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         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
 129                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 130         } else {
 131                 kfree(gl->gl_lksb.sb_lvbptr);
 132                 kmem_cache_free(gfs2_glock_cachep, gl);
 133         }
 134 }
 135
 136 void gfs2_glock_free(struct gfs2_glock *gl)
 137 {
 138         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 139
 140         BUG_ON(atomic_read(&gl->gl_revokes));
 141         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 142         smp_mb();
 143         wake_up_glock(gl);
 144         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 145         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 146                 wake_up(&sdp->sd_glock_wait);
 147 }
 148
 149 /**
 150  * gfs2_glock_hold() - increment reference count on glock
 151  * @gl: The glock to hold
 152  *
 153  */
 154
 155 void gfs2_glock_hold(struct gfs2_glock *gl)
 156 {
 157         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 158         lockref_get(&gl->gl_lockref);
 159 }
 160
 161 /**
 162  * demote_ok - Check to see if it's ok to unlock a glock
 163  * @gl: the glock
 164  *
 165  * Returns: 1 if it's ok
 166  */
 167
 168 static int demote_ok(const struct gfs2_glock *gl)
 169 {
 170         const struct gfs2_glock_operations *glops = gl->gl_ops;
 171
 172         if (gl->gl_state == LM_ST_UNLOCKED)
 173                 return 0;
 174         if (!list_empty(&gl->gl_holders))
 175                 return 0;
 176         if (glops->go_demote_ok)
 177                 return glops->go_demote_ok(gl);
 178         return 1;
 179 }
 180
 181
 182 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 183 {
 184         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 185                 return;
 186
 187         spin_lock(&lru_lock);
 188
 189         list_del(&gl->gl_lru);
 190         list_add_tail(&gl->gl_lru, &lru_list);
 191
 192         if (!test_bit(GLF_LRU, &gl->gl_flags)) {
 193                 set_bit(GLF_LRU, &gl->gl_flags);
 194                 atomic_inc(&lru_count);
 195         }
 196
 197         spin_unlock(&lru_lock);
 198 }
 199
 200 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 201 {
 202         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 203                 return;
 204
 205         spin_lock(&lru_lock);
 206         if (test_bit(GLF_LRU, &gl->gl_flags)) {
 207                 list_del_init(&gl->gl_lru);
 208                 atomic_dec(&lru_count);
 209                 clear_bit(GLF_LRU, &gl->gl_flags);
 210         }
 211         spin_unlock(&lru_lock);
 212 }
 213
 214 /*
 215  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 216  * work queue.
 217  */
 218 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 219         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 220                 /*
 221                  * We are holding the lockref spinlock, and the work was still
 222                  * queued above.  The queued work (glock_work_func) takes that
 223                  * spinlock before dropping its glock reference(s), so it
 224                  * cannot have dropped them in the meantime.
 225                  */
 226                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 227                 gl->gl_lockref.count--;
 228         }
 229 }
 230
 231 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 232         spin_lock(&gl->gl_lockref.lock);
 233         __gfs2_glock_queue_work(gl, delay);
 234         spin_unlock(&gl->gl_lockref.lock);
 235 }
 236
 237 static void __gfs2_glock_put(struct gfs2_glock *gl)
 238 {
 239         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 240         struct address_space *mapping = gfs2_glock2aspace(gl);
 241
 242         lockref_mark_dead(&gl->gl_lockref);
 243
 244         gfs2_glock_remove_from_lru(gl);
 245         spin_unlock(&gl->gl_lockref.lock);
 246         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 247         GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
 248         trace_gfs2_glock_put(gl);
 249         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 250 }
 251
 252 /*
 253  * Cause the glock to be put in work queue context.
 254  */
 255 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 256 {
 257         gfs2_glock_queue_work(gl, 0);
 258 }
 259
 260 /**
 261  * gfs2_glock_put() - Decrement reference count on glock
 262  * @gl: The glock to put
 263  *
 264  */
 265
 266 void gfs2_glock_put(struct gfs2_glock *gl)
 267 {
 268         if (lockref_put_or_lock(&gl->gl_lockref))
 269                 return;
 270
 271         __gfs2_glock_put(gl);
 272 }
 273
 274 /**
 275  * may_grant - check if its ok to grant a new lock
 276  * @gl: The glock
 277  * @gh: The lock request which we wish to grant
 278  *
 279  * Returns: true if its ok to grant the lock
 280  */
 281
 282 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 283 {
 284         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 285         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 286              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 287                 return 0;
 288         if (gl->gl_state == gh->gh_state)
 289                 return 1;
 290         if (gh->gh_flags & GL_EXACT)
 291                 return 0;
 292         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 293                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 294                         return 1;
 295                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 296                         return 1;
 297         }
 298         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 299                 return 1;
 300         return 0;
 301 }
 302
 303 static void gfs2_holder_wake(struct gfs2_holder *gh)
 304 {
 305         clear_bit(HIF_WAIT, &gh->gh_iflags);
 306         smp_mb__after_atomic();
 307         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 308 }
 309
 310 /**
 311  * do_error - Something unexpected has happened during a lock request
 312  *
 313  */
 314
 315 static void do_error(struct gfs2_glock *gl, const int ret)
 316 {
 317         struct gfs2_holder *gh, *tmp;
 318
 319         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 320                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 321                         continue;
 322                 if (ret & LM_OUT_ERROR)
 323                         gh->gh_error = -EIO;
 324                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 325                         gh->gh_error = GLR_TRYFAILED;
 326                 else
 327                         continue;
 328                 list_del_init(&gh->gh_list);
 329                 trace_gfs2_glock_queue(gh, 0);
 330                 gfs2_holder_wake(gh);
 331         }
 332 }
 333
 334 /**
 335  * do_promote - promote as many requests as possible on the current queue
 336  * @gl: The glock
 337  *
 338  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 339  *          if a type specific operation is underway.
