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/hash.h>
  19 #include <linux/jhash.h>
  20 #include <linux/kallsyms.h>
  21 #include <linux/gfs2_ondisk.h>
  22 #include <linux/list.h>
  23 #include <linux/wait.h>
  24 #include <linux/module.h>
  25 #include <linux/uaccess.h>
  26 #include <linux/seq_file.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/kthread.h>
  29 #include <linux/freezer.h>
  30 #include <linux/workqueue.h>
  31 #include <linux/jiffies.h>
  32 #include <linux/rcupdate.h>
  33 #include <linux/rculist_bl.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/percpu.h>
  36 #include <linux/list_sort.h>
  37 #include <linux/lockref.h>
  38 #include <linux/rhashtable.h>
  39
  40 #include "gfs2.h"
  41 #include "incore.h"
  42 #include "glock.h"
  43 #include "glops.h"
  44 #include "inode.h"
  45 #include "lops.h"
  46 #include "meta_io.h"
  47 #include "quota.h"
  48 #include "super.h"
  49 #include "util.h"
  50 #include "bmap.h"
  51 #define CREATE_TRACE_POINTS
  52 #include "trace_gfs2.h"
  53
  54 struct gfs2_glock_iter {
  55         struct gfs2_sbd *sdp;           /* incore superblock           */
  56         struct rhashtable_iter hti;     /* rhashtable iterator         */
  57         struct gfs2_glock *gl;          /* current glock struct        */
  58         loff_t last_pos;                /* last position               */
  59 };
  60
  61 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  62
  63 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  64
  65 static struct dentry *gfs2_root;
  66 static struct workqueue_struct *glock_workqueue;
  67 struct workqueue_struct *gfs2_delete_workqueue;
  68 static LIST_HEAD(lru_list);
  69 static atomic_t lru_count = ATOMIC_INIT(0);
  70 static DEFINE_SPINLOCK(lru_lock);
  71
  72 #define GFS2_GL_HASH_SHIFT      15
  73 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
  74
  75 static const struct rhashtable_params ht_parms = {
  76         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
  77         .key_len = offsetofend(struct lm_lockname, ln_type),
  78         .key_offset = offsetof(struct gfs2_glock, gl_name),
  79         .head_offset = offsetof(struct gfs2_glock, gl_node),
  80 };
  81
  82 static struct rhashtable gl_hash_table;
  83
  84 #define GLOCK_WAIT_TABLE_BITS 12
  85 #define GLOCK_WAIT_TABLE_SIZE (1 << GLOCK_WAIT_TABLE_BITS)
  86 static wait_queue_head_t glock_wait_table[GLOCK_WAIT_TABLE_SIZE] __cacheline_aligned;
  87
  88 struct wait_glock_queue {
  89         struct lm_lockname *name;
  90         wait_queue_entry_t wait;
  91 };
  92
  93 static int glock_wake_function(wait_queue_entry_t *wait, unsigned int mode,
  94                                int sync, void *key)
  95 {
  96         struct wait_glock_queue *wait_glock =
  97                 container_of(wait, struct wait_glock_queue, wait);
  98         struct lm_lockname *wait_name = wait_glock->name;
  99         struct lm_lockname *wake_name = key;
 100
 101         if (wake_name->ln_sbd != wait_name->ln_sbd ||
 102             wake_name->ln_number != wait_name->ln_number ||
 103             wake_name->ln_type != wait_name->ln_type)
 104                 return 0;
 105         return autoremove_wake_function(wait, mode, sync, key);
 106 }
 107
 108 static wait_queue_head_t *glock_waitqueue(struct lm_lockname *name)
 109 {
 110         u32 hash = jhash2((u32 *)name, sizeof(*name) / 4, 0);
 111
 112         return glock_wait_table + hash_32(hash, GLOCK_WAIT_TABLE_BITS);
 113 }
 114
 115 /**
 116  * wake_up_glock  -  Wake up waiters on a glock
 117  * @gl: the glock
 118  */
 119 static void wake_up_glock(struct gfs2_glock *gl)
 120 {
 121         wait_queue_head_t *wq = glock_waitqueue(&gl->gl_name);
 122
 123         if (waitqueue_active(wq))
 124                 __wake_up(wq, TASK_NORMAL, 1, &gl->gl_name);
 125 }
 126
 127 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 128 {
 129         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 130
 131         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
 132                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 133         } else {
 134                 kfree(gl->gl_lksb.sb_lvbptr);
 135                 kmem_cache_free(gfs2_glock_cachep, gl);
 136         }
 137 }
 138
 139 void gfs2_glock_free(struct gfs2_glock *gl)
 140 {
 141         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 142
 143         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 144         smp_mb();
 145         wake_up_glock(gl);
 146         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 147         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 148                 wake_up(&sdp->sd_glock_wait);
 149 }
 150
 151 /**
 152  * gfs2_glock_hold() - increment reference count on glock
 153  * @gl: The glock to hold
 154  *
 155  */
 156
 157 void gfs2_glock_hold(struct gfs2_glock *gl)
 158 {
 159         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 160         lockref_get(&gl->gl_lockref);
 161 }
 162
 163 /**
 164  * demote_ok - Check to see if it's ok to unlock a glock
 165  * @gl: the glock
 166  *
 167  * Returns: 1 if it's ok
 168  */
 169
 170 static int demote_ok(const struct gfs2_glock *gl)
 171 {
 172         const struct gfs2_glock_operations *glops = gl->gl_ops;
 173
 174         if (gl->gl_state == LM_ST_UNLOCKED)
 175                 return 0;
 176         if (!list_empty(&gl->gl_holders))
 177                 return 0;
 178         if (glops->go_demote_ok)
 179                 return glops->go_demote_ok(gl);
 180         return 1;
 181 }
 182
 183
 184 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 185 {
 186         spin_lock(&lru_lock);
 187
 188         if (!list_empty(&gl->gl_lru))
 189                 list_del_init(&gl->gl_lru);
 190         else
 191                 atomic_inc(&lru_count);
 192
 193         list_add_tail(&gl->gl_lru, &lru_list);
 194         set_bit(GLF_LRU, &gl->gl_flags);
 195         spin_unlock(&lru_lock);
 196 }
 197
 198 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 199 {
 200         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 201                 return;
 202
 203         spin_lock(&lru_lock);
 204         if (!list_empty(&gl->gl_lru)) {
 205                 list_del_init(&gl->gl_lru);
 206                 atomic_dec(&lru_count);
 207                 clear_bit(GLF_LRU, &gl->gl_flags);
 208         }
 209         spin_unlock(&lru_lock);
 210 }
 211
 212 /*
 213  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 214  * work queue.
 215  */
 216 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 217         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 218                 /*
 219                  * We are holding the lockref spinlock, and the work was still
 220                  * queued above.  The queued work (glock_work_func) takes that
 221                  * spinlock before dropping its glock reference(s), so it
 222                  * cannot have dropped them in the meantime.
 223                  */
 224                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 225                 gl->gl_lockref.count--;
 226         }
 227 }
 228
 229 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 230         spin_lock(&gl->gl_lockref.lock);
 231         __gfs2_glock_queue_work(gl, delay);
 232         spin_unlock(&gl->gl_lockref.lock);
 233 }
 234
 235 static void __gfs2_glock_put(struct gfs2_glock *gl)
 236 {
 237         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 238         struct address_space *mapping = gfs2_glock2aspace(gl);
 239
 240         lockref_mark_dead(&gl->gl_lockref);
 241
 242         gfs2_glock_remove_from_lru(gl);
 243         spin_unlock(&gl->gl_lockref.lock);
 244         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 245         GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
 246         trace_gfs2_glock_put(gl);
 247         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 248 }
 249
 250 /*
 251  * Cause the glock to be put in work queue context.
