examine_bucket(examiner, sdp, x);
1512 }
1513
1514
1515 /**
1516  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1517  * @gl: The glock to thaw
1518  *
1519  */
1520
1521 static void thaw_glock(struct gfs2_glock *gl)
1522 {
1523         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1524                 goto out;
1525         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1526         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0) {
1527 out:
1528                 gfs2_glock_put(gl);
1529         }
1530 }
1531
1532 /**
1533  * clear_glock - look at a glock and see if we can free it from glock cache
1534  * @gl: the glock to look at
1535  *
1536  */
1537
1538 static void clear_glock(struct gfs2_glock *gl)
1539 {
1540         gfs2_glock_remove_from_lru(gl);
1541
1542         spin_lock(&gl->gl_spin);
1543         if (gl->gl_state != LM_ST_UNLOCKED)
1544                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1545         spin_unlock(&gl->gl_spin);
1546         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1547                 gfs2_glock_put(gl);
1548 }
1549
1550 /**
1551  * gfs2_glock_thaw - Thaw any frozen glocks
1552  * @sdp: The super block
1553  *
1554  */
1555
1556 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1557 {
1558         glock_hash_walk(thaw_glock, sdp);
1559 }
1560
1561 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1562 {
1563         spin_lock(&gl->gl_spin);
1564         gfs2_dump_glock(seq, gl);
1565         spin_unlock(&gl->gl_spin);
1566 }
1567
1568 static void dump_glock_func(struct gfs2_glock *gl)
1569 {
1570         dump_glock(NULL, gl);
1571 }
1572
1573 /**
1574  * gfs2_gl_hash_clear - Empty out the glock hash table
1575  * @sdp: the filesystem
1576  * @wait: wait until it's all gone
1577  *
1578  * Called when unmounting the filesystem.
1579  */
1580
1581 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1582 {
1583         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1584         flush_workqueue(glock_workqueue);
1585         glock_hash_walk(clear_glock, sdp);
1586         flush_workqueue(glock_workqueue);
1587         wait_event(sdp->sd_glock_wait, atomic_read(&sdp->sd_glock_disposal) == 0);
1588         glock_hash_walk(dump_glock_func, sdp);
1589 }
1590
1591 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1592 {
1593         struct gfs2_glock *gl = ip->i_gl;
1594         int ret;
1595
1596         ret = gfs2_truncatei_resume(ip);
1597         gfs2_assert_withdraw(gl->gl_sbd, ret == 0);
1598
1599         spin_lock(&gl->gl_spin);
1600         clear_bit(GLF_LOCK, &gl->gl_flags);
1601         run_queue(gl, 1);
1602         spin_unlock(&gl->gl_spin);
1603 }
1604
1605 static const char *state2str(unsigned state)
1606 {
1607         switch(state) {
1608         case LM_ST_UNLOCKED:
1609                 return "UN";
1610         case LM_ST_SHARED:
1611                 return "SH";
1612         case LM_ST_DEFERRED:
1613                 return "DF";
1614         case LM_ST_EXCLUSIVE:
1615                 return "EX";
1616         }
1617         return "??";
1618 }
1619
1620 static const char *hflags2str(char *buf, unsigned flags, unsigned long iflags)
1621 {
1622         char *p = buf;
1623         if (flags & LM_FLAG_TRY)
1624                 *p++ = 't';
1625         if (flags & LM_FLAG_TRY_1CB)
1626                 *p++ = 'T';
1627         if (flags & LM_FLAG_NOEXP)
1628                 *p++ = 'e';
1629         if (flags & LM_FLAG_ANY)
1630                 *p++ = 'A';
1631         if (flags & LM_FLAG_PRIORITY)
1632                 *p++ = 'p';
1633         if (flags & GL_ASYNC)
1634                 *p++ = 'a';
1635         if (flags & GL_EXACT)
1636                 *p++ = 'E';
1637         if (flags & GL_NOCACHE)
1638                 *p++ = 'c';
1639         if (test_bit(HIF_HOLDER, &iflags))
1640                 *p++ = 'H';
1641         if (test_bit(HIF_WAIT, &iflags))
1642                 *p++ = 'W';
1643         if (test_bit(HIF_FIRST, &iflags))
1644                 *p++ = 'F';
1645         *p = 0;
1646         return buf;
1647 }
1648
1649 /**
1650  * dump_holder - print information about a glock holder
1651  * @seq: the seq_file struct
1652  * @gh: the glock holder
1653  *
1654  */
1655
1656 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1657 {
1658         struct task_struct *gh_owner = NULL;
1659         char flags_buf[32];
1660
1661         rcu_read_lock();
1662         if (gh->gh_owner_pid)