 340  */
 341
 342 static int do_promote(struct gfs2_glock *gl)
 343 __releases(&gl->gl_lockref.lock)
 344 __acquires(&gl->gl_lockref.lock)
 345 {
 346         const struct gfs2_glock_operations *glops = gl->gl_ops;
 347         struct gfs2_holder *gh, *tmp;
 348         int ret;
 349
 350 restart:
 351         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 352                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 353                         continue;
 354                 if (may_grant(gl, gh)) {
 355                         if (gh->gh_list.prev == &gl->gl_holders &&
 356                             glops->go_lock) {
 357                                 spin_unlock(&gl->gl_lockref.lock);
 358                                 /* FIXME: eliminate this eventually */
 359                                 ret = glops->go_lock(gh);
 360                                 spin_lock(&gl->gl_lockref.lock);
 361                                 if (ret) {
 362                                         if (ret == 1)
 363                                                 return 2;
 364                                         gh->gh_error = ret;
 365                                         list_del_init(&gh->gh_list);
 366                                         trace_gfs2_glock_queue(gh, 0);
 367                                         gfs2_holder_wake(gh);
 368                                         goto restart;
 369                                 }
 370                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 371                                 trace_gfs2_promote(gh, 1);
 372                                 gfs2_holder_wake(gh);
 373                                 goto restart;
 374                         }
 375                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 376                         trace_gfs2_promote(gh, 0);
 377                         gfs2_holder_wake(gh);
 378                         continue;
 379                 }
 380                 if (gh->gh_list.prev == &gl->gl_holders)
 381                         return 1;
 382                 do_error(gl, 0);
 383                 break;
 384         }
 385         return 0;
 386 }
 387
 388 /**
 389  * find_first_waiter - find the first gh that's waiting for the glock
 390  * @gl: the glock
 391  */
 392
 393 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 394 {
 395         struct gfs2_holder *gh;
 396
 397         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 398                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 399                         return gh;
 400         }
 401         return NULL;
 402 }
 403
 404 /**
 405  * state_change - record that the glock is now in a different state
 406  * @gl: the glock
 407  * @new_state the new state
 408  *
 409  */
 410
 411 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 412 {
 413         int held1, held2;
 414
 415         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 416         held2 = (new_state != LM_ST_UNLOCKED);
 417
 418         if (held1 != held2) {
 419                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 420                 if (held2)
 421                         gl->gl_lockref.count++;
 422                 else
 423                         gl->gl_lockref.count--;
 424         }
 425         if (held1 && held2 && list_empty(&gl->gl_holders))
 426                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 427
 428         if (new_state != gl->gl_target)
 429                 /* shorten our minimum hold time */
 430                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 431                                        GL_GLOCK_MIN_HOLD);
 432         gl->gl_state = new_state;
 433         gl->gl_tchange = jiffies;
 434 }
 435
 436 static void gfs2_demote_wake(struct gfs2_glock *gl)
 437 {
 438         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 439         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 440         smp_mb__after_atomic();
 441         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 442 }
 443
 444 /**
 445  * finish_xmote - The DLM has replied to one of our lock requests
 446  * @gl: The glock
 447  * @ret: The status from the DLM
 448  *
 449  */
 450
 451 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 452 {
 453         const struct gfs2_glock_operations *glops = gl->gl_ops;
 454         struct gfs2_holder *gh;
 455         unsigned state = ret & LM_OUT_ST_MASK;
 456         int rv;
 457
 458         spin_lock(&gl->gl_lockref.lock);
 459         trace_gfs2_glock_state_change(gl, state);
 460         state_change(gl, state);
 461         gh = find_first_waiter(gl);
 462
 463         /* Demote to UN request arrived during demote to SH or DF */
 464         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 465             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 466                 gl->gl_target = LM_ST_UNLOCKED;
 467
 468         /* Check for state != intended state */
 469         if (unlikely(state != gl->gl_target)) {
 470                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 471                         /* move to back of queue and try next entry */
 472                         if (ret & LM_OUT_CANCELED) {
 473                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 474                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 475                                 gh = find_first_waiter(gl);
 476                                 gl->gl_target = gh->gh_state;
 477                                 goto retry;
 478                         }
 479                         /* Some error or failed "try lock" - report it */
 480                         if ((ret & LM_OUT_ERROR) ||
 481                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 482                                 gl->gl_target = gl->gl_state;
 483                                 do_error(gl, ret);
 484                                 goto out;
 485                         }
 486                 }
 487                 switch(state) {
 488                 /* Unlocked due to conversion deadlock, try again */
 489                 case LM_ST_UNLOCKED:
 490 retry:
 491                         do_xmote(gl, gh, gl->gl_target);
 492                         break;
 493                 /* Conversion fails, unlock and try again */
 494                 case LM_ST_SHARED:
 495                 case LM_ST_DEFERRED:
 496                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 497                         break;
 498                 default: /* Everything else */
 499                         fs_err(gl->gl_name.ln_sbd, "wanted %u got %u\n",
 500                                gl->gl_target, state);
 501                         GLOCK_BUG_ON(gl, 1);
 502                 }
 503                 spin_unlock(&gl->gl_lockref.lock);
 504                 return;
 505         }
 506
 507         /* Fast path - we got what we asked for */
 508         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 509                 gfs2_demote_wake(gl);
 510         if (state != LM_ST_UNLOCKED) {
 511                 if (glops->go_xmote_bh) {
 512                         spin_unlock(&gl->gl_lockref.lock);
 513                         rv = glops->go_xmote_bh(gl, gh);
 514                         spin_lock(&gl->gl_lockref.lock);
 515                         if (rv) {
 516                                 do_error(gl, rv);
 517                                 goto out;
 518                         }
 519                 }
 520                 rv = do_promote(gl);
 521                 if (rv == 2)
 522                         goto out_locked;
 523         }
 524 out:
 525         clear_bit(GLF_LOCK, &gl->gl_flags);
 526 out_locked:
 527         spin_unlock(&gl->gl_lockref.lock);
 528 }
 529
 530 /**
 531  * do_xmote - Calls the DLM to change the state of a lock
 532  * @gl: The lock state
 533  * @gh: The holder (only for promotes)
 534  * @target: The target lock state
 535  *
 536  */
 537
 538 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 539 __releases(&gl->gl_lockref.lock)
 540 __acquires(&gl->gl_lockref.lock)
 541 {
 542         const struct gfs2_glock_operations *glops = gl->gl_ops;
 543         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 544         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 545         int ret;
 546
 547         if (unlikely(test_bit(SDF_WITHDRAWN, &sdp->sd_flags)) &&
 548             target != LM_ST_UNLOCKED)
 549                 return;
 550         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 551                       LM_FLAG_PRIORITY);
 552         GLOCK_BUG_ON(gl, gl->gl_state == target);
 553         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 554         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 555             glops->go_inval) {
 556                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 557                 do_error(gl, 0); /* Fail queued try locks */
 558         }
 559         gl->gl_req = target;
 560         set_bit(GLF_BLOCKING, &gl->gl_flags);
 561         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 562             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 563             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 564                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 565         spin_unlock(&gl->gl_lockref.lock);
 566         if (glops->go_sync)
 567                 glops->go_sync(gl);
 568         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 569                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 570         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 571
 572         gfs2_glock_hold(gl);
 573         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 574                 /* lock_dlm */
 575                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 576                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 577                     target == LM_ST_UNLOCKED &&
 578                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 579                         finish_xmote(gl, target);
 580                         gfs2_glock_queue_work(gl, 0);
 581                 }
 582                 else if (ret) {
 583                         fs_err(sdp, "lm_lock ret %d\n", ret);
 584                         GLOCK_BUG_ON(gl, !test_bit(SDF_WITHDRAWN,
 585                                                    &sdp->sd_flags));
 586                 }
 587         } else { /* lock_nolock */
 588                 finish_xmote(gl, target);
 589                 gfs2_glock_queue_work(gl, 0);
 590         }
 591
 592         spin_lock(&gl->gl_lockref.lock);
 593 }
 594
 595 /**
 596  * find_first_holder - find the first "holder" gh
 597  * @gl: the glock
 598  */
 599
 600 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 601 {
 602         struct gfs2_holder *gh;
 603
 604         if (!list_empty(&gl->gl_holders)) {