 252  */
 253 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 254 {
 255         gfs2_glock_queue_work(gl, 0);
 256 }
 257
 258 /**
 259  * gfs2_glock_put() - Decrement reference count on glock
 260  * @gl: The glock to put
 261  *
 262  */
 263
 264 void gfs2_glock_put(struct gfs2_glock *gl)
 265 {
 266         if (lockref_put_or_lock(&gl->gl_lockref))
 267                 return;
 268
 269         __gfs2_glock_put(gl);
 270 }
 271
 272 /**
 273  * may_grant - check if its ok to grant a new lock
 274  * @gl: The glock
 275  * @gh: The lock request which we wish to grant
 276  *
 277  * Returns: true if its ok to grant the lock
 278  */
 279
 280 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 281 {
 282         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 283         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 284              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 285                 return 0;
 286         if (gl->gl_state == gh->gh_state)
 287                 return 1;
 288         if (gh->gh_flags & GL_EXACT)
 289                 return 0;
 290         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 291                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 292                         return 1;
 293                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 294                         return 1;
 295         }
 296         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 297                 return 1;
 298         return 0;
 299 }
 300
 301 static void gfs2_holder_wake(struct gfs2_holder *gh)
 302 {
 303         clear_bit(HIF_WAIT, &gh->gh_iflags);
 304         smp_mb__after_atomic();
 305         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 306 }
 307
 308 /**
 309  * do_error - Something unexpected has happened during a lock request
 310  *
 311  */
 312
 313 static void do_error(struct gfs2_glock *gl, const int ret)
 314 {
 315         struct gfs2_holder *gh, *tmp;
 316
 317         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 318                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 319                         continue;
 320                 if (ret & LM_OUT_ERROR)
 321                         gh->gh_error = -EIO;
 322                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 323                         gh->gh_error = GLR_TRYFAILED;
 324                 else
 325                         continue;
 326                 list_del_init(&gh->gh_list);
 327                 trace_gfs2_glock_queue(gh, 0);
 328                 gfs2_holder_wake(gh);
 329         }
 330 }
 331
 332 /**
 333  * do_promote - promote as many requests as possible on the current queue
 334  * @gl: The glock
 335  *
 336  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 337  *          if a type specific operation is underway.
 338  */
 339
 340 static int do_promote(struct gfs2_glock *gl)
 341 __releases(&gl->gl_lockref.lock)
 342 __acquires(&gl->gl_lockref.lock)
 343 {
 344         const struct gfs2_glock_operations *glops = gl->gl_ops;
 345         struct gfs2_holder *gh, *tmp;
 346         int ret;
 347
 348 restart:
 349         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 350                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 351                         continue;
 352                 if (may_grant(gl, gh)) {
 353                         if (gh->gh_list.prev == &gl->gl_holders &&
 354                             glops->go_lock) {
 355                                 spin_unlock(&gl->gl_lockref.lock);
 356                                 /* FIXME: eliminate this eventually */
 357                                 ret = glops->go_lock(gh);
 358                                 spin_lock(&gl->gl_lockref.lock);
 359                                 if (ret) {
 360                                         if (ret == 1)
 361                                                 return 2;
 362                                         gh->gh_error = ret;
 363                                         list_del_init(&gh->gh_list);
 364                                         trace_gfs2_glock_queue(gh, 0);
 365                                         gfs2_holder_wake(gh);
 366                                         goto restart;
 367                                 }
 368                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 369                                 trace_gfs2_promote(gh, 1);
 370                                 gfs2_holder_wake(gh);
 371                                 goto restart;
 372                         }
 373                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 374                         trace_gfs2_promote(gh, 0);
 375                         gfs2_holder_wake(gh);
 376                         continue;
 377                 }
 378                 if (gh->gh_list.prev == &gl->gl_holders)
 379                         return 1;
 380                 do_error(gl, 0);
 381                 break;
 382         }
 383         return 0;
 384 }
 385
 386 /**
 387  * find_first_waiter - find the first gh that's waiting for the glock
 388  * @gl: the glock
 389  */
 390
 391 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 392 {
 393         struct gfs2_holder *gh;
 394
 395         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 396                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 397                         return gh;
 398         }
 399         return NULL;
 400 }
 401
 402 /**
 403  * state_change - record that the glock is now in a different state
 404  * @gl: the glock
 405  * @new_state the new state
 406  *
 407  */
 408
 409 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 410 {
 411         int held1, held2;
 412
 413         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 414         held2 = (new_state != LM_ST_UNLOCKED);
 415
 416         if (held1 != held2) {
 417                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 418                 if (held2)
 419                         gl->gl_lockref.count++;
 420                 else
 421                         gl->gl_lockref.count--;
 422         }
 423         if (held1 && held2 && list_empty(&gl->gl_holders))
 424                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 425
 426         if (new_state != gl->gl_target)
 427                 /* shorten our minimum hold time */
 428                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 429                                        GL_GLOCK_MIN_HOLD);
 430         gl->gl_state = new_state;
 431         gl->gl_tchange = jiffies;
 432 }
 433
 434 static void gfs2_demote_wake(struct gfs2_glock *gl)
 435 {
 436         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 437         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 438         smp_mb__after_atomic();
 439         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 440 }
 441
 442 /**
 443  * finish_xmote - The DLM has replied to one of our lock requests
 444  * @gl: The glock
 445  * @ret: The status from the DLM
 446  *
 447  */
 448
 449 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 450 {
 451         const struct gfs2_glock_operations *glops = gl->gl_ops;
 452         struct gfs2_holder *gh;
 453         unsigned state = ret & LM_OUT_ST_MASK;
 454         int rv;
 455
 456         spin_lock(&gl->gl_lockref.lock);
 457         trace_gfs2_glock_state_change(gl, state);
 458         state_change(gl, state);
 459         gh = find_first_waiter(gl);
 460
 461         /* Demote to UN request arrived during demote to SH or DF */
 462         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 463             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 464                 gl->gl_target = LM_ST_UNLOCKED;
 465
 466         /* Check for state != intended state */
 467         if (unlikely(state != gl->gl_target)) {
 468                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 469                         /* move to back of queue and try next entry */
 470                         if (ret & LM_OUT_CANCELED) {
 471                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 472                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 473                                 gh = find_first_waiter(gl);
 474                                 gl->gl_target = gh->gh_state;
 475                                 goto retry;
 476                         }
 477                         /* Some error or failed "try lock" - report it */
 478                         if ((ret & LM_OUT_ERROR) ||
 479                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 480                                 gl->gl_target = gl->gl_state;
 481                                 do_error(gl, ret);
 482                                 goto out;
 483                         }
 484                 }
 485                 switch(state) {
 486                 /* Unlocked due to conversion deadlock, try again */
 487                 case LM_ST_UNLOCKED:
 488 retry:
 489                         do_xmote(gl, gh, gl->gl_target);
 490                         break;
 491                 /* Conversion fails, unlock and try again */
 492                 case LM_ST_SHARED:
 493                 case LM_ST_DEFERRED:
 494                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 495                         break;
 496                 default: /* Everything else */
 497                         pr_err("wanted %u got %u\n", gl->gl_target, state);
 498                         GLOCK_BUG_ON(gl, 1);
 499                 }
 500                 spin_unlock(&gl->gl_lockref.lock);
 501                 return;
 502         }
 503
 504         /* Fast path - we got what we asked for */
 505         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 506                 gfs2_demote_wake(gl);
 507         if (state != LM_ST_UNLOCKED) {
 508                 if (glops->go_xmote_bh) {
 509                         spin_unlock(&gl->gl_lockref.lock);
 510                         rv = glops->go_xmote_bh(gl, gh);
 511                         spin_lock(&gl->gl_lockref.lock);
 512                         if (rv) {