1663                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1664         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1665                        state2str(gh->gh_state),
1666                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1667                        gh->gh_error,
1668                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1669                        gh_owner ? gh_owner->comm : "(ended)",
1670                        (void *)gh->gh_ip);
1671         rcu_read_unlock();
1672 }
1673
1674 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1675 {
1676         const unsigned long *gflags = &gl->gl_flags;
1677         char *p = buf;
1678
1679         if (test_bit(GLF_LOCK, gflags))
1680                 *p++ = 'l';
1681         if (test_bit(GLF_DEMOTE, gflags))
1682                 *p++ = 'D';
1683         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1684                 *p++ = 'd';
1685         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1686                 *p++ = 'p';
1687         if (test_bit(GLF_DIRTY, gflags))
1688                 *p++ = 'y';
1689         if (test_bit(GLF_LFLUSH, gflags))
1690                 *p++ = 'f';
1691         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1692                 *p++ = 'i';
1693         if (test_bit(GLF_REPLY_PENDING, gflags))
1694                 *p++ = 'r';
1695         if (test_bit(GLF_INITIAL, gflags))
1696                 *p++ = 'I';
1697         if (test_bit(GLF_FROZEN, gflags))
1698                 *p++ = 'F';
1699         if (test_bit(GLF_QUEUED, gflags))
1700                 *p++ = 'q';
1701         if (test_bit(GLF_LRU, gflags))
1702                 *p++ = 'L';
1703         if (gl->gl_object)
1704                 *p++ = 'o';
1705         if (test_bit(GLF_BLOCKING, gflags))
1706                 *p++ = 'b';
1707         *p = 0;
1708         return buf;
1709 }
1710
1711 /**
1712  * gfs2_dump_glock - print information about a glock
1713  * @seq: The seq_file struct
1714  * @gl: the glock
1715  *
1716  * The file format is as follows:
1717  * One line per object, capital letters are used to indicate objects
1718  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1719  * other objects are indented by a single space and follow the glock to
1720  * which they are related. Fields are indicated by lower case letters
1721  * followed by a colon and the field value, except for strings which are in
1722  * [] so that its possible to see if they are composed of spaces for
1723  * example. The field's are n = number (id of the object), f = flags,
1724  * t = type, s = state, r = refcount, e = error, p = pid.
1725  *
1726  */
1727
1728 void gfs2_dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1729 {
1730         const struct gfs2_glock_operations *glops = gl->gl_ops;
1731         unsigned long long dtime;
1732         const struct gfs2_holder *gh;
1733         char gflags_buf[32];
1734
1735         dtime = jiffies - gl->gl_demote_time;
1736         dtime *= 1000000/HZ; /* demote time in uSec */
1737         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1738                 dtime = 0;
1739         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d v:%d r:%d m:%ld\n",
1740                   state2str(gl->gl_state),
1741                   gl->gl_name.ln_type,
1742                   (unsigned long long)gl->gl_name.ln_number,
1743                   gflags2str(gflags_buf, gl),
1744                   state2str(gl->gl_target),
1745                   state2str(gl->gl_demote_state), dtime,
1746                   atomic_read(&gl->gl_ail_count),
1747                   atomic_read(&gl->gl_revokes),
1748                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1749
1750         list_for_each_entry(gh, &gl->gl_holders, gh_list)
1751                 dump_holder(seq, gh);
1752
1753         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1754                 glops->go_dump(seq, gl);
1755 }
1756
1757 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
1758 {
1759         struct gfs2_glock *gl = iter_ptr;
1760
1761         seq_printf(seq, "G: n:%u/%llx rtt:%lld/%lld rttb:%lld/%lld irt:%lld/%lld dcnt: %lld qcnt: %lld\n",
1762                    gl->gl_name.ln_type,
1763                    (unsigned long long)gl->gl_name.ln_number,
1764                    (long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
1765                    (long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