 605                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 606                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 607                         return gh;
 608         }
 609         return NULL;
 610 }
 611
 612 /**
 613  * run_queue - do all outstanding tasks related to a glock
 614  * @gl: The glock in question
 615  * @nonblock: True if we must not block in run_queue
 616  *
 617  */
 618
 619 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 620 __releases(&gl->gl_lockref.lock)
 621 __acquires(&gl->gl_lockref.lock)
 622 {
 623         struct gfs2_holder *gh = NULL;
 624         int ret;
 625
 626         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 627                 return;
 628
 629         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 630
 631         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 632             gl->gl_demote_state != gl->gl_state) {
 633                 if (find_first_holder(gl))
 634                         goto out_unlock;
 635                 if (nonblock)
 636                         goto out_sched;
 637                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 638                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 639                 gl->gl_target = gl->gl_demote_state;
 640         } else {
 641                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 642                         gfs2_demote_wake(gl);
 643                 ret = do_promote(gl);
 644                 if (ret == 0)
 645                         goto out_unlock;
 646                 if (ret == 2)
 647                         goto out;
 648                 gh = find_first_waiter(gl);
 649                 gl->gl_target = gh->gh_state;
 650                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 651                         do_error(gl, 0); /* Fail queued try locks */
 652         }
 653         do_xmote(gl, gh, gl->gl_target);
 654 out:
 655         return;
 656
 657 out_sched:
 658         clear_bit(GLF_LOCK, &gl->gl_flags);
 659         smp_mb__after_atomic();
 660         gl->gl_lockref.count++;
 661         __gfs2_glock_queue_work(gl, 0);
 662         return;
 663
 664 out_unlock:
 665         clear_bit(GLF_LOCK, &gl->gl_flags);
 666         smp_mb__after_atomic();
 667         return;
 668 }
 669
 670 static void delete_work_func(struct work_struct *work)
 671 {
 672         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 673         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 674         struct inode *inode;
 675         u64 no_addr = gl->gl_name.ln_number;
 676
 677         /* If someone's using this glock to create a new dinode, the block must
 678            have been freed by another node, then re-used, in which case our
 679            iopen callback is too late after the fact. Ignore it. */
 680         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 681                 goto out;
 682
 683         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 684         if (!IS_ERR_OR_NULL(inode)) {
 685                 d_prune_aliases(inode);
 686                 iput(inode);
 687         }
 688 out:
 689         gfs2_glock_put(gl);
 690 }
 691
 692 static void glock_work_func(struct work_struct *work)
 693 {
 694         unsigned long delay = 0;
 695         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 696         unsigned int drop_refs = 1;
 697
 698         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 699                 finish_xmote(gl, gl->gl_reply);
 700                 drop_refs++;
 701         }
 702         spin_lock(&gl->gl_lockref.lock);
 703         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 704             gl->gl_state != LM_ST_UNLOCKED &&
 705             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 706                 unsigned long holdtime, now = jiffies;
 707
 708                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 709                 if (time_before(now, holdtime))
 710                         delay = holdtime - now;
 711
 712                 if (!delay) {
 713                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 714                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 715                 }
 716         }
 717         run_queue(gl, 0);
 718         if (delay) {
 719                 /* Keep one glock reference for the work we requeue. */
 720                 drop_refs--;
 721                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 722                         delay = 0;
 723                 __gfs2_glock_queue_work(gl, delay);
 724         }
 725
 726         /*
 727          * Drop the remaining glock references manually here. (Mind that
 728          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 729          * here as well.)
 730          */
 731         gl->gl_lockref.count -= drop_refs;
 732         if (!gl->gl_lockref.count) {
 733                 __gfs2_glock_put(gl);
 734                 return;
 735         }
 736         spin_unlock(&gl->gl_lockref.lock);
 737 }
 738
 739 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 740                                             struct gfs2_glock *new)
 741 {
 742         struct wait_glock_queue wait;
 743         wait_queue_head_t *wq = glock_waitqueue(name);
 744         struct gfs2_glock *gl;
 745
 746         wait.name = name;
 747         init_wait(&wait.wait);
 748         wait.wait.func = glock_wake_function;
 749
 750 again:
 751         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 752         rcu_read_lock();
 753         if (new) {
 754                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 755                         &new->gl_node, ht_parms);
 756                 if (IS_ERR(gl))
 757                         goto out;
 758         } else {
 759                 gl = rhashtable_lookup_fast(&gl_hash_table,
 760                         name, ht_parms);
 761         }
 762         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 763                 rcu_read_unlock();
 764                 schedule();
 765                 goto again;
 766         }
 767 out:
 768         rcu_read_unlock();
 769         finish_wait(wq, &wait.wait);
 770         return gl;
 771 }
 772
 773 /**
 774  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 775  * @sdp: The GFS2 superblock
 776  * @number: the lock number
 777  * @glops: The glock_operations to use
 778  * @create: If 0, don't create the glock if it doesn't exist
 779  * @glp: the glock is returned here
 780  *
 781  * This does not lock a glock, just finds/creates structures for one.
 782  *
 783  * Returns: errno
 784  */
 785
 786 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 787                    const struct gfs2_glock_operations *glops, int create,
 788                    struct gfs2_glock **glp)
 789 {
 790         struct super_block *s = sdp->sd_vfs;
 791         struct lm_lockname name = { .ln_number = number,
 792                                     .ln_type = glops->go_type,
 793                                     .ln_sbd = sdp };
 794         struct gfs2_glock *gl, *tmp;
 795         struct address_space *mapping;
 796         struct kmem_cache *cachep;
 797         int ret = 0;
 798
 799         gl = find_insert_glock(&name, NULL);
 800         if (gl) {
 801                 *glp = gl;
 802                 return 0;
 803         }
 804         if (!create)
 805                 return -ENOENT;
 806
 807         if (glops->go_flags & GLOF_ASPACE)
 808                 cachep = gfs2_glock_aspace_cachep;
 809         else
 810                 cachep = gfs2_glock_cachep;
 811         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 812         if (!gl)
 813                 return -ENOMEM;
 814
 815         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 816
 817         if (glops->go_flags & GLOF_LVB) {
 818                 gl->gl_lksb.sb_lvbptr = kzalloc(GFS2_MIN_LVB_SIZE, GFP_NOFS);
 819                 if (!gl->gl_lksb.sb_lvbptr) {
 820                         kmem_cache_free(cachep, gl);
 821                         return -ENOMEM;
 822                 }
 823         }
 824
 825         atomic_inc(&sdp->sd_glock_disposal);
 826         gl->gl_node.next = NULL;
 827         gl->gl_flags = 0;
 828         gl->gl_name = name;
 829         gl->gl_lockref.count = 1;
 830         gl->gl_state = LM_ST_UNLOCKED;
 831         gl->gl_target = LM_ST_UNLOCKED;
 832         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 833         gl->gl_ops = glops;
 834         gl->gl_dstamp = 0;
 835         preempt_disable();
 836         /* We use the global stats to estimate the initial per-glock stats */
 837         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 838         preempt_enable();
 839         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 840         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 841         gl->gl_tchange = jiffies;
 842         gl->gl_object = NULL;
 843         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 844         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 845         INIT_WORK(&gl->gl_delete, delete_work_func);
 846
 847         mapping = gfs2_glock2aspace(gl);
 848         if (mapping) {
 849                 mapping->a_ops = &gfs2_meta_aops;
 850                 mapping->host = s->s_bdev->bd_inode;
 851                 mapping->flags = 0;
 852                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 853                 mapping->private_data = NULL;
 854                 mapping->writeback_index = 0;
 855         }
 856
 857         tmp = find_insert_glock(&name, gl);
 858         if (!tmp) {
 859                 *glp = gl;
 860                 goto out;
 861         }
 862         if (IS_ERR(tmp)) {
 863                 ret = PTR_ERR(tmp);
 864                 goto out_free;
 865         }
 866         *glp = tmp;
 867
 868 out_free:
 869         kfree(gl->gl_lksb.sb_lvbptr);
 870         kmem_cache_free(cachep, gl);
 871         atomic_dec(&sdp->sd_glock_disposal);
 872
 873 out:
 874         return ret;
 875 }
 876
 877 /**
 878  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 879  * @gl: the glock
 880  * @state: the state we're requesting
 881  * @flags: the modifier flags
 882  * @gh: the holder structure
 883  *
 884  */
 885
 886 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 887                       struct gfs2_holder *gh)
 888 {
 889         INIT_LIST_HEAD(&gh->gh_list);
 890         gh->gh_gl = gl;
 891         gh->gh_ip = _RET_IP_;
 892         gh->gh_owner_pid = get_pid(task_pid(current));
 893         gh->gh_state = state;
 894         gh->gh_flags = flags;
 895         gh->gh_error = 0;
 896         gh->gh_iflags = 0;
 897         gfs2_glock_hold(gl);
 898 }
 899
 900 /**
 901  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 902  * @state: the state we're requesting
 903  * @flags: the modifier flags
 904  * @gh: the holder structure
 905  *
 906  * Don't mess with the glock.
