fs/gfs2/glock.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   5  */
   6
   7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/sched.h>
  10 #include <linux/slab.h>
  11 #include <linux/spinlock.h>
  12 #include <linux/buffer_head.h>
  13 #include <linux/delay.h>
  14 #include <linux/sort.h>
  15 #include <linux/hash.h>
  16 #include <linux/jhash.h>
  17 #include <linux/kallsyms.h>
  18 #include <linux/gfs2_ondisk.h>
  19 #include <linux/list.h>
  20 #include <linux/wait.h>
  21 #include <linux/module.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/debugfs.h>
  25 #include <linux/kthread.h>
  26 #include <linux/freezer.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/jiffies.h>
  29 #include <linux/rcupdate.h>
  30 #include <linux/rculist_bl.h>
  31 #include <linux/bit_spinlock.h>
  32 #include <linux/percpu.h>
  33 #include <linux/list_sort.h>
  34 #include <linux/lockref.h>
  35 #include <linux/rhashtable.h>
  36
  37 #include "gfs2.h"
  38 #include "incore.h"
  39 #include "glock.h"
  40 #include "glops.h"
  41 #include "inode.h"
  42 #include "lops.h"
  43 #include "meta_io.h"
  44 #include "quota.h"
  45 #include "super.h"
  46 #include "util.h"
  47 #include "bmap.h"
  48 #define CREATE_TRACE_POINTS
  49 #include "trace_gfs2.h"
  50
  51 struct gfs2_glock_iter {
  52         struct gfs2_sbd *sdp;           /* incore superblock           */
  53         struct rhashtable_iter hti;     /* rhashtable iterator         */
  54         struct gfs2_glock *gl;          /* current glock struct        */
  55         loff_t last_pos;                /* last position               */
  56 };
  57
  58 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  59
  60 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  61
  62 static struct dentry *gfs2_root;
  63 static struct workqueue_struct *glock_workqueue;
  64 struct workqueue_struct *gfs2_delete_workqueue;
  65 static LIST_HEAD(lru_list);
  66 static atomic_t lru_count = ATOMIC_INIT(0);
  67 static DEFINE_SPINLOCK(lru_lock);
  68
  69 #define GFS2_GL_HASH_SHIFT      15
  70 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
  71
  72 static const struct rhashtable_params ht_parms = {
  73         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
  74         .key_len = offsetofend(struct lm_lockname, ln_type),
  75         .key_offset = offsetof(struct gfs2_glock, gl_name),
  76         .head_offset = offsetof(struct gfs2_glock, gl_node),
  77 };
  78
  79 static struct rhashtable gl_hash_table;
  80
  81 #define GLOCK_WAIT_TABLE_BITS 12
  82 #define GLOCK_WAIT_TABLE_SIZE (1 << GLOCK_WAIT_TABLE_BITS)
  83 static wait_queue_head_t glock_wait_table[GLOCK_WAIT_TABLE_SIZE] __cacheline_aligned;
  84
  85 struct wait_glock_queue {
  86         struct lm_lockname *name;
  87         wait_queue_entry_t wait;
  88 };
  89
  90 static int glock_wake_function(wait_queue_entry_t *wait, unsigned int mode,
  91                                int sync, void *key)
  92 {
  93         struct wait_glock_queue *wait_glock =
  94                 container_of(wait, struct wait_glock_queue, wait);
  95         struct lm_lockname *wait_name = wait_glock->name;
  96         struct lm_lockname *wake_name = key;
  97
  98         if (wake_name->ln_sbd != wait_name->ln_sbd ||
  99             wake_name->ln_number != wait_name->ln_number ||
 100             wake_name->ln_type != wait_name->ln_type)
 101                 return 0;
 102         return autoremove_wake_function(wait, mode, sync, key);
 103 }
 104
 105 static wait_queue_head_t *glock_waitqueue(struct lm_lockname *name)
 106 {
 107         u32 hash = jhash2((u32 *)name, ht_parms.key_len / 4, 0);
 108
 109         return glock_wait_table + hash_32(hash, GLOCK_WAIT_TABLE_BITS);
 110 }
 111
 112 /**
 113  * wake_up_glock  -  Wake up waiters on a glock
 114  * @gl: the glock
 115  */
 116 static void wake_up_glock(struct gfs2_glock *gl)
 117 {
 118         wait_queue_head_t *wq = glock_waitqueue(&gl->gl_name);
 119
 120         if (waitqueue_active(wq))
 121                 __wake_up(wq, TASK_NORMAL, 1, &gl->gl_name);
 122 }
 123
 124 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 125 {
 126         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 127
 128         kfree(gl->gl_lksb.sb_lvbptr);
 129         if (gl->gl_ops->go_flags & GLOF_ASPACE)
 130                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 131         else
 132                 kmem_cache_free(gfs2_glock_cachep, gl);
 133 }
 134
 135 /**
 136  * glock_blocked_by_withdraw - determine if we can still use a glock
 137  * @gl: the glock
 138  *
 139  * We need to allow some glocks to be enqueued, dequeued, promoted, and demoted
 140  * when we're withdrawn. For example, to maintain metadata integrity, we should
 141  * disallow the use of inode and rgrp glocks when withdrawn. Other glocks, like
 142  * iopen or the transaction glocks may be safely used because none of their
 143  * metadata goes through the journal. So in general, we should disallow all
 144  * glocks that are journaled, and allow all the others. One exception is:
 145  * we need to allow our active journal to be promoted and demoted so others
 146  * may recover it and we can reacquire it when they're done.
 147  */
 148 static bool glock_blocked_by_withdraw(struct gfs2_glock *gl)
 149 {
 150         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 151
 152         if (likely(!gfs2_withdrawn(sdp)))
 153                 return false;
 154         if (gl->gl_ops->go_flags & GLOF_NONDISK)
 155                 return false;
 156         if (!sdp->sd_jdesc ||
 157             gl->gl_name.ln_number == sdp->sd_jdesc->jd_no_addr)
 158                 return false;
 159         return true;
 160 }
 161
 162 void gfs2_glock_free(struct gfs2_glock *gl)
 163 {
 164         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 165
 166         gfs2_glock_assert_withdraw(gl, atomic_read(&gl->gl_revokes) == 0);
 167         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 168         smp_mb();
 169         wake_up_glock(gl);
 170         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 171         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 172                 wake_up(&sdp->sd_glock_wait);
 173 }
 174
 175 /**
 176  * gfs2_glock_hold() - increment reference count on glock
 177  * @gl: The glock to hold
 178  *
 179  */
 180
 181 void gfs2_glock_hold(struct gfs2_glock *gl)
 182 {
 183         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 184         lockref_get(&gl->gl_lockref);
 185 }
 186
 187 /**
 188  * demote_ok - Check to see if it's ok to unlock a glock
 189  * @gl: the glock
 190  *
 191  * Returns: 1 if it's ok
 192  */
 193
 194 static int demote_ok(const struct gfs2_glock *gl)
 195 {
 196         const struct gfs2_glock_operations *glops = gl->gl_ops;
 197
 198         if (gl->gl_state == LM_ST_UNLOCKED)
 199                 return 0;
 200         if (!list_empty(&gl->gl_holders))
 201                 return 0;
 202         if (glops->go_demote_ok)
 203                 return glops->go_demote_ok(gl);
 204         return 1;
 205 }
 206
 207
 208 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 209 {
 210         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 211                 return;
 212
 213         spin_lock(&lru_lock);
 214
 215         list_del(&gl->gl_lru);
 216         list_add_tail(&gl->gl_lru, &lru_list);
 217
 218         if (!test_bit(GLF_LRU, &gl->gl_flags)) {
 219                 set_bit(GLF_LRU, &gl->gl_flags);
 220                 atomic_inc(&lru_count);
 221         }
 222
 223         spin_unlock(&lru_lock);
 224 }
 225
 226 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 227 {
 228         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 229                 return;
 230
 231         spin_lock(&lru_lock);
 232         if (test_bit(GLF_LRU, &gl->gl_flags)) {
 233                 list_del_init(&gl->gl_lru);
 234                 atomic_dec(&lru_count);
 235                 clear_bit(GLF_LRU, &gl->gl_flags);
 236         }
 237         spin_unlock(&lru_lock);
 238 }
 239
 240 /*
 241  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 242  * work queue.
 243  */
 244 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 245         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 246                 /*
 247                  * We are holding the lockref spinlock, and the work was still
 248                  * queued above.  The queued work (glock_work_func) takes that
 249                  * spinlock before dropping its glock reference(s), so it
 250                  * cannot have dropped them in the meantime.
 251                  */
 252                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 253                 gl->gl_lockref.count--;
 254         }
 255 }
 256
 257 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 258         spin_lock(&gl->gl_lockref.lock);
 259         __gfs2_glock_queue_work(gl, delay);
 260         spin_unlock(&gl->gl_lockref.lock);
 261 }
 262
 263 static void __gfs2_glock_put(struct gfs2_glock *gl)
 264 {
 265         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 266         struct address_space *mapping = gfs2_glock2aspace(gl);
 267
 268         lockref_mark_dead(&gl->gl_lockref);
 269
 270         gfs2_glock_remove_from_lru(gl);
 271         spin_unlock(&gl->gl_lockref.lock);
 272         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 273         GLOCK_BUG_ON(gl, mapping && mapping->nrpages && !gfs2_withdrawn(sdp));
 274         trace_gfs2_glock_put(gl);
 275         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 276 }
 277
 278 /*
 279  * Cause the glock to be put in work queue context.