1766                    (long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
1767                    (long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
1768                    (long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
1769                    (long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
1770                    (long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
1771                    (long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
1772         return 0;
1773 }
1774
1775 static const char *gfs2_gltype[] = {
1776         "type",
1777         "reserved",
1778         "nondisk",
1779         "inode",
1780         "rgrp",
1781         "meta",
1782         "iopen",
1783         "flock",
1784         "plock",
1785         "quota",
1786         "journal",
1787 };
1788
1789 static const char *gfs2_stype[] = {
1790         [GFS2_LKS_SRTT]         = "srtt",
1791         [GFS2_LKS_SRTTVAR]      = "srttvar",
1792         [GFS2_LKS_SRTTB]        = "srttb",
1793         [GFS2_LKS_SRTTVARB]     = "srttvarb",
1794         [GFS2_LKS_SIRT]         = "sirt",
1795         [GFS2_LKS_SIRTVAR]      = "sirtvar",
1796         [GFS2_LKS_DCOUNT]       = "dlm",
1797         [GFS2_LKS_QCOUNT]       = "queue",
1798 };
1799
1800 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
1801
1802 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
1803 {
1804         struct gfs2_glock_iter *gi = seq->private;
1805         struct gfs2_sbd *sdp = gi->sdp;
1806         unsigned index = gi->hash >> 3;
1807         unsigned subindex = gi->hash & 0x07;
1808         s64 value;
1809         int i;
1810
1811         if (index == 0 && subindex != 0)
1812                 return 0;
1813
1814         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
1815                    (index == 0) ? "cpu": gfs2_stype[subindex]);
1816
1817         for_each_possible_cpu(i) {
1818                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
1819                 if (index == 0) {
1820                         value = i;
1821                 } else {
1822                         value = lkstats->lkstats[index - 1].stats[subindex];
1823                 }
1824                 seq_printf(seq, " %15lld", (long long)value);
1825         }
1826         seq_putc(seq, '\n');
1827         return 0;
1828 }
1829
1830 int __init gfs2_glock_init(void)
1831 {
1832         unsigned i;
1833         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
1834                 INIT_HLIST_BL_HEAD(&gl_hash_table[i]);
1835         }
1836
1837         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1838                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1839         if (!glock_workqueue)
1840                 return -ENOMEM;
1841         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1842                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1843                                                 0);
1844         if (!gfs2_delete_workqueue) {
1845                 destroy_workqueue(glock_workqueue);
1846                 return -ENOMEM;
1847         }
1848
1849         register_shrinker(&glock_shrinker);
1850
1851         return 0;
1852 }
1853
1854 void gfs2_glock_exit(void)
1855 {
1856         unregister_shrinker(&glock_shrinker);
1857         destroy_workqueue(glock_workqueue);
1858         destroy_workqueue(gfs2_delete_workqueue);
1859 }
1860
1861 static inline struct gfs2_glock *glock_hash_chain(unsigned hash)
1862 {
1863         return hlist_bl_entry(hlist_bl_first_rcu(&gl_hash_table[hash]),
1864                               struct gfs2_glock, gl_list);
1865 }
1866
1867 static inline struct gfs2_glock *glock_hash_next(struct gfs2_glock *gl)
1868 {
1869         return hlist_bl_entry(rcu_dereference(gl->gl_list.next),
1870                               struct gfs2_glock, gl_list);
1871 }
1872
1873 static int gfs2_glock_iter_next(struct gfs2_glock_iter *gi)
1874 {
1875         struct gfs2_glock *gl;
1876
1877         do {
1878                 gl = gi->gl;
1879                 if (gl) {
1880                         gi->gl = glock_hash_next(gl);
1881                         gi->nhash++;
1882                 } else {
1883                         if (gi->hash >= GFS2_GL_HASH_SIZE) {
1884                                 rcu_read_unlock();
1885                                 return 1;
1886                         }