 907  *
 908  */
 909
 910 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
 911 {
 912         gh->gh_state = state;
 913         gh->gh_flags = flags;
 914         gh->gh_iflags = 0;
 915         gh->gh_ip = _RET_IP_;
 916         put_pid(gh->gh_owner_pid);
 917         gh->gh_owner_pid = get_pid(task_pid(current));
 918 }
 919
 920 /**
 921  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 922  * @gh: the holder structure
 923  *
 924  */
 925
 926 void gfs2_holder_uninit(struct gfs2_holder *gh)
 927 {
 928         put_pid(gh->gh_owner_pid);
 929         gfs2_glock_put(gh->gh_gl);
 930         gfs2_holder_mark_uninitialized(gh);
 931         gh->gh_ip = 0;
 932 }
 933
 934 /**
 935  * gfs2_glock_wait - wait on a glock acquisition
 936  * @gh: the glock holder
 937  *
 938  * Returns: 0 on success
 939  */
 940
 941 int gfs2_glock_wait(struct gfs2_holder *gh)
 942 {
 943         unsigned long time1 = jiffies;
 944
 945         might_sleep();
 946         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
 947         if (time_after(jiffies, time1 + HZ)) /* have we waited > a second? */
 948                 /* Lengthen the minimum hold time. */
 949                 gh->gh_gl->gl_hold_time = min(gh->gh_gl->gl_hold_time +
 950                                               GL_GLOCK_HOLD_INCR,
 951                                               GL_GLOCK_MAX_HOLD);
 952         return gh->gh_error;
 953 }
 954
 955 /**
 956  * handle_callback - process a demote request
 957  * @gl: the glock
 958  * @state: the state the caller wants us to change to
 959  *
 960  * There are only two requests that we are going to see in actual
 961  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 962  */
 963
 964 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 965                             unsigned long delay, bool remote)
 966 {
 967         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 968
 969         set_bit(bit, &gl->gl_flags);
 970         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 971                 gl->gl_demote_state = state;
 972                 gl->gl_demote_time = jiffies;
 973         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 974                         gl->gl_demote_state != state) {
 975                 gl->gl_demote_state = LM_ST_UNLOCKED;
 976         }
 977         if (gl->gl_ops->go_callback)
 978                 gl->gl_ops->go_callback(gl, remote);
 979         trace_gfs2_demote_rq(gl, remote);
 980 }
 981
 982 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 983 {
 984         struct va_format vaf;
 985         va_list args;
 986
 987         va_start(args, fmt);
 988
 989         if (seq) {
 990                 seq_vprintf(seq, fmt, args);
 991         } else {
 992                 vaf.fmt = fmt;
 993                 vaf.va = &args;
 994
 995                 pr_err("%pV", &vaf);
 996         }
 997
 998         va_end(args);
 999 }
1000
1001 /**
1002  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1003  * @gh: the holder structure to add
1004  *
1005  * Eventually we should move the recursive locking trap to a
1006  * debugging option or something like that. This is the fast
1007  * path and needs to have the minimum number of distractions.
1008  *
1009  */
1010
1011 static inline void add_to_queue(struct gfs2_holder *gh)
1012 __releases(&gl->gl_lockref.lock)
1013 __acquires(&gl->gl_lockref.lock)
1014 {
1015         struct gfs2_glock *gl = gh->gh_gl;
1016         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1017         struct list_head *insert_pt = NULL;
1018         struct gfs2_holder *gh2;
1019         int try_futile = 0;
1020
1021         BUG_ON(gh->gh_owner_pid == NULL);
1022         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1023                 BUG();
1024
1025         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1026                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1027                         try_futile = !may_grant(gl, gh);
1028                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1029                         goto fail;
1030         }
1031
1032         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1033                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1034                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1035                         goto trap_recursive;
1036                 if (try_futile &&
1037                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1038 fail:
1039                         gh->gh_error = GLR_TRYFAILED;
1040                         gfs2_holder_wake(gh);
1041                         return;
1042                 }
1043                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1044                         continue;
1045                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1046                         insert_pt = &gh2->gh_list;
1047         }
1048         set_bit(GLF_QUEUED, &gl->gl_flags);
1049         trace_gfs2_glock_queue(gh, 1);
1050         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1051         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1052         if (likely(insert_pt == NULL)) {
1053                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1054                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1055                         goto do_cancel;
1056                 return;
1057         }
1058         list_add_tail(&gh->gh_list, insert_pt);
1059 do_cancel:
1060         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1061         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1062                 spin_unlock(&gl->gl_lockref.lock);
1063                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1064                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1065                 spin_lock(&gl->gl_lockref.lock);
1066         }
1067         return;
1068
1069 trap_recursive:
1070         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1071         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1072         fs_err(sdp, "lock type: %d req lock state : %d\n",
1073                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1074         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1075         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1076         fs_err(sdp, "lock type: %d req lock state : %d\n",
1077                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1078         gfs2_dump_glock(NULL, gl, true);
1079         BUG();
1080 }
1081
1082 /**
1083  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1084  * @gh: the holder structure
1085  *
1086  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1087  *
1088  * Returns: 0, GLR_TRYFAILED, or errno on failure
1089  */
1090
1091 int gfs2_glock_nq(struct gfs2_holder *gh)
1092 {
1093         struct gfs2_glock *gl = gh->gh_gl;
1094         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1095         int error = 0;
1096
1097         if (unlikely(test_bit(SDF_WITHDRAWN, &sdp->sd_flags)))
1098                 return -EIO;
1099
1100         if (test_bit(GLF_LRU, &gl->gl_flags))
1101                 gfs2_glock_remove_from_lru(gl);
1102
1103         spin_lock(&gl->gl_lockref.lock);
1104         add_to_queue(gh);
1105         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1106                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1107                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1108                 gl->gl_lockref.count++;
1109                 __gfs2_glock_queue_work(gl, 0);
1110         }
1111         run_queue(gl, 1);
1112         spin_unlock(&gl->gl_lockref.lock);
1113
1114         if (!(gh->gh_flags & GL_ASYNC))
1115                 error = gfs2_glock_wait(gh);
1116
1117         return error;
1118 }
1119
1120 /**
1121  * gfs2_glock_poll - poll to see if an async request has been completed
1122  * @gh: the holder
1123  *
1124  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1125  */
1126
1127 int gfs2_glock_poll(struct gfs2_holder *gh)
1128 {
1129         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1130 }
1131
1132 /**
1133  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1134  * @gh: the glock holder
1135  *
1136  */
1137
1138 void gfs2_glock_dq(struct gfs2_holder *gh)
1139 {
1140         struct gfs2_glock *gl = gh->gh_gl;
1141         const struct gfs2_glock_operations *glops = gl->gl_ops;
1142         unsigned delay = 0;
1143         int fast_path = 0;
1144
1145         spin_lock(&gl->gl_lockref.lock);
1146         if (gh->gh_flags & GL_NOCACHE)
1147                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1148
1149         list_del_init(&gh->gh_list);
1150         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1151         if (find_first_holder(gl) == NULL) {
1152                 if (glops->go_unlock) {
1153                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1154                         spin_unlock(&gl->gl_lockref.lock);
1155                         glops->go_unlock(gh);
1156                         spin_lock(&gl->gl_lockref.lock);
1157                         clear_bit(GLF_LOCK, &gl->gl_flags);
1158                 }
1159                 if (list_empty(&gl->gl_holders) &&
