fs/gfs2/glock.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/sched.h>
  13 #include <linux/slab.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/buffer_head.h>
  16 #include <linux/delay.h>
  17 #include <linux/sort.h>
  18 #include <linux/jhash.h>
  19 #include <linux/kallsyms.h>
  20 #include <linux/gfs2_ondisk.h>
  21 #include <linux/list.h>
  22 #include <linux/wait.h>
  23 #include <linux/module.h>
  24 #include <linux/uaccess.h>
  25 #include <linux/seq_file.h>
  26 #include <linux/debugfs.h>
  27 #include <linux/kthread.h>
  28 #include <linux/freezer.h>
  29 #include <linux/workqueue.h>
  30 #include <linux/jiffies.h>
  31 #include <linux/rcupdate.h>
  32 #include <linux/rculist_bl.h>
  33 #include <linux/bit_spinlock.h>
  34 #include <linux/percpu.h>
  35 #include <linux/list_sort.h>
  36 #include <linux/lockref.h>
  37 #include <linux/rhashtable.h>
  38
  39 #include "gfs2.h"
  40 #include "incore.h"
  41 #include "glock.h"
  42 #include "glops.h"
  43 #include "inode.h"
  44 #include "lops.h"
  45 #include "meta_io.h"
  46 #include "quota.h"
  47 #include "super.h"
  48 #include "util.h"
  49 #include "bmap.h"
  50 #define CREATE_TRACE_POINTS
  51 #include "trace_gfs2.h"
  52
  53 struct gfs2_glock_iter {
  54         struct gfs2_sbd *sdp;           /* incore superblock           */
  55         struct rhashtable_iter hti;     /* rhashtable iterator         */
  56         struct gfs2_glock *gl;          /* current glock struct        */
  57         loff_t last_pos;                /* last position               */
  58 };
  59
  60 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  61
  62 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  63
  64 static struct dentry *gfs2_root;
  65 static struct workqueue_struct *glock_workqueue;
  66 struct workqueue_struct *gfs2_delete_workqueue;
  67 static LIST_HEAD(lru_list);
  68 static atomic_t lru_count = ATOMIC_INIT(0);
  69 static DEFINE_SPINLOCK(lru_lock);
  70
  71 #define GFS2_GL_HASH_SHIFT      15
  72 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
  73
  74 static struct rhashtable_params ht_parms = {
  75         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
  76         .key_len = offsetofend(struct lm_lockname, ln_type),
  77         .key_offset = offsetof(struct gfs2_glock, gl_name),
  78         .head_offset = offsetof(struct gfs2_glock, gl_node),
  79 };
  80
  81 static struct rhashtable gl_hash_table;
  82
  83 static void gfs2_glock_dealloc(struct rcu_head *rcu)
  84 {
  85         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
  86
  87         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
  88                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
  89         } else {
  90                 kfree(gl->gl_lksb.sb_lvbptr);
  91                 kmem_cache_free(gfs2_glock_cachep, gl);
  92         }
  93 }
  94
  95 void gfs2_glock_free(struct gfs2_glock *gl)
  96 {
  97         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
  98
  99         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 100         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 101                 wake_up(&sdp->sd_glock_wait);
 102 }
 103
 104 /**
 105  * gfs2_glock_hold() - increment reference count on glock
 106  * @gl: The glock to hold
 107  *
 108  */
 109
 110 static void gfs2_glock_hold(struct gfs2_glock *gl)
 111 {
 112         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 113         lockref_get(&gl->gl_lockref);
 114 }
 115
 116 /**
 117  * demote_ok - Check to see if it's ok to unlock a glock
 118  * @gl: the glock
 119  *
 120  * Returns: 1 if it's ok
 121  */
 122
 123 static int demote_ok(const struct gfs2_glock *gl)
 124 {
 125         const struct gfs2_glock_operations *glops = gl->gl_ops;
 126
 127         if (gl->gl_state == LM_ST_UNLOCKED)
 128                 return 0;
 129         if (!list_empty(&gl->gl_holders))
 130                 return 0;
 131         if (glops->go_demote_ok)
 132                 return glops->go_demote_ok(gl);
 133         return 1;
 134 }
 135
 136
 137 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 138 {
 139         spin_lock(&lru_lock);
 140
 141         if (!list_empty(&gl->gl_lru))
 142                 list_del_init(&gl->gl_lru);
 143         else
 144                 atomic_inc(&lru_count);
 145
 146         list_add_tail(&gl->gl_lru, &lru_list);
 147         set_bit(GLF_LRU, &gl->gl_flags);
 148         spin_unlock(&lru_lock);
 149 }
 150
 151 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 152 {
 153         spin_lock(&lru_lock);
 154         if (!list_empty(&gl->gl_lru)) {
 155                 list_del_init(&gl->gl_lru);
 156                 atomic_dec(&lru_count);
 157                 clear_bit(GLF_LRU, &gl->gl_flags);
 158         }
 159         spin_unlock(&lru_lock);
 160 }
 161
 162 /*
 163  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 164  * work queue.