 513                                 do_error(gl, rv);
 514                                 goto out;
 515                         }
 516                 }
 517                 rv = do_promote(gl);
 518                 if (rv == 2)
 519                         goto out_locked;
 520         }
 521 out:
 522         clear_bit(GLF_LOCK, &gl->gl_flags);
 523 out_locked:
 524         spin_unlock(&gl->gl_lockref.lock);
 525 }
 526
 527 /**
 528  * do_xmote - Calls the DLM to change the state of a lock
 529  * @gl: The lock state
 530  * @gh: The holder (only for promotes)
 531  * @target: The target lock state
 532  *
 533  */
 534
 535 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 536 __releases(&gl->gl_lockref.lock)
 537 __acquires(&gl->gl_lockref.lock)
 538 {
 539         const struct gfs2_glock_operations *glops = gl->gl_ops;
 540         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 541         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 542         int ret;
 543
 544         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)) &&
 545             target != LM_ST_UNLOCKED)
 546                 return;
 547         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 548                       LM_FLAG_PRIORITY);
 549         GLOCK_BUG_ON(gl, gl->gl_state == target);
 550         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 551         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 552             glops->go_inval) {
 553                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 554                 do_error(gl, 0); /* Fail queued try locks */
 555         }
 556         gl->gl_req = target;
 557         set_bit(GLF_BLOCKING, &gl->gl_flags);
 558         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 559             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 560             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 561                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 562         spin_unlock(&gl->gl_lockref.lock);
 563         if (glops->go_sync)
 564                 glops->go_sync(gl);
 565         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 566                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 567         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 568
 569         gfs2_glock_hold(gl);
 570         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 571                 /* lock_dlm */
 572                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 573                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 574                     target == LM_ST_UNLOCKED &&
 575                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 576                         finish_xmote(gl, target);
 577                         gfs2_glock_queue_work(gl, 0);
 578                 }
 579                 else if (ret) {
 580                         pr_err("lm_lock ret %d\n", ret);
 581                         GLOCK_BUG_ON(gl, !test_bit(SDF_SHUTDOWN,
 582                                                    &sdp->sd_flags));
 583                 }
 584         } else { /* lock_nolock */
 585                 finish_xmote(gl, target);
 586                 gfs2_glock_queue_work(gl, 0);
 587         }
 588
 589         spin_lock(&gl->gl_lockref.lock);
 590 }
 591
 592 /**
 593  * find_first_holder - find the first "holder" gh
 594  * @gl: the glock
 595  */
 596
 597 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 598 {
 599         struct gfs2_holder *gh;
 600
 601         if (!list_empty(&gl->gl_holders)) {
 602                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 603                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 604                         return gh;
 605         }
 606         return NULL;
 607 }
 608
 609 /**
 610  * run_queue - do all outstanding tasks related to a glock
 611  * @gl: The glock in question
 612  * @nonblock: True if we must not block in run_queue
 613  *
 614  */
 615
 616 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 617 __releases(&gl->gl_lockref.lock)
 618 __acquires(&gl->gl_lockref.lock)
 619 {
 620         struct gfs2_holder *gh = NULL;
 621         int ret;
 622
 623         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 624                 return;
 625
 626         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 627
 628         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 629             gl->gl_demote_state != gl->gl_state) {
 630                 if (find_first_holder(gl))
 631                         goto out_unlock;
 632                 if (nonblock)
 633                         goto out_sched;
 634                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 635                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 636                 gl->gl_target = gl->gl_demote_state;
 637         } else {
 638                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 639                         gfs2_demote_wake(gl);
 640                 ret = do_promote(gl);
 641                 if (ret == 0)
 642                         goto out_unlock;
 643                 if (ret == 2)
 644                         goto out;
 645                 gh = find_first_waiter(gl);
 646                 gl->gl_target = gh->gh_state;
 647                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 648                         do_error(gl, 0); /* Fail queued try locks */
 649         }
 650         do_xmote(gl, gh, gl->gl_target);
 651 out:
 652         return;
 653
 654 out_sched:
 655         clear_bit(GLF_LOCK, &gl->gl_flags);
 656         smp_mb__after_atomic();
 657         gl->gl_lockref.count++;
 658         __gfs2_glock_queue_work(gl, 0);
 659         return;
 660
 661 out_unlock:
 662         clear_bit(GLF_LOCK, &gl->gl_flags);
 663         smp_mb__after_atomic();
 664         return;
 665 }
 666
 667 static void delete_work_func(struct work_struct *work)
 668 {
 669         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 670         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 671         struct inode *inode;
 672         u64 no_addr = gl->gl_name.ln_number;
 673
 674         /* If someone's using this glock to create a new dinode, the block must
 675            have been freed by another node, then re-used, in which case our
 676            iopen callback is too late after the fact. Ignore it. */
 677         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 678                 goto out;
 679
 680         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 681         if (inode && !IS_ERR(inode)) {
 682                 d_prune_aliases(inode);
 683                 iput(inode);
 684         }
 685 out:
 686         gfs2_glock_put(gl);
 687 }
 688
 689 static void glock_work_func(struct work_struct *work)
 690 {
 691         unsigned long delay = 0;
 692         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 693         unsigned int drop_refs = 1;
 694
 695         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 696                 finish_xmote(gl, gl->gl_reply);
 697                 drop_refs++;
 698         }
 699         spin_lock(&gl->gl_lockref.lock);
 700         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 701             gl->gl_state != LM_ST_UNLOCKED &&
 702             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 703                 unsigned long holdtime, now = jiffies;
 704
 705                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 706                 if (time_before(now, holdtime))
 707                         delay = holdtime - now;
 708
 709                 if (!delay) {
 710                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 711                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 712                 }
 713         }
 714         run_queue(gl, 0);
 715         if (delay) {
 716                 /* Keep one glock reference for the work we requeue. */
 717                 drop_refs--;
 718                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 719                         delay = 0;
 720                 __gfs2_glock_queue_work(gl, delay);
 721         }
 722
 723         /*
 724          * Drop the remaining glock references manually here. (Mind that
 725          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 726          * here as well.)
 727          */
 728         gl->gl_lockref.count -= drop_refs;
 729         if (!gl->gl_lockref.count) {
 730                 __gfs2_glock_put(gl);
 731                 return;
 732         }
 733         spin_unlock(&gl->gl_lockref.lock);
 734 }
 735
 736 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 737                                             struct gfs2_glock *new)
 738 {
 739         struct wait_glock_queue wait;
 740         wait_queue_head_t *wq = glock_waitqueue(name);
 741         struct gfs2_glock *gl;
 742
 743         wait.name = name;
 744         init_wait(&wait.wait);
 745         wait.wait.func = glock_wake_function;
 746
 747 again:
 748         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 749         rcu_read_lock();
 750         if (new) {
 751                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 752                         &new->gl_node, ht_parms);
 753                 if (IS_ERR(gl))
 754                         goto out;
 755         } else {
 756                 gl = rhashtable_lookup_fast(&gl_hash_table,
 757                         name, ht_parms);
 758         }
 759         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 760                 rcu_read_unlock();
 761                 schedule();
 762                 goto again;
 763         }
 764 out:
 765         rcu_read_unlock();
 766         finish_wait(wq, &wait.wait);
 767         return gl;
 768 }
 769
 770 /**
 771  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 772  * @sdp: The GFS2 superblock
 773  * @number: the lock number
 774  * @glops: The glock_operations to use
 775  * @create: If 0, don't create the glock if it doesn't exist
 776  * @glp: the glock is returned here
 777  *
 778  * This does not lock a glock, just finds/creates structures for one.