 280  */
 281 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 282 {
 283         gfs2_glock_queue_work(gl, 0);
 284 }
 285
 286 /**
 287  * gfs2_glock_put() - Decrement reference count on glock
 288  * @gl: The glock to put
 289  *
 290  */
 291
 292 void gfs2_glock_put(struct gfs2_glock *gl)
 293 {
 294         if (lockref_put_or_lock(&gl->gl_lockref))
 295                 return;
 296
 297         __gfs2_glock_put(gl);
 298 }
 299
 300 /**
 301  * may_grant - check if its ok to grant a new lock
 302  * @gl: The glock
 303  * @gh: The lock request which we wish to grant
 304  *
 305  * Returns: true if its ok to grant the lock
 306  */
 307
 308 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 309 {
 310         const struct gfs2_holder *gh_head = list_first_entry(&gl->gl_holders, const struct gfs2_holder, gh_list);
 311         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 312              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 313                 return 0;
 314         if (gl->gl_state == gh->gh_state)
 315                 return 1;
 316         if (gh->gh_flags & GL_EXACT)
 317                 return 0;
 318         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 319                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 320                         return 1;
 321                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 322                         return 1;
 323         }
 324         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 325                 return 1;
 326         return 0;
 327 }
 328
 329 static void gfs2_holder_wake(struct gfs2_holder *gh)
 330 {
 331         clear_bit(HIF_WAIT, &gh->gh_iflags);
 332         smp_mb__after_atomic();
 333         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 334         if (gh->gh_flags & GL_ASYNC) {
 335                 struct gfs2_sbd *sdp = gh->gh_gl->gl_name.ln_sbd;
 336
 337                 wake_up(&sdp->sd_async_glock_wait);
 338         }
 339 }
 340
 341 /**
 342  * do_error - Something unexpected has happened during a lock request
 343  *
 344  */
 345
 346 static void do_error(struct gfs2_glock *gl, const int ret)
 347 {
 348         struct gfs2_holder *gh, *tmp;
 349
 350         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 351                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 352                         continue;
 353                 if (ret & LM_OUT_ERROR)
 354                         gh->gh_error = -EIO;
 355                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 356                         gh->gh_error = GLR_TRYFAILED;
 357                 else
 358                         continue;
 359                 list_del_init(&gh->gh_list);
 360                 trace_gfs2_glock_queue(gh, 0);
 361                 gfs2_holder_wake(gh);
 362         }
 363 }
 364
 365 /**
 366  * do_promote - promote as many requests as possible on the current queue
 367  * @gl: The glock
 368  *
 369  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 370  *          if a type specific operation is underway.
 371  */
 372
 373 static int do_promote(struct gfs2_glock *gl)
 374 __releases(&gl->gl_lockref.lock)
 375 __acquires(&gl->gl_lockref.lock)
 376 {
 377         const struct gfs2_glock_operations *glops = gl->gl_ops;
 378         struct gfs2_holder *gh, *tmp;
 379         int ret;
 380
 381 restart:
 382         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 383                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 384                         continue;
 385                 if (may_grant(gl, gh)) {
 386                         if (gh->gh_list.prev == &gl->gl_holders &&
 387                             glops->go_lock) {
 388                                 spin_unlock(&gl->gl_lockref.lock);
 389                                 /* FIXME: eliminate this eventually */
 390                                 ret = glops->go_lock(gh);
 391                                 spin_lock(&gl->gl_lockref.lock);
 392                                 if (ret) {
 393                                         if (ret == 1)
 394                                                 return 2;
 395                                         gh->gh_error = ret;
 396                                         list_del_init(&gh->gh_list);
 397                                         trace_gfs2_glock_queue(gh, 0);
 398                                         gfs2_holder_wake(gh);
 399                                         goto restart;
 400                                 }
 401                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 402                                 trace_gfs2_promote(gh, 1);
 403                                 gfs2_holder_wake(gh);
 404                                 goto restart;
 405                         }
 406                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 407                         trace_gfs2_promote(gh, 0);
 408                         gfs2_holder_wake(gh);
 409                         continue;
 410                 }
 411                 if (gh->gh_list.prev == &gl->gl_holders)
 412                         return 1;
 413                 do_error(gl, 0);
 414                 break;
 415         }
 416         return 0;
 417 }
 418
 419 /**
 420  * find_first_waiter - find the first gh that's waiting for the glock
 421  * @gl: the glock
 422  */
 423
 424 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 425 {
 426         struct gfs2_holder *gh;
 427
 428         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 429                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 430                         return gh;
 431         }
 432         return NULL;
 433 }
 434
 435 /**
 436  * state_change - record that the glock is now in a different state
 437  * @gl: the glock
 438  * @new_state the new state
 439  *
 440  */
 441
 442 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 443 {
 444         int held1, held2;
 445
 446         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 447         held2 = (new_state != LM_ST_UNLOCKED);
 448
 449         if (held1 != held2) {
 450                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 451                 if (held2)
 452                         gl->gl_lockref.count++;
 453                 else
 454                         gl->gl_lockref.count--;
 455         }
 456         if (held1 && held2 && list_empty(&gl->gl_holders))
 457                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 458
 459         if (new_state != gl->gl_target)
 460                 /* shorten our minimum hold time */
 461                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 462                                        GL_GLOCK_MIN_HOLD);
 463         gl->gl_state = new_state;
 464         gl->gl_tchange = jiffies;
 465 }
 466
 467 static void gfs2_set_demote(struct gfs2_glock *gl)
 468 {
 469         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 470
 471         set_bit(GLF_DEMOTE, &gl->gl_flags);
 472         smp_mb();
 473         wake_up(&sdp->sd_async_glock_wait);
 474 }
 475
 476 static void gfs2_demote_wake(struct gfs2_glock *gl)
 477 {
 478         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 479         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 480         smp_mb__after_atomic();
 481         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 482 }
 483
 484 /**
 485  * finish_xmote - The DLM has replied to one of our lock requests
 486  * @gl: The glock
 487  * @ret: The status from the DLM
 488  *
 489  */
 490
 491 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 492 {
 493         const struct gfs2_glock_operations *glops = gl->gl_ops;
 494         struct gfs2_holder *gh;
 495         unsigned state = ret & LM_OUT_ST_MASK;
 496         int rv;
 497
 498         spin_lock(&gl->gl_lockref.lock);
 499         trace_gfs2_glock_state_change(gl, state);
 500         state_change(gl, state);
 501         gh = find_first_waiter(gl);
 502
 503         /* Demote to UN request arrived during demote to SH or DF */
 504         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 505             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 506                 gl->gl_target = LM_ST_UNLOCKED;
 507
 508         /* Check for state != intended state */
 509         if (unlikely(state != gl->gl_target)) {
 510                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 511                         /* move to back of queue and try next entry */
 512                         if (ret & LM_OUT_CANCELED) {
 513                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 514                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 515                                 gh = find_first_waiter(gl);
 516                                 gl->gl_target = gh->gh_state;
 517                                 goto retry;
 518                         }
 519                         /* Some error or failed "try lock" - report it */
 520                         if ((ret & LM_OUT_ERROR) ||
 521                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 522                                 gl->gl_target = gl->gl_state;
 523                                 do_error(gl, ret);
 524                                 goto out;
 525                         }
 526                 }
 527                 switch(state) {
 528                 /* Unlocked due to conversion deadlock, try again */
 529                 case LM_ST_UNLOCKED:
 530 retry:
 531                         do_xmote(gl, gh, gl->gl_target);
 532                         break;
 533                 /* Conversion fails, unlock and try again */
 534                 case LM_ST_SHARED:
 535                 case LM_ST_DEFERRED:
 536                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 537                         break;
 538                 default: /* Everything else */
 539                         fs_err(gl->gl_name.ln_sbd, "wanted %u got %u\n",
 540                                gl->gl_target, state);
 541                         GLOCK_BUG_ON(gl, 1);
 542                 }
 543                 spin_unlock(&gl->gl_lockref.lock);
 544                 return;
 545         }
 546
 547         /* Fast path - we got what we asked for */
 548         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 549                 gfs2_demote_wake(gl);
 550         if (state != LM_ST_UNLOCKED) {
 551                 if (glops->go_xmote_bh) {
 552                         spin_unlock(&gl->gl_lockref.lock);
 553                         rv = glops->go_xmote_bh(gl, gh);
 554                         spin_lock(&gl->gl_lockref.lock);
 555                         if (rv) {
 556                                 do_error(gl, rv);
 557                                 goto out;
 558                         }
 559                 }
 560                 rv = do_promote(gl);
 561                 if (rv == 2)
 562                         goto out_locked;
 563         }
 564 out:
 565         clear_bit(GLF_LOCK, &gl->gl_flags);
 566 out_locked:
 567         spin_unlock(&gl->gl_lockref.lock);
 568 }
 569
 570 /**
 571  * do_xmote - Calls the DLM to change the state of a lock
 572  * @gl: The lock state
 573  * @gh: The holder (only for promotes)
 574  * @target: The target lock state
 575  *
 576  */
 577
 578 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 579 __releases(&gl->gl_lockref.lock)
 580 __acquires(&gl->gl_lockref.lock)
 581 {
 582         const struct gfs2_glock_operations *glops = gl->gl_ops;
 583         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 584         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 585         int ret;
 586
 587         if (target != LM_ST_UNLOCKED && glock_blocked_by_withdraw(gl) &&
 588             gh && !(gh->gh_flags & LM_FLAG_NOEXP))
 589                 return;
 590         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 591                       LM_FLAG_PRIORITY);
 592         GLOCK_BUG_ON(gl, gl->gl_state == target);
 593         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 594         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 595             glops->go_inval) {
 596                 /*
 597                  * If another process is already doing the invalidate, let that
 598                  * finish first.  The glock state machine will get back to this
 599                  * holder again later.
 600                  */
 601                 if (test_and_set_bit(GLF_INVALIDATE_IN_PROGRESS,
 602                                      &gl->gl_flags))
 603                         return;
 604                 do_error(gl, 0); /* Fail queued try locks */
 605         }
 606         gl->gl_req = target;
 607         set_bit(GLF_BLOCKING, &gl->gl_flags);
 608         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 609             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 610             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 611                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 612         spin_unlock(&gl->gl_lockref.lock);
 613         if (glops->go_sync) {
 614                 ret = glops->go_sync(gl);
 615                 /* If we had a problem syncing (due to io errors or whatever,
 616                  * we should not invalidate the metadata or tell dlm to
 617                  * release the glock to other nodes.
 618                  */
 619                 if (ret) {
 620                         if (cmpxchg(&sdp->sd_log_error, 0, ret)) {
 621                                 fs_err(sdp, "Error %d syncing glock \n", ret);
 622                                 gfs2_dump_glock(NULL, gl, true);
 623                         }
 624                         goto skip_inval;
 625                 }
 626         }
 627         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags)) {
 628                 /*
 629                  * The call to go_sync should have cleared out the ail list.