 165  */
 166 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 167         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 168                 /*
 169                  * We are holding the lockref spinlock, and the work was still
 170                  * queued above.  The queued work (glock_work_func) takes that
 171                  * spinlock before dropping its glock reference(s), so it
 172                  * cannot have dropped them in the meantime.
 173                  */
 174                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 175                 gl->gl_lockref.count--;
 176         }
 177 }
 178
 179 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 180         spin_lock(&gl->gl_lockref.lock);
 181         __gfs2_glock_queue_work(gl, delay);
 182         spin_unlock(&gl->gl_lockref.lock);
 183 }
 184
 185 static void __gfs2_glock_put(struct gfs2_glock *gl)
 186 {
 187         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 188         struct address_space *mapping = gfs2_glock2aspace(gl);
 189
 190         lockref_mark_dead(&gl->gl_lockref);
 191
 192         gfs2_glock_remove_from_lru(gl);
 193         spin_unlock(&gl->gl_lockref.lock);
 194         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 195         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 196         GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
 197         trace_gfs2_glock_put(gl);
 198         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 199 }
 200
 201 /**
 202  * gfs2_glock_put() - Decrement reference count on glock
 203  * @gl: The glock to put
 204  *
 205  */
 206
 207 void gfs2_glock_put(struct gfs2_glock *gl)
 208 {
 209         if (lockref_put_or_lock(&gl->gl_lockref))
 210                 return;
 211
 212         __gfs2_glock_put(gl);
 213 }
 214
 215 /**
 216  * may_grant - check if its ok to grant a new lock
 217  * @gl: The glock
 218  * @gh: The lock request which we wish to grant
 219  *
 220  * Returns: true if its ok to grant the lock
 221  */
 222
 223 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 224 {
 225         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 226         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 227              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 228                 return 0;
 229         if (gl->gl_state == gh->gh_state)
 230                 return 1;
 231         if (gh->gh_flags & GL_EXACT)
 232                 return 0;
 233         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 234                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 235                         return 1;
 236                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 237                         return 1;
 238         }
 239         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 240                 return 1;
 241         return 0;
 242 }
 243
 244 static void gfs2_holder_wake(struct gfs2_holder *gh)
 245 {
 246         clear_bit(HIF_WAIT, &gh->gh_iflags);
 247         smp_mb__after_atomic();
 248         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 249 }
 250
 251 /**
 252  * do_error - Something unexpected has happened during a lock request
 253  *
 254  */
 255
 256 static void do_error(struct gfs2_glock *gl, const int ret)
 257 {
 258         struct gfs2_holder *gh, *tmp;
 259
 260         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 261                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 262                         continue;
 263                 if (ret & LM_OUT_ERROR)
 264                         gh->gh_error = -EIO;
 265                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 266                         gh->gh_error = GLR_TRYFAILED;
 267                 else
 268                         continue;
 269                 list_del_init(&gh->gh_list);
 270                 trace_gfs2_glock_queue(gh, 0);
 271                 gfs2_holder_wake(gh);
 272         }
 273 }
 274
 275 /**
 276  * do_promote - promote as many requests as possible on the current queue
 277  * @gl: The glock
 278  *
 279  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 280  *          if a type specific operation is underway.