 779  *
 780  * Returns: errno
 781  */
 782
 783 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 784                    const struct gfs2_glock_operations *glops, int create,
 785                    struct gfs2_glock **glp)
 786 {
 787         struct super_block *s = sdp->sd_vfs;
 788         struct lm_lockname name = { .ln_number = number,
 789                                     .ln_type = glops->go_type,
 790                                     .ln_sbd = sdp };
 791         struct gfs2_glock *gl, *tmp;
 792         struct address_space *mapping;
 793         struct kmem_cache *cachep;
 794         int ret = 0;
 795
 796         gl = find_insert_glock(&name, NULL);
 797         if (gl) {
 798                 *glp = gl;
 799                 return 0;
 800         }
 801         if (!create)
 802                 return -ENOENT;
 803
 804         if (glops->go_flags & GLOF_ASPACE)
 805                 cachep = gfs2_glock_aspace_cachep;
 806         else
 807                 cachep = gfs2_glock_cachep;
 808         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 809         if (!gl)
 810                 return -ENOMEM;
 811
 812         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 813
 814         if (glops->go_flags & GLOF_LVB) {
 815                 gl->gl_lksb.sb_lvbptr = kzalloc(GFS2_MIN_LVB_SIZE, GFP_NOFS);
 816                 if (!gl->gl_lksb.sb_lvbptr) {
 817                         kmem_cache_free(cachep, gl);
 818                         return -ENOMEM;
 819                 }
 820         }
 821
 822         atomic_inc(&sdp->sd_glock_disposal);
 823         gl->gl_node.next = NULL;
 824         gl->gl_flags = 0;
 825         gl->gl_name = name;
 826         gl->gl_lockref.count = 1;
 827         gl->gl_state = LM_ST_UNLOCKED;
 828         gl->gl_target = LM_ST_UNLOCKED;
 829         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 830         gl->gl_ops = glops;
 831         gl->gl_dstamp = 0;
 832         preempt_disable();
 833         /* We use the global stats to estimate the initial per-glock stats */
 834         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 835         preempt_enable();
 836         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 837         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 838         gl->gl_tchange = jiffies;
 839         gl->gl_object = NULL;
 840         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 841         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 842         INIT_WORK(&gl->gl_delete, delete_work_func);
 843
 844         mapping = gfs2_glock2aspace(gl);
 845         if (mapping) {
 846                 mapping->a_ops = &gfs2_meta_aops;
 847                 mapping->host = s->s_bdev->bd_inode;
 848                 mapping->flags = 0;
 849                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 850                 mapping->private_data = NULL;
 851                 mapping->writeback_index = 0;
 852         }
 853
 854         tmp = find_insert_glock(&name, gl);
 855         if (!tmp) {
 856                 *glp = gl;
 857                 goto out;
 858         }
 859         if (IS_ERR(tmp)) {
 860                 ret = PTR_ERR(tmp);
 861                 goto out_free;
 862         }
 863         *glp = tmp;
 864
 865 out_free:
 866         kfree(gl->gl_lksb.sb_lvbptr);
 867         kmem_cache_free(cachep, gl);
 868         atomic_dec(&sdp->sd_glock_disposal);
 869
 870 out:
 871         return ret;
 872 }
 873
 874 /**
 875  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 876  * @gl: the glock
 877  * @state: the state we're requesting
 878  * @flags: the modifier flags
 879  * @gh: the holder structure
 880  *
 881  */
 882
 883 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 884                       struct gfs2_holder *gh)
 885 {
 886         INIT_LIST_HEAD(&gh->gh_list);
 887         gh->gh_gl = gl;
 888         gh->gh_ip = _RET_IP_;
 889         gh->gh_owner_pid = get_pid(task_pid(current));
 890         gh->gh_state = state;
 891         gh->gh_flags = flags;
 892         gh->gh_error = 0;
 893         gh->gh_iflags = 0;
 894         gfs2_glock_hold(gl);
 895 }
 896
 897 /**
 898  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 899  * @state: the state we're requesting
 900  * @flags: the modifier flags
 901  * @gh: the holder structure
 902  *
 903  * Don't mess with the glock.
 904  *
 905  */
 906
 907 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
 908 {
 909         gh->gh_state = state;
 910         gh->gh_flags = flags;
 911         gh->gh_iflags = 0;
 912         gh->gh_ip = _RET_IP_;
 913         put_pid(gh->gh_owner_pid);
 914         gh->gh_owner_pid = get_pid(task_pid(current));
 915 }
 916
 917 /**
 918  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 919  * @gh: the holder structure
 920  *
 921  */
 922
 923 void gfs2_holder_uninit(struct gfs2_holder *gh)
 924 {
 925         put_pid(gh->gh_owner_pid);
 926         gfs2_glock_put(gh->gh_gl);
 927         gfs2_holder_mark_uninitialized(gh);
 928         gh->gh_ip = 0;
 929 }
 930
 931 /**
 932  * gfs2_glock_wait - wait on a glock acquisition
 933  * @gh: the glock holder
 934  *
 935  * Returns: 0 on success
 936  */
 937
 938 int gfs2_glock_wait(struct gfs2_holder *gh)
 939 {
 940         unsigned long time1 = jiffies;
 941
 942         might_sleep();
 943         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
 944         if (time_after(jiffies, time1 + HZ)) /* have we waited > a second? */
 945                 /* Lengthen the minimum hold time. */
 946                 gh->gh_gl->gl_hold_time = min(gh->gh_gl->gl_hold_time +
 947                                               GL_GLOCK_HOLD_INCR,
 948                                               GL_GLOCK_MAX_HOLD);
 949         return gh->gh_error;
 950 }
 951
 952 /**
 953  * handle_callback - process a demote request
 954  * @gl: the glock
 955  * @state: the state the caller wants us to change to
 956  *
 957  * There are only two requests that we are going to see in actual
 958  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 959  */
 960
 961 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 962                             unsigned long delay, bool remote)
 963 {
 964         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 965
 966         set_bit(bit, &gl->gl_flags);
 967         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 968                 gl->gl_demote_state = state;
 969                 gl->gl_demote_time = jiffies;
 970         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 971                         gl->gl_demote_state != state) {
 972                 gl->gl_demote_state = LM_ST_UNLOCKED;
 973         }
 974         if (gl->gl_ops->go_callback)
 975                 gl->gl_ops->go_callback(gl, remote);
 976         trace_gfs2_demote_rq(gl, remote);
 977 }
 978
 979 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 980 {
 981         struct va_format vaf;
 982         va_list args;
 983
 984         va_start(args, fmt);
 985
 986         if (seq) {
 987                 seq_vprintf(seq, fmt, args);
 988         } else {
 989                 vaf.fmt = fmt;
 990                 vaf.va = &args;
 991
 992                 pr_err("%pV", &vaf);
 993         }
 994
 995         va_end(args);
 996 }
 997
 998 /**
 999  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1000  * @gh: the holder structure to add
1001  *
1002  * Eventually we should move the recursive locking trap to a
1003  * debugging option or something like that. This is the fast
1004  * path and needs to have the minimum number of distractions.
1005  *
1006  */
1007
1008 static inline void add_to_queue(struct gfs2_holder *gh)
1009 __releases(&gl->gl_lockref.lock)
1010 __acquires(&gl->gl_lockref.lock)
1011 {
1012         struct gfs2_glock *gl = gh->gh_gl;
1013         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1014         struct list_head *insert_pt = NULL;
1015         struct gfs2_holder *gh2;
1016         int try_futile = 0;
1017
1018         BUG_ON(gh->gh_owner_pid == NULL);
1019         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1020                 BUG();
1021
1022         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1023                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1024                         try_futile = !may_grant(gl, gh);
1025                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1026                         goto fail;
1027         }
1028
1029         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1030                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1031                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1032                         goto trap_recursive;
1033                 if (try_futile &&
1034                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1035 fail:
1036                         gh->gh_error = GLR_TRYFAILED;
1037                         gfs2_holder_wake(gh);
1038                         return;
1039                 }
1040                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1041                         continue;
1042                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1043                         insert_pt = &gh2->gh_list;
1044         }
1045         set_bit(GLF_QUEUED, &gl->gl_flags);
1046         trace_gfs2_glock_queue(gh, 1);
1047         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1048         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1049         if (likely(insert_pt == NULL)) {
1050                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1051                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1052                         goto do_cancel;
1053                 return;
1054         }
1055         list_add_tail(&gh->gh_list, insert_pt);
1056 do_cancel:
1057         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1058         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1059                 spin_unlock(&gl->gl_lockref.lock);
1060                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1061                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1062                 spin_lock(&gl->gl_lockref.lock);
1063         }
1064         return;
1065
1066 trap_recursive:
1067         pr_err("original: %pSR\n", (void *)gh2->gh_ip);
1068         pr_err("pid: %d\n", pid_nr(gh2->gh_owner_pid));
1069         pr_err("lock type: %d req lock state : %d\n",
1070                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1071         pr_err("new: %pSR\n", (void *)gh->gh_ip);
1072         pr_err("pid: %d\n", pid_nr(gh->gh_owner_pid));