1160                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1161                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1162                         fast_path = 1;
1163         }
1164         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1165                 gfs2_glock_add_to_lru(gl);
1166
1167         trace_gfs2_glock_queue(gh, 0);
1168         if (unlikely(!fast_path)) {
1169                 gl->gl_lockref.count++;
1170                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1171                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1172                     gl->gl_name.ln_type == LM_TYPE_INODE)
1173                         delay = gl->gl_hold_time;
1174                 __gfs2_glock_queue_work(gl, delay);
1175         }
1176         spin_unlock(&gl->gl_lockref.lock);
1177 }
1178
1179 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1180 {
1181         struct gfs2_glock *gl = gh->gh_gl;
1182         gfs2_glock_dq(gh);
1183         might_sleep();
1184         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1185 }
1186
1187 /**
1188  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1189  * @gh: the holder structure
1190  *
1191  */
1192
1193 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1194 {
1195         gfs2_glock_dq(gh);
1196         gfs2_holder_uninit(gh);
1197 }
1198
1199 /**
1200  * gfs2_glock_nq_num - acquire a glock based on lock number
1201  * @sdp: the filesystem
1202  * @number: the lock number
1203  * @glops: the glock operations for the type of glock
1204  * @state: the state to acquire the glock in
1205  * @flags: modifier flags for the acquisition
1206  * @gh: the struct gfs2_holder
1207  *
1208  * Returns: errno
1209  */
1210
1211 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1212                       const struct gfs2_glock_operations *glops,
1213                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1214 {
1215         struct gfs2_glock *gl;
1216         int error;
1217
1218         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1219         if (!error) {
1220                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1221                 gfs2_glock_put(gl);
1222         }
1223
1224         return error;
1225 }
1226
1227 /**
1228  * glock_compare - Compare two struct gfs2_glock structures for sorting
1229  * @arg_a: the first structure
1230  * @arg_b: the second structure
1231  *
1232  */
1233
1234 static int glock_compare(const void *arg_a, const void *arg_b)
1235 {
1236         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1237         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1238         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1239         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1240
1241         if (a->ln_number > b->ln_number)
1242                 return 1;
1243         if (a->ln_number < b->ln_number)
1244                 return -1;
1245         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1246         return 0;
1247 }
1248
1249 /**
1250  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1251  * @num_gh: the number of structures
1252  * @ghs: an array of struct gfs2_holder structures
1253  *
1254  * Returns: 0 on success (all glocks acquired),
1255  *          errno on failure (no glocks acquired)
1256  */
1257
1258 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1259                      struct gfs2_holder **p)
1260 {
1261         unsigned int x;
1262         int error = 0;
1263
1264         for (x = 0; x < num_gh; x++)
1265                 p[x] = &ghs[x];
1266
1267         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1268
1269         for (x = 0; x < num_gh; x++) {
1270                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1271
1272                 error = gfs2_glock_nq(p[x]);
1273                 if (error) {
1274                         while (x--)
1275                                 gfs2_glock_dq(p[x]);
1276                         break;
1277                 }
1278         }
1279
1280         return error;
1281 }
1282
1283 /**
1284  * gfs2_glock_nq_m - acquire multiple glocks
1285  * @num_gh: the number of structures
1286  * @ghs: an array of struct gfs2_holder structures
1287  *
1288  *
1289  * Returns: 0 on success (all glocks acquired),
1290  *          errno on failure (no glocks acquired)
1291  */
1292
1293 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1294 {
1295         struct gfs2_holder *tmp[4];
1296         struct gfs2_holder **pph = tmp;
1297         int error = 0;
1298
1299         switch(num_gh) {
1300         case 0:
1301                 return 0;
1302         case 1:
1303                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1304                 return gfs2_glock_nq(ghs);
1305         default:
1306                 if (num_gh <= 4)
1307                         break;
1308                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1309                                     GFP_NOFS);
1310                 if (!pph)
1311                         return -ENOMEM;
1312         }
1313
1314         error = nq_m_sync(num_gh, ghs, pph);
1315
1316         if (pph != tmp)
1317                 kfree(pph);
1318
1319         return error;
1320 }
1321
1322 /**
1323  * gfs2_glock_dq_m - release multiple glocks
1324  * @num_gh: the number of structures
1325  * @ghs: an array of struct gfs2_holder structures
1326  *
1327  */
1328
1329 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1330 {
1331         while (num_gh--)
1332                 gfs2_glock_dq(&ghs[num_gh]);
1333 }
1334
1335 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1336 {
1337         unsigned long delay = 0;
1338         unsigned long holdtime;
1339         unsigned long now = jiffies;
1340
1341         gfs2_glock_hold(gl);
1342         holdtime = gl->gl_tchange + gl->gl_hold_time;
1343         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1344             gl->gl_name.ln_type == LM_TYPE_INODE) {
1345                 if (time_before(now, holdtime))
1346                         delay = holdtime - now;
1347                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1348                         delay = gl->gl_hold_time;
1349         }
1350
1351         spin_lock(&gl->gl_lockref.lock);
1352         handle_callback(gl, state, delay, true);
1353         __gfs2_glock_queue_work(gl, delay);
1354         spin_unlock(&gl->gl_lockref.lock);
1355 }
1356
1357 /**
1358  * gfs2_should_freeze - Figure out if glock should be frozen
1359  * @gl: The glock in question
1360  *
1361  * Glocks are not frozen if (a) the result of the dlm operation is
1362  * an error, (b) the locking operation was an unlock operation or
1363  * (c) if there is a "noexp" flagged request anywhere in the queue
1364  *
1365  * Returns: 1 if freezing should occur, 0 otherwise
1366  */
1367
1368 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1369 {
1370         const struct gfs2_holder *gh;
1371
1372         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1373                 return 0;
1374         if (gl->gl_target == LM_ST_UNLOCKED)
1375                 return 0;
1376
1377         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1378                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1379                         continue;
1380                 if (LM_FLAG_NOEXP & gh->gh_flags)
1381                         return 0;
1382         }
1383
1384         return 1;
1385 }
1386
1387 /**
1388  * gfs2_glock_complete - Callback used by locking
1389  * @gl: Pointer to the glock
1390  * @ret: The return value from the dlm
1391  *
1392  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1393  * to use a bitfield shared with other glock state fields.
1394  */
1395
1396 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1397 {
1398         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1399
1400         spin_lock(&gl->gl_lockref.lock);
1401         gl->gl_reply = ret;
1402
1403         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1404                 if (gfs2_should_freeze(gl)) {
1405                         set_bit(GLF_FROZEN, &gl->gl_flags);
1406                         spin_unlock(&gl->gl_lockref.lock);
1407                         return;
1408                 }
1409         }
1410
1411         gl->gl_lockref.count++;
1412         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1413         __gfs2_glock_queue_work(gl, 0);
1414         spin_unlock(&gl->gl_lockref.lock);
1415 }
1416
1417 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1418 {
1419         struct gfs2_glock *gla, *glb;
1420
1421         gla = list_entry(a, struct gfs2_glock, gl_lru);
1422         glb = list_entry(b, struct gfs2_glock, gl_lru);
1423
1424         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1425                 return 1;
1426         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1427                 return -1;
1428
1429         return 0;
1430 }
1431
1432 /**
1433  * gfs2_dispose_glock_lru - Demote a list of glocks
1434  * @list: The list to dispose of
1435  *
1436  * Disposing of glocks may involve disk accesses, so that here we sort
1437  * the glocks by number (i.e. disk location of the inodes) so that if
1438  * there are any such accesses, they'll be sent in order (mostly).