 281  */
 282
 283 static int do_promote(struct gfs2_glock *gl)
 284 __releases(&gl->gl_lockref.lock)
 285 __acquires(&gl->gl_lockref.lock)
 286 {
 287         const struct gfs2_glock_operations *glops = gl->gl_ops;
 288         struct gfs2_holder *gh, *tmp;
 289         int ret;
 290
 291 restart:
 292         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 293                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 294                         continue;
 295                 if (may_grant(gl, gh)) {
 296                         if (gh->gh_list.prev == &gl->gl_holders &&
 297                             glops->go_lock) {
 298                                 spin_unlock(&gl->gl_lockref.lock);
 299                                 /* FIXME: eliminate this eventually */
 300                                 ret = glops->go_lock(gh);
 301                                 spin_lock(&gl->gl_lockref.lock);
 302                                 if (ret) {
 303                                         if (ret == 1)
 304                                                 return 2;
 305                                         gh->gh_error = ret;
 306                                         list_del_init(&gh->gh_list);
 307                                         trace_gfs2_glock_queue(gh, 0);
 308                                         gfs2_holder_wake(gh);
 309                                         goto restart;
 310                                 }
 311                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 312                                 trace_gfs2_promote(gh, 1);
 313                                 gfs2_holder_wake(gh);
 314                                 goto restart;
 315                         }
 316                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 317                         trace_gfs2_promote(gh, 0);
 318                         gfs2_holder_wake(gh);
 319                         continue;
 320                 }
 321                 if (gh->gh_list.prev == &gl->gl_holders)
 322                         return 1;
 323                 do_error(gl, 0);
 324                 break;
 325         }
 326         return 0;
 327 }
 328
 329 /**
 330  * find_first_waiter - find the first gh that's waiting for the glock
 331  * @gl: the glock
 332  */
 333
 334 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 335 {
 336         struct gfs2_holder *gh;
 337
 338         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 339                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 340                         return gh;
 341         }
 342         return NULL;
 343 }
 344
 345 /**
 346  * state_change - record that the glock is now in a different state
 347  * @gl: the glock
 348  * @new_state the new state
 349  *
 350  */
 351
 352 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 353 {
 354         int held1, held2;
 355
 356         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 357         held2 = (new_state != LM_ST_UNLOCKED);
 358
 359         if (held1 != held2) {
 360                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 361                 if (held2)
 362                         gl->gl_lockref.count++;
 363                 else
 364                         gl->gl_lockref.count--;
 365         }
 366         if (held1 && held2 && list_empty(&gl->gl_holders))
 367                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 368
 369         if (new_state != gl->gl_target)
 370                 /* shorten our minimum hold time */
 371                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 372                                        GL_GLOCK_MIN_HOLD);
 373         gl->gl_state = new_state;
 374         gl->gl_tchange = jiffies;
 375 }
 376
 377 static void gfs2_demote_wake(struct gfs2_glock *gl)
 378 {
 379         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 380         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 381         smp_mb__after_atomic();
 382         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 383 }
 384
 385 /**
 386  * finish_xmote - The DLM has replied to one of our lock requests
 387  * @gl: The glock
 388  * @ret: The status from the DLM
 389  *
 390  */
 391
 392 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 393 {
 394         const struct gfs2_glock_operations *glops = gl->gl_ops;
 395         struct gfs2_holder *gh;
 396         unsigned state = ret & LM_OUT_ST_MASK;
 397         int rv;
 398
 399         spin_lock(&gl->gl_lockref.lock);
 400         trace_gfs2_glock_state_change(gl, state);
 401         state_change(gl, state);
 402         gh = find_first_waiter(gl);
 403
 404         /* Demote to UN request arrived during demote to SH or DF */
 405         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 406             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 407                 gl->gl_target = LM_ST_UNLOCKED;
 408
 409         /* Check for state != intended state */
 410         if (unlikely(state != gl->gl_target)) {
 411                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 412                         /* move to back of queue and try next entry */
 413                         if (ret & LM_OUT_CANCELED) {