1073         pr_err("lock type: %d req lock state : %d\n",
1074                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1075         gfs2_dump_glock(NULL, gl);
1076         BUG();
1077 }
1078
1079 /**
1080  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1081  * @gh: the holder structure
1082  *
1083  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1084  *
1085  * Returns: 0, GLR_TRYFAILED, or errno on failure
1086  */
1087
1088 int gfs2_glock_nq(struct gfs2_holder *gh)
1089 {
1090         struct gfs2_glock *gl = gh->gh_gl;
1091         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1092         int error = 0;
1093
1094         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1095                 return -EIO;
1096
1097         if (test_bit(GLF_LRU, &gl->gl_flags))
1098                 gfs2_glock_remove_from_lru(gl);
1099
1100         spin_lock(&gl->gl_lockref.lock);
1101         add_to_queue(gh);
1102         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1103                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1104                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1105                 gl->gl_lockref.count++;
1106                 __gfs2_glock_queue_work(gl, 0);
1107         }
1108         run_queue(gl, 1);
1109         spin_unlock(&gl->gl_lockref.lock);
1110
1111         if (!(gh->gh_flags & GL_ASYNC))
1112                 error = gfs2_glock_wait(gh);
1113
1114         return error;
1115 }
1116
1117 /**
1118  * gfs2_glock_poll - poll to see if an async request has been completed
1119  * @gh: the holder
1120  *
1121  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1122  */
1123
1124 int gfs2_glock_poll(struct gfs2_holder *gh)
1125 {
1126         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1127 }
1128
1129 /**
1130  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1131  * @gh: the glock holder
1132  *
1133  */
1134
1135 void gfs2_glock_dq(struct gfs2_holder *gh)
1136 {
1137         struct gfs2_glock *gl = gh->gh_gl;
1138         const struct gfs2_glock_operations *glops = gl->gl_ops;
1139         unsigned delay = 0;
1140         int fast_path = 0;
1141
1142         spin_lock(&gl->gl_lockref.lock);
1143         if (gh->gh_flags & GL_NOCACHE)
1144                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1145
1146         list_del_init(&gh->gh_list);
1147         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1148         if (find_first_holder(gl) == NULL) {
1149                 if (glops->go_unlock) {
1150                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1151                         spin_unlock(&gl->gl_lockref.lock);
1152                         glops->go_unlock(gh);
1153                         spin_lock(&gl->gl_lockref.lock);
1154                         clear_bit(GLF_LOCK, &gl->gl_flags);
1155                 }
1156                 if (list_empty(&gl->gl_holders) &&
1157                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1158                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1159                         fast_path = 1;
1160         }
1161         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl) &&
1162             (glops->go_flags & GLOF_LRU))
1163                 gfs2_glock_add_to_lru(gl);
1164
1165         trace_gfs2_glock_queue(gh, 0);
1166         if (unlikely(!fast_path)) {
1167                 gl->gl_lockref.count++;
1168                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1169                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1170                     gl->gl_name.ln_type == LM_TYPE_INODE)
1171                         delay = gl->gl_hold_time;
1172                 __gfs2_glock_queue_work(gl, delay);
1173         }
1174         spin_unlock(&gl->gl_lockref.lock);
1175 }
1176
1177 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1178 {
1179         struct gfs2_glock *gl = gh->gh_gl;
1180         gfs2_glock_dq(gh);
1181         might_sleep();
1182         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1183 }
1184
1185 /**
1186  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1187  * @gh: the holder structure
1188  *
1189  */
1190
1191 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1192 {
1193         gfs2_glock_dq(gh);
1194         gfs2_holder_uninit(gh);
1195 }
1196
1197 /**
1198  * gfs2_glock_nq_num - acquire a glock based on lock number
1199  * @sdp: the filesystem
1200  * @number: the lock number
1201  * @glops: the glock operations for the type of glock
1202  * @state: the state to acquire the glock in
1203  * @flags: modifier flags for the acquisition
1204  * @gh: the struct gfs2_holder
1205  *
1206  * Returns: errno
1207  */
1208
1209 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1210                       const struct gfs2_glock_operations *glops,
1211                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1212 {
1213         struct gfs2_glock *gl;
1214         int error;
1215
1216         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1217         if (!error) {
1218                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1219                 gfs2_glock_put(gl);
1220         }
1221
1222         return error;
1223 }
1224
1225 /**
1226  * glock_compare - Compare two struct gfs2_glock structures for sorting
1227  * @arg_a: the first structure
1228  * @arg_b: the second structure
1229  *
1230  */
1231
1232 static int glock_compare(const void *arg_a, const void *arg_b)
1233 {
1234         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1235         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1236         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1237         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1238
1239         if (a->ln_number > b->ln_number)
1240                 return 1;
1241         if (a->ln_number < b->ln_number)
1242                 return -1;
1243         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1244         return 0;
1245 }
1246
1247 /**
1248  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1249  * @num_gh: the number of structures
1250  * @ghs: an array of struct gfs2_holder structures
1251  *
1252  * Returns: 0 on success (all glocks acquired),
1253  *          errno on failure (no glocks acquired)
1254  */
1255
1256 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1257                      struct gfs2_holder **p)
1258 {
1259         unsigned int x;
1260         int error = 0;
1261
1262         for (x = 0; x < num_gh; x++)
1263                 p[x] = &ghs[x];
1264
1265         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1266
1267         for (x = 0; x < num_gh; x++) {
1268                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1269
1270                 error = gfs2_glock_nq(p[x]);
1271                 if (error) {
1272                         while (x--)
1273                                 gfs2_glock_dq(p[x]);
1274                         break;
1275                 }
1276         }
1277
1278         return error;
1279 }
1280
1281 /**
1282  * gfs2_glock_nq_m - acquire multiple glocks
1283  * @num_gh: the number of structures
1284  * @ghs: an array of struct gfs2_holder structures
1285  *
1286  *
1287  * Returns: 0 on success (all glocks acquired),
1288  *          errno on failure (no glocks acquired)
1289  */
1290
1291 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1292 {
1293         struct gfs2_holder *tmp[4];
1294         struct gfs2_holder **pph = tmp;
1295         int error = 0;
1296
1297         switch(num_gh) {
1298         case 0:
1299                 return 0;
1300         case 1:
1301                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1302                 return gfs2_glock_nq(ghs);
1303         default:
1304                 if (num_gh <= 4)
1305                         break;
1306                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1307                                     GFP_NOFS);
1308                 if (!pph)
1309                         return -ENOMEM;
1310         }
1311
1312         error = nq_m_sync(num_gh, ghs, pph);
1313
1314         if (pph != tmp)
1315                 kfree(pph);
1316
1317         return error;
1318 }
1319
1320 /**
1321  * gfs2_glock_dq_m - release multiple glocks
1322  * @num_gh: the number of structures
1323  * @ghs: an array of struct gfs2_holder structures
1324  *
1325  */
1326
1327 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1328 {
1329         while (num_gh--)
1330                 gfs2_glock_dq(&ghs[num_gh]);
1331 }
1332
1333 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1334 {
1335         unsigned long delay = 0;
1336         unsigned long holdtime;
1337         unsigned long now = jiffies;
1338
1339         gfs2_glock_hold(gl);
1340         holdtime = gl->gl_tchange + gl->gl_hold_time;
1341         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1342             gl->gl_name.ln_type == LM_TYPE_INODE) {
1343                 if (time_before(now, holdtime))
1344                         delay = holdtime - now;
1345                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1346                         delay = gl->gl_hold_time;
1347         }
1348
1349         spin_lock(&gl->gl_lockref.lock);
1350         handle_callback(gl, state, delay, true);
1351         __gfs2_glock_queue_work(gl, delay);
1352         spin_unlock(&gl->gl_lockref.lock);
1353 }
1354
1355 /**
1356  * gfs2_should_freeze - Figure out if glock should be frozen
1357  * @gl: The glock in question
1358  *
1359  * Glocks are not frozen if (a) the result of the dlm operation is
1360  * an error, (b) the locking operation was an unlock operation or
1361  * (c) if there is a "noexp" flagged request anywhere in the queue
1362  *
1363  * Returns: 1 if freezing should occur, 0 otherwise
1364  */
1365
1366 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1367 {
1368         const struct gfs2_holder *gh;
1369
1370         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1371                 return 0;
1372         if (gl->gl_target == LM_ST_UNLOCKED)
1373                 return 0;
1374
1375         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1376                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1377                         continue;
1378                 if (LM_FLAG_NOEXP & gh->gh_flags)
1379                         return 0;
1380         }
1381
1382         return 1;
1383 }
1384
1385 /**
1386  * gfs2_glock_complete - Callback used by locking
1387  * @gl: Pointer to the glock
1388  * @ret: The return value from the dlm
1389  *
1390  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1391  * to use a bitfield shared with other glock state fields.