 630                  * If there are still items, we have a problem. We ought to
 631                  * withdraw, but we can't because the withdraw code also uses
 632                  * glocks. Warn about the error, dump the glock, then fall
 633                  * through and wait for logd to do the withdraw for us.
 634                  */
 635                 if ((atomic_read(&gl->gl_ail_count) != 0) &&
 636                     (!cmpxchg(&sdp->sd_log_error, 0, -EIO))) {
 637                         gfs2_glock_assert_warn(gl,
 638                                                !atomic_read(&gl->gl_ail_count));
 639                         gfs2_dump_glock(NULL, gl, true);
 640                 }
 641                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 642                 clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 643         }
 644
 645 skip_inval:
 646         gfs2_glock_hold(gl);
 647         /*
 648          * Check for an error encountered since we called go_sync and go_inval.
 649          * If so, we can't withdraw from the glock code because the withdraw
 650          * code itself uses glocks (see function signal_our_withdraw) to
 651          * change the mount to read-only. Most importantly, we must not call
 652          * dlm to unlock the glock until the journal is in a known good state
 653          * (after journal replay) otherwise other nodes may use the object
 654          * (rgrp or dinode) and then later, journal replay will corrupt the
 655          * file system. The best we can do here is wait for the logd daemon
 656          * to see sd_log_error and withdraw, and in the meantime, requeue the
 657          * work for later.
 658          *
 659          * However, if we're just unlocking the lock (say, for unmount, when
 660          * gfs2_gl_hash_clear calls clear_glock) and recovery is complete
 661          * then it's okay to tell dlm to unlock it.
 662          */
 663         if (unlikely(sdp->sd_log_error && !gfs2_withdrawn(sdp)))
 664                 gfs2_withdraw_delayed(sdp);
 665         if (glock_blocked_by_withdraw(gl)) {
 666                 if (target != LM_ST_UNLOCKED ||
 667                     test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags)) {
 668                         gfs2_glock_queue_work(gl, GL_GLOCK_DFT_HOLD);
 669                         goto out;
 670                 }
 671         }
 672
 673         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 674                 /* lock_dlm */
 675                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 676                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 677                     target == LM_ST_UNLOCKED &&
 678                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 679                         finish_xmote(gl, target);
 680                         gfs2_glock_queue_work(gl, 0);
 681                 } else if (ret) {
 682                         fs_err(sdp, "lm_lock ret %d\n", ret);
 683                         GLOCK_BUG_ON(gl, !gfs2_withdrawn(sdp));
 684                 }
 685         } else { /* lock_nolock */
 686                 finish_xmote(gl, target);
 687                 gfs2_glock_queue_work(gl, 0);
 688         }
 689 out:
 690         spin_lock(&gl->gl_lockref.lock);
 691 }
 692
 693 /**
 694  * find_first_holder - find the first "holder" gh
 695  * @gl: the glock
 696  */
 697
 698 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 699 {
 700         struct gfs2_holder *gh;
 701
 702         if (!list_empty(&gl->gl_holders)) {
 703                 gh = list_first_entry(&gl->gl_holders, struct gfs2_holder, gh_list);
 704                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 705                         return gh;
 706         }
 707         return NULL;
 708 }
 709
 710 /**
 711  * run_queue - do all outstanding tasks related to a glock
 712  * @gl: The glock in question
 713  * @nonblock: True if we must not block in run_queue
 714  *
 715  */
 716
 717 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 718 __releases(&gl->gl_lockref.lock)
 719 __acquires(&gl->gl_lockref.lock)
 720 {
 721         struct gfs2_holder *gh = NULL;
 722         int ret;
 723
 724         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 725                 return;
 726
 727         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 728
 729         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 730             gl->gl_demote_state != gl->gl_state) {
 731                 if (find_first_holder(gl))
 732                         goto out_unlock;
 733                 if (nonblock)
 734                         goto out_sched;
 735                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 736                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 737                 gl->gl_target = gl->gl_demote_state;
 738         } else {
 739                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 740                         gfs2_demote_wake(gl);
 741                 ret = do_promote(gl);
 742                 if (ret == 0)
 743                         goto out_unlock;
 744                 if (ret == 2)
 745                         goto out;
 746                 gh = find_first_waiter(gl);
 747                 gl->gl_target = gh->gh_state;
 748                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 749                         do_error(gl, 0); /* Fail queued try locks */
 750         }
 751         do_xmote(gl, gh, gl->gl_target);
 752 out:
 753         return;
 754
 755 out_sched:
 756         clear_bit(GLF_LOCK, &gl->gl_flags);
 757         smp_mb__after_atomic();
 758         gl->gl_lockref.count++;
 759         __gfs2_glock_queue_work(gl, 0);
 760         return;
 761
 762 out_unlock:
 763         clear_bit(GLF_LOCK, &gl->gl_flags);
 764         smp_mb__after_atomic();
 765         return;
 766 }
 767
 768 void gfs2_inode_remember_delete(struct gfs2_glock *gl, u64 generation)
 769 {
 770         struct gfs2_inode_lvb *ri = (void *)gl->gl_lksb.sb_lvbptr;
 771
 772         if (ri->ri_magic == 0)
 773                 ri->ri_magic = cpu_to_be32(GFS2_MAGIC);
 774         if (ri->ri_magic == cpu_to_be32(GFS2_MAGIC))
 775                 ri->ri_generation_deleted = cpu_to_be64(generation);
 776 }
 777
 778 bool gfs2_inode_already_deleted(struct gfs2_glock *gl, u64 generation)
 779 {
 780         struct gfs2_inode_lvb *ri = (void *)gl->gl_lksb.sb_lvbptr;
 781
 782         if (ri->ri_magic != cpu_to_be32(GFS2_MAGIC))
 783                 return false;
 784         return generation <= be64_to_cpu(ri->ri_generation_deleted);
 785 }
 786
 787 static void gfs2_glock_poke(struct gfs2_glock *gl)
 788 {
 789         int flags = LM_FLAG_TRY_1CB | LM_FLAG_ANY | GL_SKIP;
 790         struct gfs2_holder gh;
 791         int error;
 792
 793         error = gfs2_glock_nq_init(gl, LM_ST_SHARED, flags, &gh);
 794         if (!error)
 795                 gfs2_glock_dq(&gh);
 796 }
 797
 798 static bool gfs2_try_evict(struct gfs2_glock *gl)
 799 {
 800         struct gfs2_inode *ip;
 801         bool evicted = false;
 802
 803         /*
 804          * If there is contention on the iopen glock and we have an inode, try
 805          * to grab and release the inode so that it can be evicted.  This will
 806          * allow the remote node to go ahead and delete the inode without us
 807          * having to do it, which will avoid rgrp glock thrashing.
 808          *
 809          * The remote node is likely still holding the corresponding inode
 810          * glock, so it will run before we get to verify that the delete has
 811          * happened below.
 812          */
 813         spin_lock(&gl->gl_lockref.lock);
 814         ip = gl->gl_object;
 815         if (ip && !igrab(&ip->i_inode))
 816                 ip = NULL;
 817         spin_unlock(&gl->gl_lockref.lock);
 818         if (ip) {
 819                 struct gfs2_glock *inode_gl = NULL;
 820
 821                 gl->gl_no_formal_ino = ip->i_no_formal_ino;
 822                 set_bit(GIF_DEFERRED_DELETE, &ip->i_flags);
 823                 d_prune_aliases(&ip->i_inode);
 824                 iput(&ip->i_inode);
 825
 826                 /* If the inode was evicted, gl->gl_object will now be NULL. */
 827                 spin_lock(&gl->gl_lockref.lock);
 828                 ip = gl->gl_object;
 829                 if (ip) {
 830                         inode_gl = ip->i_gl;
 831                         lockref_get(&inode_gl->gl_lockref);
 832                         clear_bit(GIF_DEFERRED_DELETE, &ip->i_flags);
 833                 }
 834                 spin_unlock(&gl->gl_lockref.lock);
 835                 if (inode_gl) {
 836                         gfs2_glock_poke(inode_gl);
 837                         gfs2_glock_put(inode_gl);
 838                 }
 839                 evicted = !ip;
 840         }
 841         return evicted;
 842 }
 843
 844 static void delete_work_func(struct work_struct *work)
 845 {
 846         struct delayed_work *dwork = to_delayed_work(work);
 847         struct gfs2_glock *gl = container_of(dwork, struct gfs2_glock, gl_delete);
 848         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 849         struct inode *inode;
 850         u64 no_addr = gl->gl_name.ln_number;
 851
 852         spin_lock(&gl->gl_lockref.lock);
 853         clear_bit(GLF_PENDING_DELETE, &gl->gl_flags);
 854         spin_unlock(&gl->gl_lockref.lock);
 855
 856         /* If someone's using this glock to create a new dinode, the block must
 857            have been freed by another node, then re-used, in which case our
 858            iopen callback is too late after the fact. Ignore it. */
 859         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 860                 goto out;
 861
 862         if (test_bit(GLF_DEMOTE, &gl->gl_flags)) {
 863                 /*
 864                  * If we can evict the inode, give the remote node trying to
 865                  * delete the inode some time before verifying that the delete
 866                  * has happened.  Otherwise, if we cause contention on the inode glock
 867                  * immediately, the remote node will think that we still have
 868                  * the inode in use, and so it will give up waiting.
 869                  *
 870                  * If we can't evict the inode, signal to the remote node that
 871                  * the inode is still in use.  We'll later try to delete the
 872                  * inode locally in gfs2_evict_inode.
 873                  *
 874                  * FIXME: We only need to verify that the remote node has
 875                  * deleted the inode because nodes before this remote delete
 876                  * rework won't cooperate.  At a later time, when we no longer
 877                  * care about compatibility with such nodes, we can skip this
 878                  * step entirely.