 414                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 415                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 416                                 gh = find_first_waiter(gl);
 417                                 gl->gl_target = gh->gh_state;
 418                                 goto retry;
 419                         }
 420                         /* Some error or failed "try lock" - report it */
 421                         if ((ret & LM_OUT_ERROR) ||
 422                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 423                                 gl->gl_target = gl->gl_state;
 424                                 do_error(gl, ret);
 425                                 goto out;
 426                         }
 427                 }
 428                 switch(state) {
 429                 /* Unlocked due to conversion deadlock, try again */
 430                 case LM_ST_UNLOCKED:
 431 retry:
 432                         do_xmote(gl, gh, gl->gl_target);
 433                         break;
 434                 /* Conversion fails, unlock and try again */
 435                 case LM_ST_SHARED:
 436                 case LM_ST_DEFERRED:
 437                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 438                         break;
 439                 default: /* Everything else */
 440                         pr_err("wanted %u got %u\n", gl->gl_target, state);
 441                         GLOCK_BUG_ON(gl, 1);
 442                 }
 443                 spin_unlock(&gl->gl_lockref.lock);
 444                 return;
 445         }
 446
 447         /* Fast path - we got what we asked for */
 448         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 449                 gfs2_demote_wake(gl);
 450         if (state != LM_ST_UNLOCKED) {
 451                 if (glops->go_xmote_bh) {
 452                         spin_unlock(&gl->gl_lockref.lock);
 453                         rv = glops->go_xmote_bh(gl, gh);
 454                         spin_lock(&gl->gl_lockref.lock);
 455                         if (rv) {
 456                                 do_error(gl, rv);
 457                                 goto out;
 458                         }
 459                 }
 460                 rv = do_promote(gl);
 461                 if (rv == 2)
 462                         goto out_locked;
 463         }
 464 out:
 465         clear_bit(GLF_LOCK, &gl->gl_flags);
 466 out_locked:
 467         spin_unlock(&gl->gl_lockref.lock);
 468 }
 469
 470 /**
 471  * do_xmote - Calls the DLM to change the state of a lock
 472  * @gl: The lock state
 473  * @gh: The holder (only for promotes)
 474  * @target: The target lock state
 475  *
 476  */
 477
 478 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 479 __releases(&gl->gl_lockref.lock)
 480 __acquires(&gl->gl_lockref.lock)
 481 {
 482         const struct gfs2_glock_operations *glops = gl->gl_ops;
 483         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 484         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 485         int ret;
 486
 487         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)) &&
 488             target != LM_ST_UNLOCKED)
 489                 return;
 490         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 491                       LM_FLAG_PRIORITY);
 492         GLOCK_BUG_ON(gl, gl->gl_state == target);
 493         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 494         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 495             glops->go_inval) {
 496                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 497                 do_error(gl, 0); /* Fail queued try locks */
 498         }
 499         gl->gl_req = target;
 500         set_bit(GLF_BLOCKING, &gl->gl_flags);
 501         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 502             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 503             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 504                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 505         spin_unlock(&gl->gl_lockref.lock);
 506         if (glops->go_sync)
 507                 glops->go_sync(gl);
 508         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 509                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 510         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 511
 512         gfs2_glock_hold(gl);
 513         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 514                 /* lock_dlm */
 515                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 516                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 517                     target == LM_ST_UNLOCKED &&
 518                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 519                         finish_xmote(gl, target);
 520                         gfs2_glock_queue_work(gl, 0);
 521                 }
 522                 else if (ret) {
 523                         pr_err("lm_lock ret %d\n", ret);
 524                         GLOCK_BUG_ON(gl, !test_bit(SDF_SHUTDOWN,
 525                                                    &sdp->sd_flags));
 526                 }
 527         } else { /* lock_nolock */
 528                 finish_xmote(gl, target);
 529                 gfs2_glock_queue_work(gl, 0);
 530         }
 531
 532         spin_lock(&gl->gl_lockref.lock);