1887                         gi->gl = glock_hash_chain(gi->hash);
1888                         gi->nhash = 0;
1889                 }
1890                 while (gi->gl == NULL) {
1891                         gi->hash++;
1892                         if (gi->hash >= GFS2_GL_HASH_SIZE) {
1893                                 rcu_read_unlock();
1894                                 return 1;
1895                         }
1896                         gi->gl = glock_hash_chain(gi->hash);
1897                         gi->nhash = 0;
1898                 }
1899         /* Skip entries for other sb and dead entries */
1900         } while (gi->sdp != gi->gl->gl_sbd ||
1901                  __lockref_is_dead(&gi->gl->gl_lockref));
1902
1903         return 0;
1904 }
1905
1906 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1907 {
1908         struct gfs2_glock_iter *gi = seq->private;
1909         loff_t n = *pos;
1910
1911         if (gi->last_pos <= *pos)
1912                 n = gi->nhash + (*pos - gi->last_pos);
1913         else
1914                 gi->hash = 0;
1915
1916         gi->nhash = 0;
1917         rcu_read_lock();
1918
1919         do {
1920                 if (gfs2_glock_iter_next(gi))
1921                         return NULL;
1922         } while (n--);
1923
1924         gi->last_pos = *pos;
1925         return gi->gl;
1926 }
1927
1928 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1929                                  loff_t *pos)
1930 {
1931         struct gfs2_glock_iter *gi = seq->private;
1932
1933         (*pos)++;
1934         gi->last_pos = *pos;
1935         if (gfs2_glock_iter_next(gi))
1936                 return NULL;
1937
1938         return gi->gl;
1939 }
1940
1941 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1942 {
1943         struct gfs2_glock_iter *gi = seq->private;
1944
1945         if (gi->gl)
1946                 rcu_read_unlock();
1947         gi->gl = NULL;
1948 }
1949
1950 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
1951 {
1952         dump_glock(seq, iter_ptr);
1953         return 0;
1954 }
1955
1956 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
1957 {
1958         struct gfs2_glock_iter *gi = seq->private;
1959
1960         gi->hash = *pos;
1961         if (*pos >= GFS2_NR_SBSTATS)
1962                 return NULL;
1963         preempt_disable();
1964         return SEQ_START_TOKEN;
1965 }
1966
1967 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
1968                                    loff_t *pos)
1969 {
1970         struct gfs2_glock_iter *gi = seq->private;
1971         (*pos)++;
1972         gi->hash++;
1973         if (gi->hash >= GFS2_NR_SBSTATS) {
1974                 preempt_enable();
1975                 return NULL;
1976         }
1977         return SEQ_START_TOKEN;
1978 }
1979
1980 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
1981 {
1982         preempt_enable();
1983 }
1984
1985 static const struct seq_operations gfs2_glock_seq_ops = {
1986         .start = gfs2_glock_seq_start,
1987         .next  = gfs2_glock_seq_next,
1988         .stop  = gfs2_glock_seq_stop,
1989         .show  = gfs2_glock_seq_show,
1990 };
1991
1992 static const struct seq_operations gfs2_glstats_seq_ops = {
1993         .start = gfs2_glock_seq_start,
1994         .next  = gfs2_glock_seq_next,
1995         .stop  = gfs2_glock_seq_stop,
1996         .show  = gfs2_glstats_seq_show,
1997 };
1998
1999 static const struct seq_operations gfs2_sbstats_seq_ops = {
2000         .start = gfs2_sbstats_seq_start,
2001         .next  = gfs2_sbstats_seq_next,
2002         .stop  = gfs2_sbstats_seq_stop,
2003         .show  = gfs2_sbstats_seq_show,
2004 };
2005
2006 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2007
2008 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2009 {
2010         int ret = seq_open_private(file, &gfs2_glock_seq_ops,
2011                                    sizeof(struct gfs2_glock_iter));
2012         if (ret == 0) {
2013                 struct seq_file *seq = file->private_data;
2014                 struct gfs2_glock_iter *gi = seq->private;
2015                 gi->sdp = inode->i_private;
2016                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2017                 if (seq->buf)
2018                         seq->size = GFS2_SEQ_GOODSIZE;
2019         }
2020         return ret;
2021 }
2022
2023 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2024 {
2025         int ret = seq_open_private(file, &gfs2_glstats_seq_ops,