 879                  */
 880                 if (gfs2_try_evict(gl)) {
 881                         if (gfs2_queue_delete_work(gl, 5 * HZ))
 882                                 return;
 883                 }
 884                 goto out;
 885         }
 886
 887         inode = gfs2_lookup_by_inum(sdp, no_addr, gl->gl_no_formal_ino,
 888                                     GFS2_BLKST_UNLINKED);
 889         if (!IS_ERR_OR_NULL(inode)) {
 890                 d_prune_aliases(inode);
 891                 iput(inode);
 892         }
 893 out:
 894         gfs2_glock_put(gl);
 895 }
 896
 897 static void glock_work_func(struct work_struct *work)
 898 {
 899         unsigned long delay = 0;
 900         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 901         unsigned int drop_refs = 1;
 902
 903         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 904                 finish_xmote(gl, gl->gl_reply);
 905                 drop_refs++;
 906         }
 907         spin_lock(&gl->gl_lockref.lock);
 908         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 909             gl->gl_state != LM_ST_UNLOCKED &&
 910             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 911                 unsigned long holdtime, now = jiffies;
 912
 913                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 914                 if (time_before(now, holdtime))
 915                         delay = holdtime - now;
 916
 917                 if (!delay) {
 918                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 919                         gfs2_set_demote(gl);
 920                 }
 921         }
 922         run_queue(gl, 0);
 923         if (delay) {
 924                 /* Keep one glock reference for the work we requeue. */
 925                 drop_refs--;
 926                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 927                         delay = 0;
 928                 __gfs2_glock_queue_work(gl, delay);
 929         }
 930
 931         /*
 932          * Drop the remaining glock references manually here. (Mind that
 933          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 934          * here as well.)
 935          */
 936         gl->gl_lockref.count -= drop_refs;
 937         if (!gl->gl_lockref.count) {
 938                 __gfs2_glock_put(gl);
 939                 return;
 940         }
 941         spin_unlock(&gl->gl_lockref.lock);
 942 }
 943
 944 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 945                                             struct gfs2_glock *new)
 946 {
 947         struct wait_glock_queue wait;
 948         wait_queue_head_t *wq = glock_waitqueue(name);
 949         struct gfs2_glock *gl;
 950
 951         wait.name = name;
 952         init_wait(&wait.wait);
 953         wait.wait.func = glock_wake_function;
 954
 955 again:
 956         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 957         rcu_read_lock();
 958         if (new) {
 959                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 960                         &new->gl_node, ht_parms);
 961                 if (IS_ERR(gl))
 962                         goto out;
 963         } else {
 964                 gl = rhashtable_lookup_fast(&gl_hash_table,
 965                         name, ht_parms);
 966         }
 967         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 968                 rcu_read_unlock();
 969                 schedule();
 970                 goto again;
 971         }
 972 out:
 973         rcu_read_unlock();
 974         finish_wait(wq, &wait.wait);
 975         return gl;
 976 }
 977
 978 /**
 979  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 980  * @sdp: The GFS2 superblock
 981  * @number: the lock number
 982  * @glops: The glock_operations to use
 983  * @create: If 0, don't create the glock if it doesn't exist
 984  * @glp: the glock is returned here
 985  *
 986  * This does not lock a glock, just finds/creates structures for one.
 987  *
 988  * Returns: errno
 989  */
 990
 991 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 992                    const struct gfs2_glock_operations *glops, int create,
 993                    struct gfs2_glock **glp)
 994 {
 995         struct super_block *s = sdp->sd_vfs;
 996         struct lm_lockname name = { .ln_number = number,
 997                                     .ln_type = glops->go_type,
 998                                     .ln_sbd = sdp };
 999         struct gfs2_glock *gl, *tmp;
1000         struct address_space *mapping;
1001         struct kmem_cache *cachep;
1002         int ret = 0;
1003
1004         gl = find_insert_glock(&name, NULL);
1005         if (gl) {
1006                 *glp = gl;
1007                 return 0;
1008         }
1009         if (!create)
1010                 return -ENOENT;
1011
1012         if (glops->go_flags & GLOF_ASPACE)
1013                 cachep = gfs2_glock_aspace_cachep;
1014         else
1015                 cachep = gfs2_glock_cachep;
1016         gl = kmem_cache_alloc(cachep, GFP_NOFS);
1017         if (!gl)
1018                 return -ENOMEM;
1019
1020         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
1021
1022         if (glops->go_flags & GLOF_LVB) {
1023                 gl->gl_lksb.sb_lvbptr = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1024                 if (!gl->gl_lksb.sb_lvbptr) {
1025                         kmem_cache_free(cachep, gl);
1026                         return -ENOMEM;
1027                 }
1028         }
1029
1030         atomic_inc(&sdp->sd_glock_disposal);
1031         gl->gl_node.next = NULL;
1032         gl->gl_flags = 0;
1033         gl->gl_name = name;
1034         gl->gl_lockref.count = 1;
1035         gl->gl_state = LM_ST_UNLOCKED;
1036         gl->gl_target = LM_ST_UNLOCKED;
1037         gl->gl_demote_state = LM_ST_EXCLUSIVE;
1038         gl->gl_ops = glops;
1039         gl->gl_dstamp = 0;
1040         preempt_disable();
1041         /* We use the global stats to estimate the initial per-glock stats */
1042         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
1043         preempt_enable();
1044         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
1045         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
1046         gl->gl_tchange = jiffies;
1047         gl->gl_object = NULL;
1048         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
1049         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
1050         INIT_DELAYED_WORK(&gl->gl_delete, delete_work_func);
1051
1052         mapping = gfs2_glock2aspace(gl);
1053         if (mapping) {
1054                 mapping->a_ops = &gfs2_meta_aops;
1055                 mapping->host = s->s_bdev->bd_inode;
1056                 mapping->flags = 0;
1057                 mapping_set_gfp_mask(mapping, GFP_NOFS);
1058                 mapping->private_data = NULL;
1059                 mapping->writeback_index = 0;
1060         }
1061
1062         tmp = find_insert_glock(&name, gl);
1063         if (!tmp) {
1064                 *glp = gl;
1065                 goto out;
1066         }
1067         if (IS_ERR(tmp)) {
1068                 ret = PTR_ERR(tmp);
1069                 goto out_free;
1070         }
1071         *glp = tmp;
1072
1073 out_free:
1074         kfree(gl->gl_lksb.sb_lvbptr);
1075         kmem_cache_free(cachep, gl);
1076         atomic_dec(&sdp->sd_glock_disposal);
1077
1078 out:
1079         return ret;
1080 }
1081
1082 /**
1083  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
1084  * @gl: the glock
1085  * @state: the state we're requesting
1086  * @flags: the modifier flags
1087  * @gh: the holder structure
1088  *
1089  */
1090
1091 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
1092                       struct gfs2_holder *gh)
1093 {
1094         INIT_LIST_HEAD(&gh->gh_list);
1095         gh->gh_gl = gl;
1096         gh->gh_ip = _RET_IP_;
1097         gh->gh_owner_pid = get_pid(task_pid(current));
1098         gh->gh_state = state;
1099         gh->gh_flags = flags;
1100         gh->gh_error = 0;
1101         gh->gh_iflags = 0;
1102         gfs2_glock_hold(gl);
1103 }
1104
1105 /**
1106  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
1107  * @state: the state we're requesting
1108  * @flags: the modifier flags
1109  * @gh: the holder structure
1110  *
1111  * Don't mess with the glock.
1112  *
1113  */
1114
1115 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
1116 {
1117         gh->gh_state = state;
1118         gh->gh_flags = flags;
1119         gh->gh_iflags = 0;
1120         gh->gh_ip = _RET_IP_;
1121         put_pid(gh->gh_owner_pid);
1122         gh->gh_owner_pid = get_pid(task_pid(current));
1123 }
1124
1125 /**
1126  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
1127  * @gh: the holder structure
1128  *
1129  */
1130
1131 void gfs2_holder_uninit(struct gfs2_holder *gh)
1132 {
1133         put_pid(gh->gh_owner_pid);
1134         gfs2_glock_put(gh->gh_gl);
1135         gfs2_holder_mark_uninitialized(gh);
1136         gh->gh_ip = 0;
1137 }
1138
1139 static void gfs2_glock_update_hold_time(struct gfs2_glock *gl,
1140                                         unsigned long start_time)
1141 {
1142         /* Have we waited longer that a second? */
1143         if (time_after(jiffies, start_time + HZ)) {
1144                 /* Lengthen the minimum hold time. */
1145                 gl->gl_hold_time = min(gl->gl_hold_time + GL_GLOCK_HOLD_INCR,
1146                                        GL_GLOCK_MAX_HOLD);
1147         }
1148 }
1149
1150 /**
1151  * gfs2_glock_wait - wait on a glock acquisition
1152  * @gh: the glock holder
1153  *
1154  * Returns: 0 on success
1155  */
1156
1157 int gfs2_glock_wait(struct gfs2_holder *gh)
1158 {
1159         unsigned long start_time = jiffies;
1160
1161         might_sleep();
1162         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
1163         gfs2_glock_update_hold_time(gh->gh_gl, start_time);
1164         return gh->gh_error;
1165 }
1166
1167 static int glocks_pending(unsigned int num_gh, struct gfs2_holder *ghs)
1168 {
1169         int i;
1170
1171         for (i = 0; i < num_gh; i++)
1172                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags))
1173                         return 1;
1174         return 0;
1175 }
1176
1177 /**
1178  * gfs2_glock_async_wait - wait on multiple asynchronous glock acquisitions
1179  * @num_gh: the number of holders in the array
1180  * @ghs: the glock holder array
1181  *
1182  * Returns: 0 on success, meaning all glocks have been granted and are held.
1183  *          -ESTALE if the request timed out, meaning all glocks were released,
1184  *          and the caller should retry the operation.
1185  */
1186
1187 int gfs2_glock_async_wait(unsigned int num_gh, struct gfs2_holder *ghs)
1188 {
1189         struct gfs2_sbd *sdp = ghs[0].gh_gl->gl_name.ln_sbd;
1190         int i, ret = 0, timeout = 0;
1191         unsigned long start_time = jiffies;
1192         bool keep_waiting;
1193
1194         might_sleep();
1195         /*
1196          * Total up the (minimum hold time * 2) of all glocks and use that to
1197          * determine the max amount of time we should wait.