 533 }
 534
 535 /**
 536  * find_first_holder - find the first "holder" gh
 537  * @gl: the glock
 538  */
 539
 540 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 541 {
 542         struct gfs2_holder *gh;
 543
 544         if (!list_empty(&gl->gl_holders)) {
 545                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 546                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 547                         return gh;
 548         }
 549         return NULL;
 550 }
 551
 552 /**
 553  * run_queue - do all outstanding tasks related to a glock
 554  * @gl: The glock in question
 555  * @nonblock: True if we must not block in run_queue
 556  *
 557  */
 558
 559 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 560 __releases(&gl->gl_lockref.lock)
 561 __acquires(&gl->gl_lockref.lock)
 562 {
 563         struct gfs2_holder *gh = NULL;
 564         int ret;
 565
 566         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 567                 return;
 568
 569         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 570
 571         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 572             gl->gl_demote_state != gl->gl_state) {
 573                 if (find_first_holder(gl))
 574                         goto out_unlock;
 575                 if (nonblock)
 576                         goto out_sched;
 577                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 578                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 579                 gl->gl_target = gl->gl_demote_state;
 580         } else {
 581                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 582                         gfs2_demote_wake(gl);
 583                 ret = do_promote(gl);
 584                 if (ret == 0)
 585                         goto out_unlock;
 586                 if (ret == 2)
 587                         goto out;
 588                 gh = find_first_waiter(gl);
 589                 gl->gl_target = gh->gh_state;
 590                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 591                         do_error(gl, 0); /* Fail queued try locks */
 592         }
 593         do_xmote(gl, gh, gl->gl_target);
 594 out:
 595         return;
 596
 597 out_sched:
 598         clear_bit(GLF_LOCK, &gl->gl_flags);
 599         smp_mb__after_atomic();
 600         gl->gl_lockref.count++;
 601         __gfs2_glock_queue_work(gl, 0);
 602         return;
 603
 604 out_unlock:
 605         clear_bit(GLF_LOCK, &gl->gl_flags);
 606         smp_mb__after_atomic();
 607         return;
 608 }
 609
 610 static void delete_work_func(struct work_struct *work)
 611 {
 612         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 613         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 614         struct inode *inode;
 615         u64 no_addr = gl->gl_name.ln_number;
 616
 617         /* If someone's using this glock to create a new dinode, the block must
 618            have been freed by another node, then re-used, in which case our
 619            iopen callback is too late after the fact. Ignore it. */
 620         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 621                 goto out;
 622
 623         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 624         if (inode && !IS_ERR(inode)) {
 625                 d_prune_aliases(inode);
 626                 iput(inode);
 627         }
 628 out:
 629         gfs2_glock_put(gl);
 630 }
 631
 632 static void glock_work_func(struct work_struct *work)
 633 {
 634         unsigned long delay = 0;
 635         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 636         unsigned int drop_refs = 1;
 637
 638         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 639                 finish_xmote(gl, gl->gl_reply);
 640                 drop_refs++;
 641         }
 642         spin_lock(&gl->gl_lockref.lock);
 643         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 644             gl->gl_state != LM_ST_UNLOCKED &&
 645             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 646                 unsigned long holdtime, now = jiffies;
 647
 648                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 649                 if (time_before(now, holdtime))
 650                         delay = holdtime - now;
 651
 652                 if (!delay) {
 653                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 654                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 655                 }
 656         }
 657         run_queue(gl, 0);
 658         if (delay) {
 659                 /* Keep one glock reference for the work we requeue. */
 660                 drop_refs--;
 661                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 662                         delay = 0;
 663                 __gfs2_glock_queue_work(gl, delay);
 664         }
 665
 666         /*
 667          * Drop the remaining glock references manually here. (Mind that
 668          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 669          * here as well.)
 670          */
 671         gl->gl_lockref.count -= drop_refs;
 672         if (!gl->gl_lockref.count) {
 673                 __gfs2_glock_put(gl);
 674                 return;
 675         }