1439  *
1440  * Must be called under the lru_lock, but may drop and retake this
1441  * lock. While the lru_lock is dropped, entries may vanish from the
1442  * list, but no new entries will appear on the list (since it is
1443  * private)
1444  */
1445
1446 static void gfs2_dispose_glock_lru(struct list_head *list)
1447 __releases(&lru_lock)
1448 __acquires(&lru_lock)
1449 {
1450         struct gfs2_glock *gl;
1451
1452         list_sort(NULL, list, glock_cmp);
1453
1454         while(!list_empty(list)) {
1455                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1456                 list_del_init(&gl->gl_lru);
1457                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1458 add_back_to_lru:
1459                         list_add(&gl->gl_lru, &lru_list);
1460                         set_bit(GLF_LRU, &gl->gl_flags);
1461                         atomic_inc(&lru_count);
1462                         continue;
1463                 }
1464                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1465                         spin_unlock(&gl->gl_lockref.lock);
1466                         goto add_back_to_lru;
1467                 }
1468                 gl->gl_lockref.count++;
1469                 if (demote_ok(gl))
1470                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1471                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1472                 __gfs2_glock_queue_work(gl, 0);
1473                 spin_unlock(&gl->gl_lockref.lock);
1474                 cond_resched_lock(&lru_lock);
1475         }
1476 }
1477
1478 /**
1479  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1480  * @nr: The number of entries to scan
1481  *
1482  * This function selects the entries on the LRU which are able to
1483  * be demoted, and then kicks off the process by calling
1484  * gfs2_dispose_glock_lru() above.
1485  */
1486
1487 static long gfs2_scan_glock_lru(int nr)
1488 {
1489         struct gfs2_glock *gl;
1490         LIST_HEAD(skipped);
1491         LIST_HEAD(dispose);
1492         long freed = 0;
1493
1494         spin_lock(&lru_lock);
1495         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1496                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1497
1498                 /* Test for being demotable */
1499                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1500                         list_move(&gl->gl_lru, &dispose);
1501                         atomic_dec(&lru_count);
1502                         clear_bit(GLF_LRU, &gl->gl_flags);
1503                         freed++;
1504                         continue;
1505                 }
1506
1507                 list_move(&gl->gl_lru, &skipped);
1508         }
1509         list_splice(&skipped, &lru_list);
1510         if (!list_empty(&dispose))
1511                 gfs2_dispose_glock_lru(&dispose);
1512         spin_unlock(&lru_lock);
1513
1514         return freed;
1515 }
1516
1517 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1518                                             struct shrink_control *sc)
1519 {
1520         if (!(sc->gfp_mask & __GFP_FS))
1521                 return SHRINK_STOP;
1522         return gfs2_scan_glock_lru(sc->nr_to_scan);
1523 }
1524
1525 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1526                                              struct shrink_control *sc)
1527 {
1528         return vfs_pressure_ratio(atomic_read(&lru_count));
1529 }
1530
1531 static struct shrinker glock_shrinker = {
1532         .seeks = DEFAULT_SEEKS,
1533         .count_objects = gfs2_glock_shrink_count,
1534         .scan_objects = gfs2_glock_shrink_scan,
1535 };
1536
1537 /**
1538  * examine_bucket - Call a function for glock in a hash bucket
1539  * @examiner: the function
1540  * @sdp: the filesystem
1541  * @bucket: the bucket
1542  *
1543  * Note that the function can be called multiple times on the same
1544  * object.  So the user must ensure that the function can cope with
1545  * that.
1546  */
1547
1548 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1549 {
1550         struct gfs2_glock *gl;
1551         struct rhashtable_iter iter;
1552
1553         rhashtable_walk_enter(&gl_hash_table, &iter);
1554
1555         do {
1556                 rhashtable_walk_start(&iter);
1557
1558                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1559                         if (gl->gl_name.ln_sbd == sdp &&
1560                             lockref_get_not_dead(&gl->gl_lockref))
1561                                 examiner(gl);
1562
1563                 rhashtable_walk_stop(&iter);
1564         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1565
1566         rhashtable_walk_exit(&iter);
1567 }
1568
1569 /**
1570  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1571  * @gl: The glock to thaw
1572  *
1573  */
1574
1575 static void thaw_glock(struct gfs2_glock *gl)
1576 {
1577         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1578                 gfs2_glock_put(gl);
1579                 return;
1580         }
1581         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1582         gfs2_glock_queue_work(gl, 0);
1583 }
1584
1585 /**
1586  * clear_glock - look at a glock and see if we can free it from glock cache
1587  * @gl: the glock to look at
1588  *
1589  */
1590
1591 static void clear_glock(struct gfs2_glock *gl)
1592 {
1593         gfs2_glock_remove_from_lru(gl);
1594
1595         spin_lock(&gl->gl_lockref.lock);
1596         if (gl->gl_state != LM_ST_UNLOCKED)
1597                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1598         __gfs2_glock_queue_work(gl, 0);
1599         spin_unlock(&gl->gl_lockref.lock);
1600 }
1601
1602 /**
1603  * gfs2_glock_thaw - Thaw any frozen glocks
1604  * @sdp: The super block
1605  *
1606  */
1607
1608 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1609 {
1610         glock_hash_walk(thaw_glock, sdp);
1611 }
1612
1613 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1614 {
1615         spin_lock(&gl->gl_lockref.lock);
1616         gfs2_dump_glock(seq, gl, fsid);
1617         spin_unlock(&gl->gl_lockref.lock);
1618 }
1619
1620 static void dump_glock_func(struct gfs2_glock *gl)
1621 {
1622         dump_glock(NULL, gl, true);
1623 }
1624
1625 /**
1626  * gfs2_gl_hash_clear - Empty out the glock hash table
1627  * @sdp: the filesystem
1628  * @wait: wait until it's all gone
1629  *
1630  * Called when unmounting the filesystem.