1392  */
1393
1394 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1395 {
1396         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1397
1398         spin_lock(&gl->gl_lockref.lock);
1399         gl->gl_reply = ret;
1400
1401         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1402                 if (gfs2_should_freeze(gl)) {
1403                         set_bit(GLF_FROZEN, &gl->gl_flags);
1404                         spin_unlock(&gl->gl_lockref.lock);
1405                         return;
1406                 }
1407         }
1408
1409         gl->gl_lockref.count++;
1410         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1411         __gfs2_glock_queue_work(gl, 0);
1412         spin_unlock(&gl->gl_lockref.lock);
1413 }
1414
1415 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1416 {
1417         struct gfs2_glock *gla, *glb;
1418
1419         gla = list_entry(a, struct gfs2_glock, gl_lru);
1420         glb = list_entry(b, struct gfs2_glock, gl_lru);
1421
1422         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1423                 return 1;
1424         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1425                 return -1;
1426
1427         return 0;
1428 }
1429
1430 /**
1431  * gfs2_dispose_glock_lru - Demote a list of glocks
1432  * @list: The list to dispose of
1433  *
1434  * Disposing of glocks may involve disk accesses, so that here we sort
1435  * the glocks by number (i.e. disk location of the inodes) so that if
1436  * there are any such accesses, they'll be sent in order (mostly).
1437  *
1438  * Must be called under the lru_lock, but may drop and retake this
1439  * lock. While the lru_lock is dropped, entries may vanish from the
1440  * list, but no new entries will appear on the list (since it is
1441  * private)
1442  */
1443
1444 static void gfs2_dispose_glock_lru(struct list_head *list)
1445 __releases(&lru_lock)
1446 __acquires(&lru_lock)
1447 {
1448         struct gfs2_glock *gl;
1449
1450         list_sort(NULL, list, glock_cmp);
1451
1452         while(!list_empty(list)) {
1453                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1454                 list_del_init(&gl->gl_lru);
1455                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1456 add_back_to_lru:
1457                         list_add(&gl->gl_lru, &lru_list);
1458                         atomic_inc(&lru_count);
1459                         continue;
1460                 }
1461                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1462                         spin_unlock(&gl->gl_lockref.lock);
1463                         goto add_back_to_lru;
1464                 }
1465                 clear_bit(GLF_LRU, &gl->gl_flags);
1466                 gl->gl_lockref.count++;
1467                 if (demote_ok(gl))
1468                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1469                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1470                 __gfs2_glock_queue_work(gl, 0);
1471                 spin_unlock(&gl->gl_lockref.lock);
1472                 cond_resched_lock(&lru_lock);
1473         }
1474 }
1475
1476 /**
1477  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1478  * @nr: The number of entries to scan
1479  *
1480  * This function selects the entries on the LRU which are able to
1481  * be demoted, and then kicks off the process by calling
1482  * gfs2_dispose_glock_lru() above.
1483  */
1484
1485 static long gfs2_scan_glock_lru(int nr)
1486 {
1487         struct gfs2_glock *gl;
1488         LIST_HEAD(skipped);
1489         LIST_HEAD(dispose);
1490         long freed = 0;
1491
1492         spin_lock(&lru_lock);
1493         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1494                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1495
1496                 /* Test for being demotable */
1497                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1498                         list_move(&gl->gl_lru, &dispose);
1499                         atomic_dec(&lru_count);
1500                         freed++;
1501                         continue;
1502                 }
1503
1504                 list_move(&gl->gl_lru, &skipped);
1505         }
1506         list_splice(&skipped, &lru_list);
1507         if (!list_empty(&dispose))
1508                 gfs2_dispose_glock_lru(&dispose);
1509         spin_unlock(&lru_lock);
1510
1511         return freed;
1512 }
1513
1514 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1515                                             struct shrink_control *sc)
1516 {
1517         if (!(sc->gfp_mask & __GFP_FS))
1518                 return SHRINK_STOP;
1519         return gfs2_scan_glock_lru(sc->nr_to_scan);
1520 }
1521
1522 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1523                                              struct shrink_control *sc)
1524 {
1525         return vfs_pressure_ratio(atomic_read(&lru_count));
1526 }
1527
1528 static struct shrinker glock_shrinker = {
1529         .seeks = DEFAULT_SEEKS,
1530         .count_objects = gfs2_glock_shrink_count,
1531         .scan_objects = gfs2_glock_shrink_scan,
1532 };
1533
1534 /**
1535  * examine_bucket - Call a function for glock in a hash bucket
1536  * @examiner: the function
1537  * @sdp: the filesystem
1538  * @bucket: the bucket
1539  *
1540  * Note that the function can be called multiple times on the same
1541  * object.  So the user must ensure that the function can cope with
1542  * that.
1543  */
1544
1545 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1546 {
1547         struct gfs2_glock *gl;
1548         struct rhashtable_iter iter;
1549
1550         rhashtable_walk_enter(&gl_hash_table, &iter);
1551
1552         do {
1553                 rhashtable_walk_start(&iter);
1554
1555                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1556                         if (gl->gl_name.ln_sbd == sdp &&
1557                             lockref_get_not_dead(&gl->gl_lockref))
1558                                 examiner(gl);
1559
1560                 rhashtable_walk_stop(&iter);
1561         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1562
1563         rhashtable_walk_exit(&iter);
1564 }
1565
1566 /**
1567  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1568  * @gl: The glock to thaw
1569  *
1570  */
1571
1572 static void thaw_glock(struct gfs2_glock *gl)
1573 {
1574         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1575                 gfs2_glock_put(gl);
1576                 return;
1577         }
1578         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1579         gfs2_glock_queue_work(gl, 0);
1580 }
1581
1582 /**
1583  * clear_glock - look at a glock and see if we can free it from glock cache
1584  * @gl: the glock to look at
1585  *
1586  */
1587
1588 static void clear_glock(struct gfs2_glock *gl)
1589 {
1590         gfs2_glock_remove_from_lru(gl);
1591
1592         spin_lock(&gl->gl_lockref.lock);
1593         if (gl->gl_state != LM_ST_UNLOCKED)
1594                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1595         __gfs2_glock_queue_work(gl, 0);
1596         spin_unlock(&gl->gl_lockref.lock);
1597 }
1598
1599 /**
1600  * gfs2_glock_thaw - Thaw any frozen glocks
1601  * @sdp: The super block
1602  *
1603  */
1604
1605 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1606 {
1607         glock_hash_walk(thaw_glock, sdp);
1608 }
1609
1610 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1611 {
1612         spin_lock(&gl->gl_lockref.lock);
1613         gfs2_dump_glock(seq, gl);
1614         spin_unlock(&gl->gl_lockref.lock);
1615 }
1616
1617 static void dump_glock_func(struct gfs2_glock *gl)
1618 {
1619         dump_glock(NULL, gl);
1620 }
1621
1622 /**
1623  * gfs2_gl_hash_clear - Empty out the glock hash table
1624  * @sdp: the filesystem
1625  * @wait: wait until it's all gone
1626  *
1627  * Called when unmounting the filesystem.