 676         spin_unlock(&gl->gl_lockref.lock);
 677 }
 678
 679 /**
 680  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 681  * @sdp: The GFS2 superblock
 682  * @number: the lock number
 683  * @glops: The glock_operations to use
 684  * @create: If 0, don't create the glock if it doesn't exist
 685  * @glp: the glock is returned here
 686  *
 687  * This does not lock a glock, just finds/creates structures for one.
 688  *
 689  * Returns: errno
 690  */
 691
 692 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 693                    const struct gfs2_glock_operations *glops, int create,
 694                    struct gfs2_glock **glp)
 695 {
 696         struct super_block *s = sdp->sd_vfs;
 697         struct lm_lockname name = { .ln_number = number,
 698                                     .ln_type = glops->go_type,
 699                                     .ln_sbd = sdp };
 700         struct gfs2_glock *gl, *tmp;
 701         struct address_space *mapping;
 702         struct kmem_cache *cachep;
 703         int ret = 0;
 704
 705         rcu_read_lock();
 706         gl = rhashtable_lookup_fast(&gl_hash_table, &name, ht_parms);
 707         if (gl && !lockref_get_not_dead(&gl->gl_lockref))
 708                 gl = NULL;
 709         rcu_read_unlock();
 710
 711         *glp = gl;
 712         if (gl)
 713                 return 0;
 714         if (!create)
 715                 return -ENOENT;
 716
 717         if (glops->go_flags & GLOF_ASPACE)
 718                 cachep = gfs2_glock_aspace_cachep;
 719         else
 720                 cachep = gfs2_glock_cachep;
 721         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 722         if (!gl)
 723                 return -ENOMEM;
 724
 725         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 726
 727         if (glops->go_flags & GLOF_LVB) {
 728                 gl->gl_lksb.sb_lvbptr = kzalloc(GFS2_MIN_LVB_SIZE, GFP_NOFS);
 729                 if (!gl->gl_lksb.sb_lvbptr) {
 730                         kmem_cache_free(cachep, gl);
 731                         return -ENOMEM;
 732                 }
 733         }
 734
 735         atomic_inc(&sdp->sd_glock_disposal);
 736         gl->gl_node.next = NULL;
 737         gl->gl_flags = 0;
 738         gl->gl_name = name;
 739         gl->gl_lockref.count = 1;
 740         gl->gl_state = LM_ST_UNLOCKED;
 741         gl->gl_target = LM_ST_UNLOCKED;
 742         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 743         gl->gl_ops = glops;
 744         gl->gl_dstamp = 0;
 745         preempt_disable();
 746         /* We use the global stats to estimate the initial per-glock stats */
 747         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 748         preempt_enable();
 749         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 750         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 751         gl->gl_tchange = jiffies;
 752         gl->gl_object = NULL;
 753         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 754         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 755         INIT_WORK(&gl->gl_delete, delete_work_func);
 756
 757         mapping = gfs2_glock2aspace(gl);
 758         if (mapping) {
 759                 mapping->a_ops = &gfs2_meta_aops;
 760                 mapping->host = s->s_bdev->bd_inode;
 761                 mapping->flags = 0;
 762                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 763                 mapping->private_data = NULL;
 764                 mapping->writeback_index = 0;
 765         }
 766
 767 again:
 768         rcu_read_lock();
 769         tmp = rhashtable_lookup_get_insert_fast(&gl_hash_table, &gl->gl_node,
 770                                                 ht_parms);
 771         if (!tmp) {
 772                 *glp = gl;
 773                 goto out;
 774         }
 775         if (IS_ERR(tmp)) {
 776                 ret = PTR_ERR(tmp);
 777                 goto out_free;
 778         }
 779         if (lockref_get_not_dead(&tmp->gl_lockref)) {
 780                 *glp = tmp;
 781                 goto out_free;
 782         }
 783         rcu_read_unlock();
 784         cond_resched();
 785         goto again;
 786
 787 out_free:
 788         kfree(gl->gl_lksb.sb_lvbptr);
 789         kmem_cache_free(cachep, gl);
 790         atomic_dec(&sdp->sd_glock_disposal);
 791
 792 out:
 793         rcu_read_unlock();
 794         return ret;
 795 }
 796
 797 /**
 798  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 799  * @gl: the glock
 800  * @state: the state we're requesting
 801  * @flags: the modifier flags
 802  * @gh: the holder structure
 803  *
 804  */
 805
 806 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 807                       struct gfs2_holder *gh)
 808 {
 809         INIT_LIST_HEAD(&gh->gh_list);
 810         gh->gh_gl = gl;
 811         gh->gh_ip = _RET_IP_;
 812         gh->gh_owner_pid = get_pid(task_pid(current));
 813         gh->gh_state = state;
 814         gh->gh_flags = flags;
 815         gh->gh_error = 0;
 816         gh->gh_iflags = 0;
 817         gfs2_glock_hold(gl);
 818 }
 819
 820 /**
 821  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 822  * @state: the state we're requesting
 823  * @flags: the modifier flags
 824  * @gh: the holder structure
 825  *
 826  * Don't mess with the glock.