1631  */
1632
1633 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1634 {
1635         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1636         flush_workqueue(glock_workqueue);
1637         glock_hash_walk(clear_glock, sdp);
1638         flush_workqueue(glock_workqueue);
1639         wait_event_timeout(sdp->sd_glock_wait,
1640                            atomic_read(&sdp->sd_glock_disposal) == 0,
1641                            HZ * 600);
1642         glock_hash_walk(dump_glock_func, sdp);
1643 }
1644
1645 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1646 {
1647         struct gfs2_glock *gl = ip->i_gl;
1648         int ret;
1649
1650         ret = gfs2_truncatei_resume(ip);
1651         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1652
1653         spin_lock(&gl->gl_lockref.lock);
1654         clear_bit(GLF_LOCK, &gl->gl_flags);
1655         run_queue(gl, 1);
1656         spin_unlock(&gl->gl_lockref.lock);
1657 }
1658
1659 static const char *state2str(unsigned state)
1660 {
1661         switch(state) {
1662         case LM_ST_UNLOCKED:
1663                 return "UN";
1664         case LM_ST_SHARED:
1665                 return "SH";
1666         case LM_ST_DEFERRED:
1667                 return "DF";
1668         case LM_ST_EXCLUSIVE:
1669                 return "EX";
1670         }
1671         return "??";
1672 }
1673
1674 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1675 {
1676         char *p = buf;
1677         if (flags & LM_FLAG_TRY)
1678                 *p++ = 't';
1679         if (flags & LM_FLAG_TRY_1CB)
1680                 *p++ = 'T';
1681         if (flags & LM_FLAG_NOEXP)
1682                 *p++ = 'e';
1683         if (flags & LM_FLAG_ANY)
1684                 *p++ = 'A';
1685         if (flags & LM_FLAG_PRIORITY)
1686                 *p++ = 'p';
1687         if (flags & GL_ASYNC)
1688                 *p++ = 'a';
1689         if (flags & GL_EXACT)
1690                 *p++ = 'E';
1691         if (flags & GL_NOCACHE)
1692                 *p++ = 'c';
1693         if (test_bit(HIF_HOLDER, &iflags))
1694                 *p++ = 'H';
1695         if (test_bit(HIF_WAIT, &iflags))
1696                 *p++ = 'W';
1697         if (test_bit(HIF_FIRST, &iflags))
1698                 *p++ = 'F';
1699         *p = 0;
1700         return buf;
1701 }
1702
1703 /**
1704  * dump_holder - print information about a glock holder
1705  * @seq: the seq_file struct
1706  * @gh: the glock holder
1707  * @fs_id_buf: pointer to file system id (if requested)
1708  *
1709  */
1710
1711 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh,
1712                         const char *fs_id_buf)
1713 {
1714         struct task_struct *gh_owner = NULL;
1715         char flags_buf[32];
1716
1717         rcu_read_lock();
1718         if (gh->gh_owner_pid)
1719                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1720         gfs2_print_dbg(seq, "%s H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1721                        fs_id_buf, state2str(gh->gh_state),
1722                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1723                        gh->gh_error,
1724                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1725                        gh_owner ? gh_owner->comm : "(ended)",
1726                        (void *)gh->gh_ip);
1727         rcu_read_unlock();
1728 }
1729
1730 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1731 {
1732         const unsigned long *gflags = &gl->gl_flags;
1733         char *p = buf;
1734
1735         if (test_bit(GLF_LOCK, gflags))
1736                 *p++ = 'l';
1737         if (test_bit(GLF_DEMOTE, gflags))
1738                 *p++ = 'D';
1739         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1740                 *p++ = 'd';
1741         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1742                 *p++ = 'p';
1743         if (test_bit(GLF_DIRTY, gflags))
1744                 *p++ = 'y';
1745         if (test_bit(GLF_LFLUSH, gflags))
1746                 *p++ = 'f';
1747         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1748                 *p++ = 'i';
1749         if (test_bit(GLF_REPLY_PENDING, gflags))
1750                 *p++ = 'r';
1751         if (test_bit(GLF_INITIAL, gflags))
1752                 *p++ = 'I';
1753         if (test_bit(GLF_FROZEN, gflags))
1754                 *p++ = 'F';
1755         if (test_bit(GLF_QUEUED, gflags))
1756                 *p++ = 'q';
1757         if (test_bit(GLF_LRU, gflags))
1758                 *p++ = 'L';
1759         if (gl->gl_object)
1760                 *p++ = 'o';
1761         if (test_bit(GLF_BLOCKING, gflags))
1762                 *p++ = 'b';
1763         *p = 0;
1764         return buf;
1765 }
1766
1767 /**
1768  * gfs2_dump_glock - print information about a glock
1769  * @seq: The seq_file struct
1770  * @gl: the glock
1771  * @fsid: If true, also dump the file system id
1772  *
1773  * The file format is as follows:
1774  * One line per object, capital letters are used to indicate objects
1775  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1776  * other objects are indented by a single space and follow the glock to
1777  * which they are related. Fields are indicated by lower case letters
1778  * followed by a colon and the field value, except for strings which are in
1779  * [] so that its possible to see if they are composed of spaces for
1780  * example. The field's are n = number (id of the object), f = flags,
1781  * t = type, s = state, r = refcount, e = error, p = pid.
1782  *
1783  */
1784
1785 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1786 {
1787         const struct gfs2_glock_operations *glops = gl->gl_ops;
1788         unsigned long long dtime;
1789         const struct gfs2_holder *gh;
1790         char gflags_buf[32];
1791         char fs_id_buf[GFS2_FSNAME_LEN + 3 * sizeof(int) + 2];
1792         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1793
1794         memset(fs_id_buf, 0, sizeof(fs_id_buf));
1795         if (fsid && sdp) /* safety precaution */
1796                 sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
1797         dtime = jiffies - gl->gl_demote_time;
1798         dtime *= 1000000/HZ; /* demote time in uSec */
1799         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1800                 dtime = 0;
1801         gfs2_print_dbg(seq, "%sG:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d "
1802                        "v:%d r:%d m:%ld\n", fs_id_buf, state2str(gl->gl_state),
1803                   gl->gl_name.ln_type,
1804                   (unsigned long long)gl->gl_name.ln_number,
1805                   gflags2str(gflags_buf, gl),
1806                   state2str(gl->gl_target),
1807                   state2str(gl->gl_demote_state), dtime,
1808                   atomic_read(&gl->gl_ail_count),
1809                   atomic_read(&gl->gl_revokes),
1810                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1811
1812         list_for_each_entry(gh, &gl->gl_holders, gh_list)
1813                 dump_holder(seq, gh, fs_id_buf);
1814
1815         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1816                 glops->go_dump(seq, gl, fs_id_buf);
1817 }
1818
1819 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
1820 {
1821         struct gfs2_glock *gl = iter_ptr;
1822
1823         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
1824                    gl->gl_name.ln_type,
1825                    (unsigned long long)gl->gl_name.ln_number,
1826                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
1827                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
1828                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
1829                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
1830                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
1831                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
1832                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
1833                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
1834         return 0;
1835 }
1836
1837 static const char *gfs2_gltype[] = {
1838         "type",
1839         "reserved",
1840         "nondisk",
1841         "inode",
1842         "rgrp",
1843         "meta",
1844         "iopen",
1845         "flock",
1846         "plock",
1847         "quota",
1848         "journal",
1849 };
1850
1851 static const char *gfs2_stype[] = {
1852         [GFS2_LKS_SRTT]         = "srtt",
1853         [GFS2_LKS_SRTTVAR]      = "srttvar",
1854         [GFS2_LKS_SRTTB]        = "srttb",
1855         [GFS2_LKS_SRTTVARB]     = "srttvarb",
1856         [GFS2_LKS_SIRT]         = "sirt",
1857         [GFS2_LKS_SIRTVAR]      = "sirtvar",
1858         [GFS2_LKS_DCOUNT]       = "dlm",
1859         [GFS2_LKS_QCOUNT]       = "queue",
1860 };
1861
1862 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
1863
1864 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
1865 {
1866         struct gfs2_sbd *sdp = seq->private;
1867         loff_t pos = *(loff_t *)iter_ptr;
1868         unsigned index = pos >> 3;
1869         unsigned subindex = pos & 0x07;
1870         int i;
1871
1872         if (index == 0 && subindex != 0)
1873                 return 0;
1874
1875         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
1876                    (index == 0) ? "cpu": gfs2_stype[subindex]);
1877
1878         for_each_possible_cpu(i) {
1879                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
1880
1881                 if (index == 0)
1882                         seq_printf(seq, " %15u", i);
1883                 else
1884                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
1885                                    lkstats[index - 1].stats[subindex]);
1886         }
1887         seq_putc(seq, '\n');
1888         return 0;
1889 }
1890
1891 int __init gfs2_glock_init(void)
1892 {
1893         int i, ret;
1894
1895         ret = rhashtable_init(&gl_hash_table, &ht_parms);
1896         if (ret < 0)
1897                 return ret;
1898
1899         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1900                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1901         if (!glock_workqueue) {
1902                 rhashtable_destroy(&gl_hash_table);
1903                 return -ENOMEM;
1904         }
1905         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1906                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1907                                                 0);
1908         if (!gfs2_delete_workqueue) {
1909                 destroy_workqueue(glock_workqueue);
1910                 rhashtable_destroy(&gl_hash_table);
1911                 return -ENOMEM;
1912         }
1913
1914         ret = register_shrinker(&glock_shrinker);
1915         if (ret) {
1916                 destroy_workqueue(gfs2_delete_workqueue);
1917                 destroy_workqueue(glock_workqueue);
1918                 rhashtable_destroy(&gl_hash_table);
1919                 return ret;
1920         }
1921
1922         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
1923                 init_waitqueue_head(glock_wait_table + i);
1924
1925         return 0;
1926 }
1927
1928 void gfs2_glock_exit(void)
1929 {
1930         unregister_shrinker(&glock_shrinker);
1931         rhashtable_destroy(&gl_hash_table);
1932         destroy_workqueue(glock_workqueue);
1933         destroy_workqueue(gfs2_delete_workqueue);
1934 }
1935
1936 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
1937 {
1938         struct gfs2_glock *gl = gi->gl;
1939
1940         if (gl) {
1941                 if (n == 0)
1942                         return;
1943                 if (!lockref_put_not_zero(&gl->gl_lockref))
1944                         gfs2_glock_queue_put(gl);
1945         }
1946         for (;;) {
1947                 gl = rhashtable_walk_next(&gi->hti);
1948                 if (IS_ERR_OR_NULL(gl)) {
1949                         if (gl == ERR_PTR(-EAGAIN)) {
1950                                 n = 1;
1951                                 continue;
1952                         }
1953                         gl = NULL;
1954                         break;
1955                 }
1956                 if (gl->gl_name.ln_sbd != gi->sdp)
1957                         continue;
1958                 if (n <= 1) {
1959                         if (!lockref_get_not_dead(&gl->gl_lockref))
1960                                 continue;
1961                         break;
1962                 } else {
1963                         if (__lockref_is_dead(&gl->gl_lockref))
1964                                 continue;
1965                         n--;
1966                 }
1967         }
1968         gi->gl = gl;
1969 }
1970
1971 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1972         __acquires(RCU)
1973 {
1974         struct gfs2_glock_iter *gi = seq->private;
1975         loff_t n;
1976
1977         /*
1978          * We can either stay where we are, skip to the next hash table
1979          * entry, or start from the beginning.