1198          */
1199         for (i = 0; i < num_gh; i++)
1200                 timeout += ghs[i].gh_gl->gl_hold_time << 1;
1201
1202 wait_for_dlm:
1203         if (!wait_event_timeout(sdp->sd_async_glock_wait,
1204                                 !glocks_pending(num_gh, ghs), timeout))
1205                 ret = -ESTALE; /* request timed out. */
1206
1207         /*
1208          * If dlm granted all our requests, we need to adjust the glock
1209          * minimum hold time values according to how long we waited.
1210          *
1211          * If our request timed out, we need to repeatedly release any held
1212          * glocks we acquired thus far to allow dlm to acquire the remaining
1213          * glocks without deadlocking.  We cannot currently cancel outstanding
1214          * glock acquisitions.
1215          *
1216          * The HIF_WAIT bit tells us which requests still need a response from
1217          * dlm.
1218          *
1219          * If dlm sent us any errors, we return the first error we find.
1220          */
1221         keep_waiting = false;
1222         for (i = 0; i < num_gh; i++) {
1223                 /* Skip holders we have already dequeued below. */
1224                 if (!gfs2_holder_queued(&ghs[i]))
1225                         continue;
1226                 /* Skip holders with a pending DLM response. */
1227                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags)) {
1228                         keep_waiting = true;
1229                         continue;
1230                 }
1231
1232                 if (test_bit(HIF_HOLDER, &ghs[i].gh_iflags)) {
1233                         if (ret == -ESTALE)
1234                                 gfs2_glock_dq(&ghs[i]);
1235                         else
1236                                 gfs2_glock_update_hold_time(ghs[i].gh_gl,
1237                                                             start_time);
1238                 }
1239                 if (!ret)
1240                         ret = ghs[i].gh_error;
1241         }
1242
1243         if (keep_waiting)
1244                 goto wait_for_dlm;
1245
1246         /*
1247          * At this point, we've either acquired all locks or released them all.
1248          */
1249         return ret;
1250 }
1251
1252 /**
1253  * handle_callback - process a demote request
1254  * @gl: the glock
1255  * @state: the state the caller wants us to change to
1256  *
1257  * There are only two requests that we are going to see in actual
1258  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
1259  */
1260
1261 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
1262                             unsigned long delay, bool remote)
1263 {
1264         if (delay)
1265                 set_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
1266         else
1267                 gfs2_set_demote(gl);
1268         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
1269                 gl->gl_demote_state = state;
1270                 gl->gl_demote_time = jiffies;
1271         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
1272                         gl->gl_demote_state != state) {
1273                 gl->gl_demote_state = LM_ST_UNLOCKED;
1274         }
1275         if (gl->gl_ops->go_callback)
1276                 gl->gl_ops->go_callback(gl, remote);
1277         trace_gfs2_demote_rq(gl, remote);
1278 }
1279
1280 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
1281 {
1282         struct va_format vaf;
1283         va_list args;
1284
1285         va_start(args, fmt);
1286
1287         if (seq) {
1288                 seq_vprintf(seq, fmt, args);
1289         } else {
1290                 vaf.fmt = fmt;
1291                 vaf.va = &args;
1292
1293                 pr_err("%pV", &vaf);
1294         }
1295
1296         va_end(args);
1297 }
1298
1299 /**
1300  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1301  * @gh: the holder structure to add
1302  *
1303  * Eventually we should move the recursive locking trap to a
1304  * debugging option or something like that. This is the fast
1305  * path and needs to have the minimum number of distractions.
1306  *
1307  */
1308
1309 static inline void add_to_queue(struct gfs2_holder *gh)
1310 __releases(&gl->gl_lockref.lock)
1311 __acquires(&gl->gl_lockref.lock)
1312 {
1313         struct gfs2_glock *gl = gh->gh_gl;
1314         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1315         struct list_head *insert_pt = NULL;
1316         struct gfs2_holder *gh2;
1317         int try_futile = 0;
1318
1319         GLOCK_BUG_ON(gl, gh->gh_owner_pid == NULL);
1320         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1321                 GLOCK_BUG_ON(gl, true);
1322
1323         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1324                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1325                         try_futile = !may_grant(gl, gh);
1326                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1327                         goto fail;
1328         }
1329
1330         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1331                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1332                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1333                         goto trap_recursive;
1334                 if (try_futile &&
1335                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1336 fail:
1337                         gh->gh_error = GLR_TRYFAILED;
1338                         gfs2_holder_wake(gh);
1339                         return;
1340                 }
1341                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1342                         continue;
1343                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1344                         insert_pt = &gh2->gh_list;
1345         }
1346         set_bit(GLF_QUEUED, &gl->gl_flags);
1347         trace_gfs2_glock_queue(gh, 1);
1348         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1349         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1350         if (likely(insert_pt == NULL)) {
1351                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1352                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1353                         goto do_cancel;
1354                 return;
1355         }
1356         list_add_tail(&gh->gh_list, insert_pt);
1357 do_cancel:
1358         gh = list_first_entry(&gl->gl_holders, struct gfs2_holder, gh_list);
1359         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1360                 spin_unlock(&gl->gl_lockref.lock);
1361                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1362                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1363                 spin_lock(&gl->gl_lockref.lock);
1364         }
1365         return;
1366
1367 trap_recursive:
1368         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1369         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1370         fs_err(sdp, "lock type: %d req lock state : %d\n",
1371                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1372         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1373         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1374         fs_err(sdp, "lock type: %d req lock state : %d\n",
1375                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1376         gfs2_dump_glock(NULL, gl, true);
1377         BUG();
1378 }
1379
1380 /**
1381  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1382  * @gh: the holder structure
1383  *
1384  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1385  *
1386  * Returns: 0, GLR_TRYFAILED, or errno on failure
1387  */
1388
1389 int gfs2_glock_nq(struct gfs2_holder *gh)
1390 {
1391         struct gfs2_glock *gl = gh->gh_gl;
1392         int error = 0;
1393
1394         if (glock_blocked_by_withdraw(gl) && !(gh->gh_flags & LM_FLAG_NOEXP))
1395                 return -EIO;
1396
1397         if (test_bit(GLF_LRU, &gl->gl_flags))
1398                 gfs2_glock_remove_from_lru(gl);
1399
1400         spin_lock(&gl->gl_lockref.lock);
1401         add_to_queue(gh);
1402         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1403                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1404                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1405                 gl->gl_lockref.count++;
1406                 __gfs2_glock_queue_work(gl, 0);
1407         }
1408         run_queue(gl, 1);
1409         spin_unlock(&gl->gl_lockref.lock);
1410
1411         if (!(gh->gh_flags & GL_ASYNC))
1412                 error = gfs2_glock_wait(gh);
1413
1414         return error;
1415 }
1416
1417 /**
1418  * gfs2_glock_poll - poll to see if an async request has been completed
1419  * @gh: the holder
1420  *
1421  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1422  */
1423
1424 int gfs2_glock_poll(struct gfs2_holder *gh)
1425 {
1426         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1427 }
1428
1429 /**
1430  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1431  * @gh: the glock holder
1432  *
1433  */
1434
1435 void gfs2_glock_dq(struct gfs2_holder *gh)
1436 {
1437         struct gfs2_glock *gl = gh->gh_gl;
1438         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1439         unsigned delay = 0;
1440         int fast_path = 0;
1441
1442         spin_lock(&gl->gl_lockref.lock);
1443         /*
1444          * If we're in the process of file system withdraw, we cannot just
1445          * dequeue any glocks until our journal is recovered, lest we
1446          * introduce file system corruption. We need two exceptions to this
1447          * rule: We need to allow unlocking of nondisk glocks and the glock
1448          * for our own journal that needs recovery.