1628  */
1629
1630 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1631 {
1632         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1633         flush_workqueue(glock_workqueue);
1634         glock_hash_walk(clear_glock, sdp);
1635         flush_workqueue(glock_workqueue);
1636         wait_event_timeout(sdp->sd_glock_wait,
1637                            atomic_read(&sdp->sd_glock_disposal) == 0,
1638                            HZ * 600);
1639         glock_hash_walk(dump_glock_func, sdp);
1640 }
1641
1642 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1643 {
1644         struct gfs2_glock *gl = ip->i_gl;
1645         int ret;
1646
1647         ret = gfs2_truncatei_resume(ip);
1648         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1649
1650         spin_lock(&gl->gl_lockref.lock);
1651         clear_bit(GLF_LOCK, &gl->gl_flags);
1652         run_queue(gl, 1);
1653         spin_unlock(&gl->gl_lockref.lock);
1654 }
1655
1656 static const char *state2str(unsigned state)
1657 {
1658         switch(state) {
1659         case LM_ST_UNLOCKED:
1660                 return "UN";
1661         case LM_ST_SHARED:
1662                 return "SH";
1663         case LM_ST_DEFERRED:
1664                 return "DF";
1665         case LM_ST_EXCLUSIVE:
1666                 return "EX";
1667         }
1668         return "??";
1669 }
1670
1671 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1672 {
1673         char *p = buf;
1674         if (flags & LM_FLAG_TRY)
1675                 *p++ = 't';
1676         if (flags & LM_FLAG_TRY_1CB)
1677                 *p++ = 'T';
1678         if (flags & LM_FLAG_NOEXP)
1679                 *p++ = 'e';
1680         if (flags & LM_FLAG_ANY)
1681                 *p++ = 'A';
1682         if (flags & LM_FLAG_PRIORITY)
1683                 *p++ = 'p';
1684         if (flags & GL_ASYNC)
1685                 *p++ = 'a';
1686         if (flags & GL_EXACT)
1687                 *p++ = 'E';
1688         if (flags & GL_NOCACHE)
1689                 *p++ = 'c';
1690         if (test_bit(HIF_HOLDER, &iflags))
1691                 *p++ = 'H';
1692         if (test_bit(HIF_WAIT, &iflags))
1693                 *p++ = 'W';
1694         if (test_bit(HIF_FIRST, &iflags))
1695                 *p++ = 'F';
1696         *p = 0;
1697         return buf;
1698 }
1699
1700 /**
1701  * dump_holder - print information about a glock holder
1702  * @seq: the seq_file struct
1703  * @gh: the glock holder
1704  *
1705  */
1706
1707 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1708 {
1709         struct task_struct *gh_owner = NULL;
1710         char flags_buf[32];
1711
1712         rcu_read_lock();
1713         if (gh->gh_owner_pid)
1714                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1715         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1716                        state2str(gh->gh_state),
1717                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1718                        gh->gh_error,
1719                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1720                        gh_owner ? gh_owner->comm : "(ended)",
1721                        (void *)gh->gh_ip);
1722         rcu_read_unlock();
1723 }
1724
1725 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1726 {
1727         const unsigned long *gflags = &gl->gl_flags;
1728         char *p = buf;
1729
1730         if (test_bit(GLF_LOCK, gflags))
1731                 *p++ = 'l';
1732         if (test_bit(GLF_DEMOTE, gflags))
1733                 *p++ = 'D';
1734         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1735                 *p++ = 'd';
1736         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1737                 *p++ = 'p';
1738         if (test_bit(GLF_DIRTY, gflags))
1739                 *p++ = 'y';
1740         if (test_bit(GLF_LFLUSH, gflags))
1741                 *p++ = 'f';
1742         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1743                 *p++ = 'i';
1744         if (test_bit(GLF_REPLY_PENDING, gflags))
1745                 *p++ = 'r';
1746         if (test_bit(GLF_INITIAL, gflags))
1747                 *p++ = 'I';
1748         if (test_bit(GLF_FROZEN, gflags))
1749                 *p++ = 'F';
1750         if (test_bit(GLF_QUEUED, gflags))
1751                 *p++ = 'q';
1752         if (test_bit(GLF_LRU, gflags))
1753                 *p++ = 'L';
1754         if (gl->gl_object)
1755                 *p++ = 'o';
1756         if (test_bit(GLF_BLOCKING, gflags))
1757                 *p++ = 'b';
1758         *p = 0;
1759         return buf;
1760 }
1761
1762 /**
1763  * gfs2_dump_glock - print information about a glock
1764  * @seq: The seq_file struct
1765  * @gl: the glock
1766  *
1767  * The file format is as follows:
1768  * One line per object, capital letters are used to indicate objects
1769  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1770  * other objects are indented by a single space and follow the glock to
1771  * which they are related. Fields are indicated by lower case letters
1772  * followed by a colon and the field value, except for strings which are in
1773  * [] so that its possible to see if they are composed of spaces for
1774  * example. The field's are n = number (id of the object), f = flags,
1775  * t = type, s = state, r = refcount, e = error, p = pid.
1776  *
1777  */
1778
1779 void gfs2_dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1780 {
1781         const struct gfs2_glock_operations *glops = gl->gl_ops;
1782         unsigned long long dtime;
1783         const struct gfs2_holder *gh;
1784         char gflags_buf[32];
1785
1786         dtime = jiffies - gl->gl_demote_time;
1787         dtime *= 1000000/HZ; /* demote time in uSec */
1788         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1789                 dtime = 0;
1790         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",
1791                   state2str(gl->gl_state),
1792                   gl->gl_name.ln_type,
1793                   (unsigned long long)gl->gl_name.ln_number,
1794                   gflags2str(gflags_buf, gl),
1795                   state2str(gl->gl_target),
1796                   state2str(gl->gl_demote_state), dtime,
1797                   atomic_read(&gl->gl_ail_count),
1798                   atomic_read(&gl->gl_revokes),
1799                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1800
1801         list_for_each_entry(gh, &gl->gl_holders, gh_list)
1802                 dump_holder(seq, gh);
1803
1804         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1805                 glops->go_dump(seq, gl);
1806 }
1807
1808 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
1809 {
1810         struct gfs2_glock *gl = iter_ptr;
1811
1812         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
1813                    gl->gl_name.ln_type,
1814                    (unsigned long long)gl->gl_name.ln_number,
1815                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
1816                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
1817                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
1818                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
1819                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
1820                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
1821                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
1822                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
1823         return 0;
1824 }
1825
1826 static const char *gfs2_gltype[] = {
1827         "type",
1828         "reserved",
1829         "nondisk",
1830         "inode",
1831         "rgrp",
1832         "meta",
1833         "iopen",
1834         "flock",
1835         "plock",
1836         "quota",
1837         "journal",
1838 };
1839
1840 static const char *gfs2_stype[] = {
1841         [GFS2_LKS_SRTT]         = "srtt",
1842         [GFS2_LKS_SRTTVAR]      = "srttvar",
1843         [GFS2_LKS_SRTTB]        = "srttb",
1844         [GFS2_LKS_SRTTVARB]     = "srttvarb",
1845         [GFS2_LKS_SIRT]         = "sirt",
1846         [GFS2_LKS_SIRTVAR]      = "sirtvar",
1847         [GFS2_LKS_DCOUNT]       = "dlm",
1848         [GFS2_LKS_QCOUNT]       = "queue",
1849 };
1850
1851 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
1852
1853 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
1854 {
1855         struct gfs2_sbd *sdp = seq->private;
1856         loff_t pos = *(loff_t *)iter_ptr;
1857         unsigned index = pos >> 3;
1858         unsigned subindex = pos & 0x07;
1859         int i;
1860
1861         if (index == 0 && subindex != 0)
1862                 return 0;
1863
1864         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
1865                    (index == 0) ? "cpu": gfs2_stype[subindex]);
1866
1867         for_each_possible_cpu(i) {
1868                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
1869
1870                 if (index == 0)
1871                         seq_printf(seq, " %15u", i);
1872                 else
1873                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
1874                                    lkstats[index - 1].stats[subindex]);
1875         }
1876         seq_putc(seq, '\n');
1877         return 0;
1878 }
1879
1880 int __init gfs2_glock_init(void)
1881 {
1882         int i, ret;
1883
1884         ret = rhashtable_init(&gl_hash_table, &ht_parms);
1885         if (ret < 0)
1886                 return ret;
1887
1888         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1889                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1890         if (!glock_workqueue) {
1891                 rhashtable_destroy(&gl_hash_table);
1892                 return -ENOMEM;
1893         }
1894         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1895                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1896                                                 0);
1897         if (!gfs2_delete_workqueue) {
1898                 destroy_workqueue(glock_workqueue);
1899                 rhashtable_destroy(&gl_hash_table);
1900                 return -ENOMEM;
1901         }
1902
1903         ret = register_shrinker(&glock_shrinker);
1904         if (ret) {
1905                 destroy_workqueue(gfs2_delete_workqueue);
1906                 destroy_workqueue(glock_workqueue);
1907                 rhashtable_destroy(&gl_hash_table);
1908                 return ret;
1909         }
1910
1911         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
1912                 init_waitqueue_head(glock_wait_table + i);
1913
1914         return 0;
1915 }
1916
1917 void gfs2_glock_exit(void)
1918 {
1919         unregister_shrinker(&glock_shrinker);
1920         rhashtable_destroy(&gl_hash_table);
1921         destroy_workqueue(glock_workqueue);
1922         destroy_workqueue(gfs2_delete_workqueue);
1923 }
1924
1925 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
1926 {
1927         struct gfs2_glock *gl = gi->gl;
1928
1929         if (gl) {
1930                 if (n == 0)
1931                         return;
1932                 if (!lockref_put_not_zero(&gl->gl_lockref))
1933                         gfs2_glock_queue_put(gl);
1934         }
1935         for (;;) {
1936                 gl = rhashtable_walk_next(&gi->hti);
1937                 if (IS_ERR_OR_NULL(gl)) {
1938                         if (gl == ERR_PTR(-EAGAIN)) {
1939                                 n = 1;
1940                                 continue;
1941                         }
1942                         gl = NULL;
1943                         break;
1944                 }
1945                 if (gl->gl_name.ln_sbd != gi->sdp)
1946                         continue;
1947                 if (n <= 1) {
1948                         if (!lockref_get_not_dead(&gl->gl_lockref))
1949                                 continue;
1950                         break;
1951                 } else {
1952                         if (__lockref_is_dead(&gl->gl_lockref))
1953                                 continue;
1954                         n--;
1955                 }
1956         }
1957         gi->gl = gl;
1958 }
1959
1960 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1961         __acquires(RCU)
1962 {
1963         struct gfs2_glock_iter *gi = seq->private;
1964         loff_t n;
1965
1966         /*
1967          * We can either stay where we are, skip to the next hash table
1968          * entry, or start from the beginning.