1980          */
1981         if (*pos < gi->last_pos) {
1982                 rhashtable_walk_exit(&gi->hti);
1983                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
1984                 n = *pos + 1;
1985         } else {
1986                 n = *pos - gi->last_pos;
1987         }
1988
1989         rhashtable_walk_start(&gi->hti);
1990
1991         gfs2_glock_iter_next(gi, n);
1992         gi->last_pos = *pos;
1993         return gi->gl;
1994 }
1995
1996 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1997                                  loff_t *pos)
1998 {
1999         struct gfs2_glock_iter *gi = seq->private;
2000
2001         (*pos)++;
2002         gi->last_pos = *pos;
2003         gfs2_glock_iter_next(gi, 1);
2004         return gi->gl;
2005 }
2006
2007 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2008         __releases(RCU)
2009 {
2010         struct gfs2_glock_iter *gi = seq->private;
2011
2012         rhashtable_walk_stop(&gi->hti);
2013 }
2014
2015 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2016 {
2017         dump_glock(seq, iter_ptr, false);
2018         return 0;
2019 }
2020
2021 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2022 {
2023         preempt_disable();
2024         if (*pos >= GFS2_NR_SBSTATS)
2025                 return NULL;
2026         return pos;
2027 }
2028
2029 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2030                                    loff_t *pos)
2031 {
2032         (*pos)++;
2033         if (*pos >= GFS2_NR_SBSTATS)
2034                 return NULL;
2035         return pos;
2036 }
2037
2038 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2039 {
2040         preempt_enable();
2041 }
2042
2043 static const struct seq_operations gfs2_glock_seq_ops = {
2044         .start = gfs2_glock_seq_start,
2045         .next  = gfs2_glock_seq_next,
2046         .stop  = gfs2_glock_seq_stop,
2047         .show  = gfs2_glock_seq_show,
2048 };
2049
2050 static const struct seq_operations gfs2_glstats_seq_ops = {
2051         .start = gfs2_glock_seq_start,
2052         .next  = gfs2_glock_seq_next,
2053         .stop  = gfs2_glock_seq_stop,
2054         .show  = gfs2_glstats_seq_show,
2055 };
2056
2057 static const struct seq_operations gfs2_sbstats_seq_ops = {
2058         .start = gfs2_sbstats_seq_start,
2059         .next  = gfs2_sbstats_seq_next,
2060         .stop  = gfs2_sbstats_seq_stop,
2061         .show  = gfs2_sbstats_seq_show,
2062 };
2063
2064 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2065
2066 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2067                               const struct seq_operations *ops)
2068 {
2069         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2070         if (ret == 0) {
2071                 struct seq_file *seq = file->private_data;
2072                 struct gfs2_glock_iter *gi = seq->private;
2073
2074                 gi->sdp = inode->i_private;
2075                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2076                 if (seq->buf)
2077                         seq->size = GFS2_SEQ_GOODSIZE;
2078                 /*
2079                  * Initially, we are "before" the first hash table entry; the
2080                  * first call to rhashtable_walk_next gets us the first entry.
2081                  */
2082                 gi->last_pos = -1;
2083                 gi->gl = NULL;
2084                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2085         }
2086         return ret;
2087 }
2088
2089 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2090 {
2091         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2092 }
2093
2094 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2095 {
2096         struct seq_file *seq = file->private_data;
2097         struct gfs2_glock_iter *gi = seq->private;
2098
2099         if (gi->gl)
2100                 gfs2_glock_put(gi->gl);
2101         rhashtable_walk_exit(&gi->hti);
2102         return seq_release_private(inode, file);
2103 }
2104
2105 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2106 {
2107         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2108 }
2109
2110 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2111 {
2112         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2113         if (ret == 0) {
2114                 struct seq_file *seq = file->private_data;
2115                 seq->private = inode->i_private;  /* sdp */
2116         }
2117         return ret;
2118 }
2119
2120 static const struct file_operations gfs2_glocks_fops = {
2121         .owner   = THIS_MODULE,
2122         .open    = gfs2_glocks_open,
2123         .read    = seq_read,
2124         .llseek  = seq_lseek,
2125         .release = gfs2_glocks_release,
2126 };
2127
2128 static const struct file_operations gfs2_glstats_fops = {
2129         .owner   = THIS_MODULE,
2130         .open    = gfs2_glstats_open,
2131         .read    = seq_read,
2132         .llseek  = seq_lseek,
2133         .release = gfs2_glocks_release,
2134 };
2135
2136 static const struct file_operations gfs2_sbstats_fops = {
2137         .owner   = THIS_MODULE,
2138         .open    = gfs2_sbstats_open,
2139         .read    = seq_read,
2140         .llseek  = seq_lseek,
2141         .release = seq_release,
2142 };
2143
2144 void gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2145 {
2146         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2147
2148         debugfs_create_file("glocks", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2149                             &gfs2_glocks_fops);
2150
2151         debugfs_create_file("glstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2152                             &gfs2_glstats_fops);
2153
2154         debugfs_create_file("sbstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2155                             &gfs2_sbstats_fops);
2156 }
2157
2158 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2159 {
2160         debugfs_remove_recursive(sdp->debugfs_dir);
2161         sdp->debugfs_dir = NULL;
2162 }
2163
2164 void gfs2_register_debugfs(void)
2165 {
2166         gfs2_root = debugfs_create_dir("gfs2", NULL);
2167 }
2168
2169 void gfs2_unregister_debugfs(void)
2170 {
2171         debugfs_remove(gfs2_root);
2172         gfs2_root = NULL;
2173 }