1449          */
1450         if (test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags) &&
1451             glock_blocked_by_withdraw(gl) &&
1452             gh->gh_gl != sdp->sd_jinode_gl) {
1453                 sdp->sd_glock_dqs_held++;
1454                 might_sleep();
1455                 wait_on_bit(&sdp->sd_flags, SDF_WITHDRAW_RECOVERY,
1456                             TASK_UNINTERRUPTIBLE);
1457         }
1458         if (gh->gh_flags & GL_NOCACHE)
1459                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1460
1461         list_del_init(&gh->gh_list);
1462         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1463         if (find_first_holder(gl) == NULL) {
1464                 if (list_empty(&gl->gl_holders) &&
1465                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1466                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1467                         fast_path = 1;
1468         }
1469         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1470                 gfs2_glock_add_to_lru(gl);
1471
1472         trace_gfs2_glock_queue(gh, 0);
1473         if (unlikely(!fast_path)) {
1474                 gl->gl_lockref.count++;
1475                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1476                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1477                     gl->gl_name.ln_type == LM_TYPE_INODE)
1478                         delay = gl->gl_hold_time;
1479                 __gfs2_glock_queue_work(gl, delay);
1480         }
1481         spin_unlock(&gl->gl_lockref.lock);
1482 }
1483
1484 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1485 {
1486         struct gfs2_glock *gl = gh->gh_gl;
1487         gfs2_glock_dq(gh);
1488         might_sleep();
1489         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1490 }
1491
1492 /**
1493  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1494  * @gh: the holder structure
1495  *
1496  */
1497
1498 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1499 {
1500         gfs2_glock_dq(gh);
1501         gfs2_holder_uninit(gh);
1502 }
1503
1504 /**
1505  * gfs2_glock_nq_num - acquire a glock based on lock number
1506  * @sdp: the filesystem
1507  * @number: the lock number
1508  * @glops: the glock operations for the type of glock
1509  * @state: the state to acquire the glock in
1510  * @flags: modifier flags for the acquisition
1511  * @gh: the struct gfs2_holder
1512  *
1513  * Returns: errno
1514  */
1515
1516 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1517                       const struct gfs2_glock_operations *glops,
1518                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1519 {
1520         struct gfs2_glock *gl;
1521         int error;
1522
1523         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1524         if (!error) {
1525                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1526                 gfs2_glock_put(gl);
1527         }
1528
1529         return error;
1530 }
1531
1532 /**
1533  * glock_compare - Compare two struct gfs2_glock structures for sorting
1534  * @arg_a: the first structure
1535  * @arg_b: the second structure
1536  *
1537  */
1538
1539 static int glock_compare(const void *arg_a, const void *arg_b)
1540 {
1541         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1542         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1543         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1544         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1545
1546         if (a->ln_number > b->ln_number)
1547                 return 1;
1548         if (a->ln_number < b->ln_number)
1549                 return -1;
1550         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1551         return 0;
1552 }
1553
1554 /**
1555  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1556  * @num_gh: the number of structures
1557  * @ghs: an array of struct gfs2_holder structures
1558  *
1559  * Returns: 0 on success (all glocks acquired),
1560  *          errno on failure (no glocks acquired)
1561  */
1562
1563 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1564                      struct gfs2_holder **p)
1565 {
1566         unsigned int x;
1567         int error = 0;
1568
1569         for (x = 0; x < num_gh; x++)
1570                 p[x] = &ghs[x];
1571
1572         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1573
1574         for (x = 0; x < num_gh; x++) {
1575                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1576
1577                 error = gfs2_glock_nq(p[x]);
1578                 if (error) {
1579                         while (x--)
1580                                 gfs2_glock_dq(p[x]);
1581                         break;
1582                 }
1583         }
1584
1585         return error;
1586 }
1587
1588 /**
1589  * gfs2_glock_nq_m - acquire multiple glocks
1590  * @num_gh: the number of structures
1591  * @ghs: an array of struct gfs2_holder structures
1592  *
1593  *
1594  * Returns: 0 on success (all glocks acquired),
1595  *          errno on failure (no glocks acquired)
1596  */
1597
1598 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1599 {
1600         struct gfs2_holder *tmp[4];
1601         struct gfs2_holder **pph = tmp;
1602         int error = 0;
1603
1604         switch(num_gh) {
1605         case 0:
1606                 return 0;
1607         case 1:
1608                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1609                 return gfs2_glock_nq(ghs);
1610         default:
1611                 if (num_gh <= 4)
1612                         break;
1613                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1614                                     GFP_NOFS);
1615                 if (!pph)
1616                         return -ENOMEM;
1617         }
1618
1619         error = nq_m_sync(num_gh, ghs, pph);
1620
1621         if (pph != tmp)
1622                 kfree(pph);
1623
1624         return error;
1625 }
1626
1627 /**
1628  * gfs2_glock_dq_m - release multiple glocks
1629  * @num_gh: the number of structures
1630  * @ghs: an array of struct gfs2_holder structures
1631  *
1632  */
1633
1634 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1635 {
1636         while (num_gh--)
1637                 gfs2_glock_dq(&ghs[num_gh]);
1638 }
1639
1640 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1641 {
1642         unsigned long delay = 0;
1643         unsigned long holdtime;
1644         unsigned long now = jiffies;
1645
1646         gfs2_glock_hold(gl);
1647         holdtime = gl->gl_tchange + gl->gl_hold_time;
1648         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1649             gl->gl_name.ln_type == LM_TYPE_INODE) {
1650                 if (time_before(now, holdtime))
1651                         delay = holdtime - now;
1652                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1653                         delay = gl->gl_hold_time;
1654         }
1655
1656         spin_lock(&gl->gl_lockref.lock);
1657         handle_callback(gl, state, delay, true);
1658         __gfs2_glock_queue_work(gl, delay);
1659         spin_unlock(&gl->gl_lockref.lock);
1660 }
1661
1662 /**
1663  * gfs2_should_freeze - Figure out if glock should be frozen
1664  * @gl: The glock in question
1665  *
1666  * Glocks are not frozen if (a) the result of the dlm operation is
1667  * an error, (b) the locking operation was an unlock operation or
1668  * (c) if there is a "noexp" flagged request anywhere in the queue
1669  *
1670  * Returns: 1 if freezing should occur, 0 otherwise
1671  */
1672
1673 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1674 {
1675         const struct gfs2_holder *gh;
1676
1677         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1678                 return 0;
1679         if (gl->gl_target == LM_ST_UNLOCKED)
1680                 return 0;
1681
1682         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1683                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1684                         continue;
1685                 if (LM_FLAG_NOEXP & gh->gh_flags)
1686                         return 0;
1687         }
1688
1689         return 1;
1690 }
1691
1692 /**
1693  * gfs2_glock_complete - Callback used by locking
1694  * @gl: Pointer to the glock
1695  * @ret: The return value from the dlm
1696  *
1697  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1698  * to use a bitfield shared with other glock state fields.
1699  */
1700
1701 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1702 {
1703         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1704
1705         spin_lock(&gl->gl_lockref.lock);
1706         gl->gl_reply = ret;
1707
1708         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1709                 if (gfs2_should_freeze(gl)) {
1710                         set_bit(GLF_FROZEN, &gl->gl_flags);
1711                         spin_unlock(&gl->gl_lockref.lock);
1712                         return;
1713                 }
1714         }
1715
1716         gl->gl_lockref.count++;
1717         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1718         __gfs2_glock_queue_work(gl, 0);
1719         spin_unlock(&gl->gl_lockref.lock);
1720 }
1721
1722 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1723 {
1724         struct gfs2_glock *gla, *glb;
1725
1726         gla = list_entry(a, struct gfs2_glock, gl_lru);
1727         glb = list_entry(b, struct gfs2_glock, gl_lru);
1728
1729         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1730                 return 1;
1731         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1732                 return -1;
1733
1734         return 0;
1735 }
1736
1737 /**
1738  * gfs2_dispose_glock_lru - Demote a list of glocks
1739  * @list: The list to dispose of
1740  *
1741  * Disposing of glocks may involve disk accesses, so that here we sort
1742  * the glocks by number (i.e. disk location of the inodes) so that if
1743  * there are any such accesses, they'll be sent in order (mostly).
1744  *
1745  * Must be called under the lru_lock, but may drop and retake this
1746  * lock. While the lru_lock is dropped, entries may vanish from the
1747  * list, but no new entries will appear on the list (since it is
1748  * private)
1749  */
1750
1751 static void gfs2_dispose_glock_lru(struct list_head *list)
1752 __releases(&lru_lock)
1753 __acquires(&lru_lock)
1754 {
1755         struct gfs2_glock *gl;
1756
1757         list_sort(NULL, list, glock_cmp);
1758
1759         while(!list_empty(list)) {
1760                 gl = list_first_entry(list, struct gfs2_glock, gl_lru);
1761                 list_del_init(&gl->gl_lru);
1762                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1763 add_back_to_lru:
1764                         list_add(&gl->gl_lru, &lru_list);
1765                         set_bit(GLF_LRU, &gl->gl_flags);
1766                         atomic_inc(&lru_count);
1767                         continue;
1768                 }
1769                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1770                         spin_unlock(&gl->gl_lockref.lock);
1771                         goto add_back_to_lru;
1772                 }
1773                 gl->gl_lockref.count++;
1774                 if (demote_ok(gl))
1775                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1776                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1777                 __gfs2_glock_queue_work(gl, 0);
1778                 spin_unlock(&gl->gl_lockref.lock);
1779                 cond_resched_lock(&lru_lock);
1780         }
1781 }
1782
1783 /**
1784  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1785  * @nr: The number of entries to scan
1786  *
1787  * This function selects the entries on the LRU which are able to
1788  * be demoted, and then kicks off the process by calling
1789  * gfs2_dispose_glock_lru() above.
1790  */
1791
1792 static long gfs2_scan_glock_lru(int nr)
1793 {
1794         struct gfs2_glock *gl;
1795         LIST_HEAD(skipped);
1796         LIST_HEAD(dispose);
1797         long freed = 0;
1798
1799         spin_lock(&lru_lock);
1800         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1801                 gl = list_first_entry(&lru_list, struct gfs2_glock, gl_lru);
1802
1803                 /* Test for being demotable */
1804                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1805                         list_move(&gl->gl_lru, &dispose);
1806                         atomic_dec(&lru_count);
1807                         clear_bit(GLF_LRU, &gl->gl_flags);
1808                         freed++;
1809                         continue;
1810                 }
1811
1812                 list_move(&gl->gl_lru, &skipped);
1813         }
1814         list_splice(&skipped, &lru_list);
1815         if (!list_empty(&dispose))
1816                 gfs2_dispose_glock_lru(&dispose);
1817         spin_unlock(&lru_lock);
1818
1819         return freed;
1820 }
1821
1822 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1823                                             struct shrink_control *sc)
1824 {
1825         if (!(sc->gfp_mask & __GFP_FS))
1826                 return SHRINK_STOP;
1827         return gfs2_scan_glock_lru(sc->nr_to_scan);
1828 }
1829
1830 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1831                                              struct shrink_control *sc)
1832 {
1833         return vfs_pressure_ratio(atomic_read(&lru_count));
1834 }
1835
1836 static struct shrinker glock_shrinker = {
1837         .seeks = DEFAULT_SEEKS,
1838         .count_objects = gfs2_glock_shrink_count,
1839         .scan_objects = gfs2_glock_shrink_scan,
1840 };
1841
1842 /**
1843  * examine_bucket - Call a function for glock in a hash bucket
1844  * @examiner: the function
1845  * @sdp: the filesystem
1846  * @bucket: the bucket
1847  *
1848  * Note that the function can be called multiple times on the same
1849  * object.  So the user must ensure that the function can cope with
1850  * that.