2026                                    sizeof(struct gfs2_glock_iter));
2027         if (ret == 0) {
2028                 struct seq_file *seq = file->private_data;
2029                 struct gfs2_glock_iter *gi = seq->private;
2030                 gi->sdp = inode->i_private;
2031                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2032                 if (seq->buf)
2033                         seq->size = GFS2_SEQ_GOODSIZE;
2034         }
2035         return ret;
2036 }
2037
2038 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2039 {
2040         int ret = seq_open_private(file, &gfs2_sbstats_seq_ops,
2041                                    sizeof(struct gfs2_glock_iter));
2042         if (ret == 0) {
2043                 struct seq_file *seq = file->private_data;
2044                 struct gfs2_glock_iter *gi = seq->private;
2045                 gi->sdp = inode->i_private;
2046         }
2047         return ret;
2048 }
2049
2050 static const struct file_operations gfs2_glocks_fops = {
2051         .owner   = THIS_MODULE,
2052         .open    = gfs2_glocks_open,
2053         .read    = seq_read,
2054         .llseek  = seq_lseek,
2055         .release = seq_release_private,
2056 };
2057
2058 static const struct file_operations gfs2_glstats_fops = {
2059         .owner   = THIS_MODULE,
2060         .open    = gfs2_glstats_open,
2061         .read    = seq_read,
2062         .llseek  = seq_lseek,
2063         .release = seq_release_private,
2064 };
2065
2066 static const struct file_operations gfs2_sbstats_fops = {
2067         .owner   = THIS_MODULE,
2068         .open    = gfs2_sbstats_open,
2069         .read    = seq_read,
2070         .llseek  = seq_lseek,
2071         .release = seq_release_private,
2072 };
2073
2074 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2075 {
2076         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2077         if (!sdp->debugfs_dir)
2078                 return -ENOMEM;
2079         sdp->debugfs_dentry_glocks = debugfs_create_file("glocks",
2080                                                          S_IFREG | S_IRUGO,
2081                                                          sdp->debugfs_dir, sdp,
2082                                                          &gfs2_glocks_fops);
2083         if (!sdp->debugfs_dentry_glocks)
2084                 goto fail;
2085
2086         sdp->debugfs_dentry_glstats = debugfs_create_file("glstats",
2087                                                         S_IFREG | S_IRUGO,
2088                                                         sdp->debugfs_dir, sdp,
2089                                                         &gfs2_glstats_fops);
2090         if (!sdp->debugfs_dentry_glstats)
2091                 goto fail;
2092
2093         sdp->debugfs_dentry_sbstats = debugfs_create_file("sbstats",
2094                                                         S_IFREG | S_IRUGO,
2095                                                         sdp->debugfs_dir, sdp,
2096                                                         &gfs2_sbstats_fops);
2097         if (!sdp->debugfs_dentry_sbstats)
2098                 goto fail;
2099
2100         return 0;
2101 fail:
2102         gfs2_delete_debugfs_file(sdp);
2103         return -ENOMEM;
2104 }
2105
2106 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2107 {
2108         if (sdp->debugfs_dir) {
2109                 if (sdp->debugfs_dentry_glocks) {
2110                         debugfs_remove(sdp->debugfs_dentry_glocks);
2111                         sdp->debugfs_dentry_glocks = NULL;
2112                 }
2113                 if (sdp->debugfs_dentry_glstats) {
2114                         debugfs_remove(sdp->debugfs_dentry_glstats);
2115                         sdp->debugfs_dentry_glstats = NULL;
2116                 }
2117                 if (sdp->debugfs_dentry_sbstats) {
2118                         debugfs_remove(sdp->debugfs_dentry_sbstats);
2119                         sdp->debugfs_dentry_sbstats = NULL;
2120                 }
2121                 debugfs_remove(sdp->debugfs_dir);
2122                 sdp->debugfs_dir = NULL;
2123         }
2124 }
2125
2126 int gfs2_register_debugfs(void)
2127 {
2128         gfs2_root = debugfs_create_dir("gfs2", NULL);
2129         return gfs2_root ? 0 : -ENOMEM;
2130 }
2131
2132 void gfs2_unregister_debugfs(void)
2133 {
2134         debugfs_remove(gfs2_root);
2135         gfs2_root = NULL;
2136 }