1969          */
1970         if (*pos < gi->last_pos) {
1971                 rhashtable_walk_exit(&gi->hti);
1972                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
1973                 n = *pos + 1;
1974         } else {
1975                 n = *pos - gi->last_pos;
1976         }
1977
1978         rhashtable_walk_start(&gi->hti);
1979
1980         gfs2_glock_iter_next(gi, n);
1981         gi->last_pos = *pos;
1982         return gi->gl;
1983 }
1984
1985 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1986                                  loff_t *pos)
1987 {
1988         struct gfs2_glock_iter *gi = seq->private;
1989
1990         (*pos)++;
1991         gi->last_pos = *pos;
1992         gfs2_glock_iter_next(gi, 1);
1993         return gi->gl;
1994 }
1995
1996 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1997         __releases(RCU)
1998 {
1999         struct gfs2_glock_iter *gi = seq->private;
2000
2001         rhashtable_walk_stop(&gi->hti);
2002 }
2003
2004 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2005 {
2006         dump_glock(seq, iter_ptr);
2007         return 0;
2008 }
2009
2010 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2011 {
2012         preempt_disable();
2013         if (*pos >= GFS2_NR_SBSTATS)
2014                 return NULL;
2015         return pos;
2016 }
2017
2018 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2019                                    loff_t *pos)
2020 {
2021         (*pos)++;
2022         if (*pos >= GFS2_NR_SBSTATS)
2023                 return NULL;
2024         return pos;
2025 }
2026
2027 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2028 {
2029         preempt_enable();
2030 }
2031
2032 static const struct seq_operations gfs2_glock_seq_ops = {
2033         .start = gfs2_glock_seq_start,
2034         .next  = gfs2_glock_seq_next,
2035         .stop  = gfs2_glock_seq_stop,
2036         .show  = gfs2_glock_seq_show,
2037 };
2038
2039 static const struct seq_operations gfs2_glstats_seq_ops = {
2040         .start = gfs2_glock_seq_start,
2041         .next  = gfs2_glock_seq_next,
2042         .stop  = gfs2_glock_seq_stop,
2043         .show  = gfs2_glstats_seq_show,
2044 };
2045
2046 static const struct seq_operations gfs2_sbstats_seq_ops = {
2047         .start = gfs2_sbstats_seq_start,
2048         .next  = gfs2_sbstats_seq_next,
2049         .stop  = gfs2_sbstats_seq_stop,
2050         .show  = gfs2_sbstats_seq_show,
2051 };
2052
2053 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2054
2055 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2056                               const struct seq_operations *ops)
2057 {
2058         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2059         if (ret == 0) {
2060                 struct seq_file *seq = file->private_data;
2061                 struct gfs2_glock_iter *gi = seq->private;
2062
2063                 gi->sdp = inode->i_private;
2064                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2065                 if (seq->buf)
2066                         seq->size = GFS2_SEQ_GOODSIZE;
2067                 /*
2068                  * Initially, we are "before" the first hash table entry; the
2069                  * first call to rhashtable_walk_next gets us the first entry.
2070                  */
2071                 gi->last_pos = -1;
2072                 gi->gl = NULL;
2073                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2074         }
2075         return ret;
2076 }
2077
2078 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2079 {
2080         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2081 }
2082
2083 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2084 {
2085         struct seq_file *seq = file->private_data;
2086         struct gfs2_glock_iter *gi = seq->private;
2087
2088         if (gi->gl)
2089                 gfs2_glock_put(gi->gl);
2090         rhashtable_walk_exit(&gi->hti);
2091         return seq_release_private(inode, file);
2092 }
2093
2094 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2095 {
2096         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2097 }
2098
2099 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2100 {
2101         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2102         if (ret == 0) {
2103                 struct seq_file *seq = file->private_data;
2104                 seq->private = inode->i_private;  /* sdp */
2105         }
2106         return ret;
2107 }
2108
2109 static const struct file_operations gfs2_glocks_fops = {
2110         .owner   = THIS_MODULE,
2111         .open    = gfs2_glocks_open,
2112         .read    = seq_read,
2113         .llseek  = seq_lseek,
2114         .release = gfs2_glocks_release,
2115 };
2116
2117 static const struct file_operations gfs2_glstats_fops = {
2118         .owner   = THIS_MODULE,
2119         .open    = gfs2_glstats_open,
2120         .read    = seq_read,
2121         .llseek  = seq_lseek,
2122         .release = gfs2_glocks_release,
2123 };
2124
2125 static const struct file_operations gfs2_sbstats_fops = {
2126         .owner   = THIS_MODULE,
2127         .open    = gfs2_sbstats_open,
2128         .read    = seq_read,
2129         .llseek  = seq_lseek,
2130         .release = seq_release,
2131 };
2132
2133 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2134 {
2135         struct dentry *dent;
2136
2137         dent = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2138         if (IS_ERR_OR_NULL(dent))
2139                 goto fail;
2140         sdp->debugfs_dir = dent;
2141
2142         dent = debugfs_create_file("glocks",
2143                                    S_IFREG | S_IRUGO,
2144                                    sdp->debugfs_dir, sdp,
2145                                    &gfs2_glocks_fops);
2146         if (IS_ERR_OR_NULL(dent))
2147                 goto fail;
2148         sdp->debugfs_dentry_glocks = dent;
2149
2150         dent = debugfs_create_file("glstats",
2151                                    S_IFREG | S_IRUGO,
2152                                    sdp->debugfs_dir, sdp,
2153                                    &gfs2_glstats_fops);
2154         if (IS_ERR_OR_NULL(dent))
2155                 goto fail;
2156         sdp->debugfs_dentry_glstats = dent;
2157
2158         dent = debugfs_create_file("sbstats",
2159                                    S_IFREG | S_IRUGO,
2160                                    sdp->debugfs_dir, sdp,
2161                                    &gfs2_sbstats_fops);
2162         if (IS_ERR_OR_NULL(dent))
2163                 goto fail;
2164         sdp->debugfs_dentry_sbstats = dent;
2165
2166         return 0;
2167 fail:
2168         gfs2_delete_debugfs_file(sdp);
2169         return dent ? PTR_ERR(dent) : -ENOMEM;
2170 }
2171
2172 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2173 {
2174         if (sdp->debugfs_dir) {
2175                 if (sdp->debugfs_dentry_glocks) {
2176                         debugfs_remove(sdp->debugfs_dentry_glocks);
2177                         sdp->debugfs_dentry_glocks = NULL;
2178                 }
2179                 if (sdp->debugfs_dentry_glstats) {
2180                         debugfs_remove(sdp->debugfs_dentry_glstats);
2181                         sdp->debugfs_dentry_glstats = NULL;
2182                 }
2183                 if (sdp->debugfs_dentry_sbstats) {
2184                         debugfs_remove(sdp->debugfs_dentry_sbstats);
2185                         sdp->debugfs_dentry_sbstats = NULL;
2186                 }
2187                 debugfs_remove(sdp->debugfs_dir);
2188                 sdp->debugfs_dir = NULL;
2189         }
2190 }
2191
2192 int gfs2_register_debugfs(void)
2193 {
2194         gfs2_root = debugfs_create_dir("gfs2", NULL);
2195         if (IS_ERR(gfs2_root))
2196                 return PTR_ERR(gfs2_root);
2197         return gfs2_root ? 0 : -ENOMEM;
2198 }
2199
2200 void gfs2_unregister_debugfs(void)
2201 {
2202         debugfs_remove(gfs2_root);
2203         gfs2_root = NULL;
2204 }