1851  */
1852
1853 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1854 {
1855         struct gfs2_glock *gl;
1856         struct rhashtable_iter iter;
1857
1858         rhashtable_walk_enter(&gl_hash_table, &iter);
1859
1860         do {
1861                 rhashtable_walk_start(&iter);
1862
1863                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1864                         if (gl->gl_name.ln_sbd == sdp &&
1865                             lockref_get_not_dead(&gl->gl_lockref))
1866                                 examiner(gl);
1867
1868                 rhashtable_walk_stop(&iter);
1869         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1870
1871         rhashtable_walk_exit(&iter);
1872 }
1873
1874 bool gfs2_queue_delete_work(struct gfs2_glock *gl, unsigned long delay)
1875 {
1876         bool queued;
1877
1878         spin_lock(&gl->gl_lockref.lock);
1879         queued = queue_delayed_work(gfs2_delete_workqueue,
1880                                     &gl->gl_delete, delay);
1881         if (queued)
1882                 set_bit(GLF_PENDING_DELETE, &gl->gl_flags);
1883         spin_unlock(&gl->gl_lockref.lock);
1884         return queued;
1885 }
1886
1887 void gfs2_cancel_delete_work(struct gfs2_glock *gl)
1888 {
1889         if (cancel_delayed_work_sync(&gl->gl_delete)) {
1890                 clear_bit(GLF_PENDING_DELETE, &gl->gl_flags);
1891                 gfs2_glock_put(gl);
1892         }
1893 }
1894
1895 bool gfs2_delete_work_queued(const struct gfs2_glock *gl)
1896 {
1897         return test_bit(GLF_PENDING_DELETE, &gl->gl_flags);
1898 }
1899
1900 static void flush_delete_work(struct gfs2_glock *gl)
1901 {
1902         flush_delayed_work(&gl->gl_delete);
1903         gfs2_glock_queue_work(gl, 0);
1904 }
1905
1906 void gfs2_flush_delete_work(struct gfs2_sbd *sdp)
1907 {
1908         glock_hash_walk(flush_delete_work, sdp);
1909         flush_workqueue(gfs2_delete_workqueue);
1910 }
1911
1912 /**
1913  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1914  * @gl: The glock to thaw
1915  *
1916  */
1917
1918 static void thaw_glock(struct gfs2_glock *gl)
1919 {
1920         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1921                 gfs2_glock_put(gl);
1922                 return;
1923         }
1924         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1925         gfs2_glock_queue_work(gl, 0);
1926 }
1927
1928 /**
1929  * clear_glock - look at a glock and see if we can free it from glock cache
1930  * @gl: the glock to look at
1931  *
1932  */
1933
1934 static void clear_glock(struct gfs2_glock *gl)
1935 {
1936         gfs2_glock_remove_from_lru(gl);
1937
1938         spin_lock(&gl->gl_lockref.lock);
1939         if (gl->gl_state != LM_ST_UNLOCKED)
1940                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1941         __gfs2_glock_queue_work(gl, 0);
1942         spin_unlock(&gl->gl_lockref.lock);
1943 }
1944
1945 /**
1946  * gfs2_glock_thaw - Thaw any frozen glocks
1947  * @sdp: The super block
1948  *
1949  */
1950
1951 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1952 {
1953         glock_hash_walk(thaw_glock, sdp);
1954 }
1955
1956 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1957 {
1958         spin_lock(&gl->gl_lockref.lock);
1959         gfs2_dump_glock(seq, gl, fsid);
1960         spin_unlock(&gl->gl_lockref.lock);
1961 }
1962
1963 static void dump_glock_func(struct gfs2_glock *gl)
1964 {
1965         dump_glock(NULL, gl, true);
1966 }
1967
1968 /**
1969  * gfs2_gl_hash_clear - Empty out the glock hash table
1970  * @sdp: the filesystem
1971  * @wait: wait until it's all gone
1972  *
1973  * Called when unmounting the filesystem.
1974  */
1975
1976 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1977 {
1978         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1979         flush_workqueue(glock_workqueue);
1980         glock_hash_walk(clear_glock, sdp);
1981         flush_workqueue(glock_workqueue);
1982         wait_event_timeout(sdp->sd_glock_wait,
1983                            atomic_read(&sdp->sd_glock_disposal) == 0,
1984                            HZ * 600);
1985         glock_hash_walk(dump_glock_func, sdp);
1986 }
1987
1988 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1989 {
1990         struct gfs2_glock *gl = ip->i_gl;
1991         int ret;
1992
1993         ret = gfs2_truncatei_resume(ip);
1994         gfs2_glock_assert_withdraw(gl, ret == 0);
1995
1996         spin_lock(&gl->gl_lockref.lock);
1997         clear_bit(GLF_LOCK, &gl->gl_flags);
1998         run_queue(gl, 1);
1999         spin_unlock(&gl->gl_lockref.lock);
2000 }
2001
2002 static const char *state2str(unsigned state)
2003 {
2004         switch(state) {
2005         case LM_ST_UNLOCKED:
2006                 return "UN";
2007         case LM_ST_SHARED:
2008                 return "SH";
2009         case LM_ST_DEFERRED:
2010                 return "DF";
2011         case LM_ST_EXCLUSIVE:
2012                 return "EX";
2013         }
2014         return "??";
2015 }
2016
2017 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
2018 {
2019         char *p = buf;
2020         if (flags & LM_FLAG_TRY)
2021                 *p++ = 't';
2022         if (flags & LM_FLAG_TRY_1CB)
2023                 *p++ = 'T';
2024         if (flags & LM_FLAG_NOEXP)
2025                 *p++ = 'e';
2026         if (flags & LM_FLAG_ANY)
2027                 *p++ = 'A';
2028         if (flags & LM_FLAG_PRIORITY)
2029                 *p++ = 'p';
2030         if (flags & GL_ASYNC)
2031                 *p++ = 'a';
2032         if (flags & GL_EXACT)
2033                 *p++ = 'E';
2034         if (flags & GL_NOCACHE)
2035                 *p++ = 'c';
2036         if (test_bit(HIF_HOLDER, &iflags))
2037                 *p++ = 'H';
2038         if (test_bit(HIF_WAIT, &iflags))
2039                 *p++ = 'W';
2040         if (test_bit(HIF_FIRST, &iflags))
2041                 *p++ = 'F';
2042         *p = 0;
2043         return buf;
2044 }
2045
2046 /**
2047  * dump_holder - print information about a glock holder
2048  * @seq: the seq_file struct
2049  * @gh: the glock holder
2050  * @fs_id_buf: pointer to file system id (if requested)
2051  *
2052  */
2053
2054 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh,
2055                         const char *fs_id_buf)
2056 {
2057         struct task_struct *gh_owner = NULL;
2058         char flags_buf[32];
2059
2060         rcu_read_lock();
2061         if (gh->gh_owner_pid)
2062                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
2063         gfs2_print_dbg(seq, "%s H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
2064                        fs_id_buf, state2str(gh->gh_state),
2065                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
2066                        gh->gh_error,
2067                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
2068                        gh_owner ? gh_owner->comm : "(ended)",
2069                        (void *)gh->gh_ip);
2070         rcu_read_unlock();
2071 }
2072
2073 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
2074 {
2075         const unsigned long *gflags = &gl->gl_flags;
2076         char *p = buf;
2077
2078         if (test_bit(GLF_LOCK, gflags))
2079                 *p++ = 'l';
2080         if (test_bit(GLF_DEMOTE, gflags))
2081                 *p++ = 'D';
2082         if (test_bit(GLF_PENDING_DEMOTE, gflags))
2083                 *p++ = 'd';
2084         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
2085                 *p++ = 'p';
2086         if (test_bit(GLF_DIRTY, gflags))
2087                 *p++ = 'y';
2088         if (test_bit(GLF_LFLUSH, gflags))
2089                 *p++ = 'f';
2090         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
2091                 *p++ = 'i';
2092         if (test_bit(GLF_REPLY_PENDING, gflags))
2093                 *p++ = 'r';
2094         if (test_bit(GLF_INITIAL, gflags))
2095                 *p++ = 'I';
2096         if (test_bit(GLF_FROZEN, gflags))
2097                 *p++ = 'F';
2098         if (test_bit(GLF_QUEUED, gflags))
2099                 *p++ = 'q';
2100         if (test_bit(GLF_LRU, gflags))
2101                 *p++ = 'L';
2102         if (gl->gl_object)
2103                 *p++ = 'o';
2104         if (test_bit(GLF_BLOCKING, gflags))
2105                 *p++ = 'b';
2106         *p = 0;
2107         return buf;
2108 }
2109
2110 /**
2111  * gfs2_dump_glock - print information about a glock
2112  * @seq: The seq_file struct
2113  * @gl: the glock
2114  * @fsid: If true, also dump the file system id
2115  *
2116  * The file format is as follows:
2117  * One line per object, capital letters are used to indicate objects
2118  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
2119  * other objects are indented by a single space and follow the glock to
2120  * which they are related. Fields are indicated by lower case letters
2121  * followed by a colon and the field value, except for strings which are in
2122  * [] so that its possible to see if they are composed of spaces for
2123  * example. The field's are n = number (id of the object), f = flags,
2124  * t = type, s = state, r = refcount, e = error, p = pid.
2125  *
2126  */
2127
2128 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
2129 {
2130         const struct gfs2_glock_operations *glops = gl->gl_ops;
2131         unsigned long long dtime;
2132         const struct gfs2_holder *gh;
2133         char gflags_buf[32];
2134         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
2135         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
2136         unsigned long nrpages = 0;
2137
2138         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
2139                 struct address_space *mapping = gfs2_glock2aspace(gl);
2140
2141                 nrpages = mapping->nrpages;
2142         }
2143         memset(fs_id_buf, 0, sizeof(fs_id_buf));
2144         if (fsid && sdp) /* safety precaution */
2145                 sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
2146         dtime = jiffies - gl->gl_demote_time;
2147         dtime *= 1000000/HZ; /* demote time in uSec */
2148         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
2149                 dtime = 0;
2150         gfs2_print_dbg(seq, "%sG:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d "
2151                        "v:%d r:%d m:%ld p:%lu\n",
2152                        fs_id_buf, state2str(gl->gl_state),
2153                        gl->gl_name.ln_type,
2154                        (unsigned long long)gl->gl_name.ln_number,
2155                        gflags2str(gflags_buf, gl),
2156                        state2str(gl->gl_target),
2157                        state2str(gl->gl_demote_state), dtime,
2158                        atomic_read(&gl->gl_ail_count),
2159                        atomic_read(&gl->gl_revokes),
2160                        (int)gl->gl_lockref.count, gl->gl_hold_time, nrpages);
2161
2162         list_for_each_entry(gh, &gl->gl_holders, gh_list)
2163                 dump_holder(seq, gh, fs_id_buf);
2164
2165         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
2166                 glops->go_dump(seq, gl, fs_id_buf);
2167 }
2168
2169 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
2170 {
2171         struct gfs2_glock *gl = iter_ptr;
2172
2173         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
2174                    gl->gl_name.ln_type,
2175                    (unsigned long long)gl->gl_name.ln_number,
2176                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
2177                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
2178                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
2179                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
2180                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
2181                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
2182                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
2183                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
2184         return 0;
2185 }
2186
2187 static const char *gfs2_gltype[] = {
2188         "type",
2189         "reserved",
2190         "nondisk",
2191         "inode",
2192         "rgrp",
2193         "meta",
2194         "iopen",
2195         "flock",
2196         "plock",
2197         "quota",
2198         "journal",
2199 };
2200
2201 static const char *gfs2_stype[] = {
2202         [GFS2_LKS_SRTT]         = "srtt",
2203         [GFS2_LKS_SRTTVAR]      = "srttvar",
2204         [GFS2_LKS_SRTTB]        = "srttb",
2205         [GFS2_LKS_SRTTVARB]     = "srttvarb",
2206         [GFS2_LKS_SIRT]         = "sirt",
2207         [GFS2_LKS_SIRTVAR]      = "sirtvar",
2208         [GFS2_LKS_DCOUNT]       = "dlm",
2209         [GFS2_LKS_QCOUNT]       = "queue",
2210 };
2211
2212 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
2213
2214 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
2215 {
2216         struct gfs2_sbd *sdp = seq->private;
2217         loff_t pos = *(loff_t *)iter_ptr;
2218         unsigned index = pos >> 3;
2219         unsigned subindex = pos & 0x07;
2220         int i;
2221
2222         if (index == 0 && subindex != 0)
2223                 return 0;
2224
2225         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
2226                    (index == 0) ? "cpu": gfs2_stype[subindex]);
2227
2228         for_each_possible_cpu(i) {
2229                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
2230
2231                 if (index == 0)
2232                         seq_printf(seq, " %15u", i);
2233                 else
2234                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
2235                                    lkstats[index - 1].stats[subindex]);
2236         }
2237         seq_putc(seq, '\n');
2238         return 0;
2239 }
2240
2241 int __init gfs2_glock_init(void)
2242 {
2243         int i, ret;
2244
2245         ret = rhashtable_init(&gl_hash_table, &ht_parms);
2246         if (ret < 0)
2247                 return ret;
2248
2249         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
2250                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
2251         if (!glock_workqueue) {
2252                 rhashtable_destroy(&gl_hash_table);
2253                 return -ENOMEM;
2254         }
2255         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
2256                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
2257                                                 0);
2258         if (!gfs2_delete_workqueue) {
2259                 destroy_workqueue(glock_workqueue);
2260                 rhashtable_destroy(&gl_hash_table);
2261                 return -ENOMEM;
2262         }
2263
2264         ret = register_shrinker(&glock_shrinker);
2265         if (ret) {
2266                 destroy_workqueue(gfs2_delete_workqueue);
2267                 destroy_workqueue(glock_workqueue);
2268                 rhashtable_destroy(&gl_hash_table);
2269                 return ret;
2270         }
2271
2272         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
2273                 init_waitqueue_head(glock_wait_table + i);
2274
2275         return 0;
2276 }
2277
2278 void gfs2_glock_exit(void)
2279 {
2280         unregister_shrinker(&glock_shrinker);
2281         rhashtable_destroy(&gl_hash_table);
2282         destroy_workqueue(glock_workqueue);
2283         destroy_workqueue(gfs2_delete_workqueue);
2284 }
2285
2286 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
2287 {
2288         struct gfs2_glock *gl = gi->gl;
2289
2290         if (gl) {
2291                 if (n == 0)
2292                         return;
2293                 if (!lockref_put_not_zero(&gl->gl_lockref))
2294                         gfs2_glock_queue_put(gl);
2295         }
2296         for (;;) {
2297                 gl = rhashtable_walk_next(&gi->hti);
2298                 if (IS_ERR_OR_NULL(gl)) {
2299                         if (gl == ERR_PTR(-EAGAIN)) {
2300                                 n = 1;
2301                                 continue;
2302                         }
2303                         gl = NULL;
2304                         break;
2305                 }
2306                 if (gl->gl_name.ln_sbd != gi->sdp)
2307                         continue;
2308                 if (n <= 1) {
2309                         if (!lockref_get_not_dead(&gl->gl_lockref))
2310                                 continue;
2311                         break;
2312                 } else {
2313                         if (__lockref_is_dead(&gl->gl_lockref))
2314                                 continue;
2315                         n--;
2316                 }
2317         }
2318         gi->gl = gl;
2319 }
2320
2321 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
2322         __acquires(RCU)
2323 {
2324         struct gfs2_glock_iter *gi = seq->private;
2325         loff_t n;
2326
2327         /*
2328          * We can either stay where we are, skip to the next hash table
2329          * entry, or start from the beginning.
2330          */
2331         if (*pos < gi->last_pos) {
2332                 rhashtable_walk_exit(&gi->hti);
2333                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2334                 n = *pos + 1;
2335         } else {
2336                 n = *pos - gi->last_pos;
2337         }
2338
2339         rhashtable_walk_start(&gi->hti);
2340
2341         gfs2_glock_iter_next(gi, n);
2342         gi->last_pos = *pos;
2343         return gi->gl;
2344 }
2345
2346 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
2347                                  loff_t *pos)
2348 {
2349         struct gfs2_glock_iter *gi = seq->private;
2350
2351         (*pos)++;
2352         gi->last_pos = *pos;
2353         gfs2_glock_iter_next(gi, 1);
2354         return gi->gl;
2355 }
2356
2357 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2358         __releases(RCU)
2359 {
2360         struct gfs2_glock_iter *gi = seq->private;
2361
2362         rhashtable_walk_stop(&gi->hti);
2363 }
2364
2365 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2366 {
2367         dump_glock(seq, iter_ptr, false);
2368         return 0;
2369 }
2370
2371 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2372 {
2373         preempt_disable();
2374         if (*pos >= GFS2_NR_SBSTATS)
2375                 return NULL;
2376         return pos;
2377 }
2378
2379 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2380                                    loff_t *pos)
2381 {
2382         (*pos)++;
2383         if (*pos >= GFS2_NR_SBSTATS)
2384                 return NULL;
2385         return pos;
2386 }
2387
2388 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2389 {
2390         preempt_enable();
2391 }
2392
2393 static const struct seq_operations gfs2_glock_seq_ops = {
2394         .start = gfs2_glock_seq_start,
2395         .next  = gfs2_glock_seq_next,
2396         .stop  = gfs2_glock_seq_stop,
2397         .show  = gfs2_glock_seq_show,
2398 };
2399
2400 static const struct seq_operations gfs2_glstats_seq_ops = {
2401         .start = gfs2_glock_seq_start,
2402         .next  = gfs2_glock_seq_next,
2403         .stop  = gfs2_glock_seq_stop,
2404         .show  = gfs2_glstats_seq_show,
2405 };
2406
2407 static const struct seq_operations gfs2_sbstats_seq_ops = {
2408         .start = gfs2_sbstats_seq_start,
2409         .next  = gfs2_sbstats_seq_next,
2410         .stop  = gfs2_sbstats_seq_stop,
2411         .show  = gfs2_sbstats_seq_show,
2412 };
2413
2414 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2415
2416 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2417                               const struct seq_operations *ops)
2418 {
2419         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2420         if (ret == 0) {
2421                 struct seq_file *seq = file->private_data;
2422                 struct gfs2_glock_iter *gi = seq->private;
2423
2424                 gi->sdp = inode->i_private;
2425                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2426                 if (seq->buf)
2427                         seq->size = GFS2_SEQ_GOODSIZE;
2428                 /*
2429                  * Initially, we are "before" the first hash table entry; the
2430                  * first call to rhashtable_walk_next gets us the first entry.
2431                  */
2432                 gi->last_pos = -1;
2433                 gi->gl = NULL;
2434                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2435         }
2436         return ret;
2437 }
2438
2439 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2440 {
2441         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2442 }
2443
2444 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2445 {
2446         struct seq_file *seq = file->private_data;
2447         struct gfs2_glock_iter *gi = seq->private;
2448
2449         if (gi->gl)
2450                 gfs2_glock_put(gi->gl);
2451         rhashtable_walk_exit(&gi->hti);
2452         return seq_release_private(inode, file);
2453 }
2454
2455 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2456 {
2457         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2458 }
2459
2460 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2461 {
2462         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2463         if (ret == 0) {
2464                 struct seq_file *seq = file->private_data;
2465                 seq->private = inode->i_private;  /* sdp */
2466         }
2467         return ret;
2468 }
2469
2470 static const struct file_operations gfs2_glocks_fops = {
2471         .owner   = THIS_MODULE,
2472         .open    = gfs2_glocks_open,
2473         .read    = seq_read,
2474         .llseek  = seq_lseek,
2475         .release = gfs2_glocks_release,
2476 };
2477
2478 static const struct file_operations gfs2_glstats_fops = {
2479         .owner   = THIS_MODULE,
2480         .open    = gfs2_glstats_open,
2481         .read    = seq_read,
2482         .llseek  = seq_lseek,
2483         .release = gfs2_glocks_release,
2484 };
2485
2486 static const struct file_operations gfs2_sbstats_fops = {
2487         .owner   = THIS_MODULE,
2488         .open    = gfs2_sbstats_open,
2489         .read    = seq_read,
2490         .llseek  = seq_lseek,
2491         .release = seq_release,
2492 };
2493
2494 void gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2495 {
2496         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2497
2498         debugfs_create_file("glocks", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2499                             &gfs2_glocks_fops);
2500
2501         debugfs_create_file("glstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2502                             &gfs2_glstats_fops);
2503
2504         debugfs_create_file("sbstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2505                             &gfs2_sbstats_fops);
2506 }
2507
2508 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2509 {
2510         debugfs_remove_recursive(sdp->debugfs_dir);
2511         sdp->debugfs_dir = NULL;
2512 }
2513
2514 void gfs2_register_debugfs(void)
2515 {
2516         gfs2_root = debugfs_create_dir("gfs2", NULL);
2517 }
2518
2519 void gfs2_unregister_debugfs(void)
2520 {
2521         debugfs_remove(gfs2_root);
2522         gfs2_root = NULL;
2523 }