 827  *
 828  */
 829
 830 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
 831 {
 832         gh->gh_state = state;
 833         gh->gh_flags = flags;
 834         gh->gh_iflags = 0;
 835         gh->gh_ip = _RET_IP_;
 836         put_pid(gh->gh_owner_pid);
 837         gh->gh_owner_pid = get_pid(task_pid(current));
 838 }
 839
 840 /**
 841  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 842  * @gh: the holder structure
 843  *
 844  */
 845
 846 void gfs2_holder_uninit(struct gfs2_holder *gh)
 847 {
 848         put_pid(gh->gh_owner_pid);
 849         gfs2_glock_put(gh->gh_gl);
 850         gfs2_holder_mark_uninitialized(gh);
 851         gh->gh_ip = 0;
 852 }
 853
 854 /**
 855  * gfs2_glock_wait - wait on a glock acquisition
 856  * @gh: the glock holder
 857  *
 858  * Returns: 0 on success
 859  */
 860
 861 int gfs2_glock_wait(struct gfs2_holder *gh)
 862 {
 863         unsigned long time1 = jiffies;
 864
 865         might_sleep();
 866         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
 867         if (time_after(jiffies, time1 + HZ)) /* have we waited > a second? */
 868                 /* Lengthen the minimum hold time. */
 869                 gh->gh_gl->gl_hold_time = min(gh->gh_gl->gl_hold_time +
 870                                               GL_GLOCK_HOLD_INCR,
 871                                               GL_GLOCK_MAX_HOLD);
 872         return gh->gh_error;
 873 }
 874
 875 /**
 876  * handle_callback - process a demote request
 877  * @gl: the glock
 878  * @state: the state the caller wants us to change to
 879  *
 880  * There are only two requests that we are going to see in actual
 881  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 882  */
 883
 884 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 885                             unsigned long delay, bool remote)
 886 {
 887         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 888
 889         set_bit(bit, &gl->gl_flags);
 890         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 891                 gl->gl_demote_state = state;
 892                 gl->gl_demote_time = jiffies;
 893         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 894                         gl->gl_demote_state != state) {
 895                 gl->gl_demote_state = LM_ST_UNLOCKED;
 896         }
 897         if (gl->gl_ops->go_callback)
 898                 gl->gl_ops->go_callback(gl, remote);
 899         trace_gfs2_demote_rq(gl, remote);
 900 }
 901
 902 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 903 {
 904         struct va_format vaf;
 905         va_list args;
 906
 907         va_start(args, fmt);
 908
 909         if (seq) {
 910                 seq_vprintf(seq, fmt, args);
 911         } else {
 912                 vaf.fmt = fmt;
 913                 vaf.va = &args;
 914
 915                 pr_err("%pV", &vaf);
 916         }
 917
 918         va_end(args);
 919 }
 920
 921 /**
 922  * add_to_queue - Add a holder to the wait queue (but look for recursion)
 923  * @gh: the holder structure to add
 924  *
 925  * Eventually we should move the recursive locking trap to a
 926  * debugging option or something like that. This is the fast
 927  * path and needs to have the minimum number of distractions.
 928  *
 929  */
 930
 931 static inline void add_to_queue(struct gfs2_holder *gh)
 932 __releases(&gl->gl_lockref.lock)
 933 __acquires(&gl->gl_lockref.lock)
 934 {
 935         struct gfs2_glock *gl = gh->gh_gl;
 936         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 937         struct list_head *insert_pt = NULL;
 938         struct gfs2_holder *gh2;
 939         int try_futile = 0;
 940
 941         BUG_ON(gh->gh_owner_pid == NULL);
 942         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
 943                 BUG();
 944
 945         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 946                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 947                         try_futile = !may_grant(gl, gh);
 948                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 949                         goto fail;
 950         }
 951
 952         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
 953                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
 954                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
 955                         goto trap_recursive;
 956                 if (try_futile &&
 957                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 958 fail:
 959                         gh->gh_error = GLR_TRYFAILED;
 960                         gfs2_holder_wake(gh);
 961                         return;
 962                 }
 963                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
 964                         continue;
 965                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
 966                         insert_pt = &gh2->gh_list;
 967         }
 968         set_bit(GLF_QUEUED, &gl->gl_flags);
 969         trace_gfs2_glock_queue(gh, 1);
 970         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
 971         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
 972         if (likely(insert_pt == NULL)) {
 973                 list_add_tail(&gh->gh_list, &gl->gl_holders);
 974                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
 975                         goto do_cancel;
 976                 return;
 977         }
 978         list_add_tail(&gh->gh_list, insert_pt);
 979 do_cancel:
 980         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 981         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
 982                 spin_unlock(&gl->gl_lockref.lock);
 983                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
 984                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
 985                 spin_lock(&gl->gl_lockref.lock);
 986         }
 987         return;
 988
 989 trap_recursive:
 990         pr_err("original: %pSR\n", (void *)gh2->gh_ip);
 991         pr_err("pid: %d\n", pid_nr(gh2->gh_owner_pid));
 992         pr_err("lock type: %d req lock state : %d\n",
 993                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
 994         pr_err("new: %pSR\n", (void *)gh->gh_ip);
 995         pr_err("pid: %d\n", pid_nr(gh->gh_owner_pid));
 996         pr_err("lock type: %d req lock state : %d\n",
 997                gh->gh_gl->gl_name.ln_type, gh->gh_state);
 998         gfs2_dump_glock(NULL, gl);
 999         BUG();
1000 }
1001
1002 /**
1003  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1004  * @gh: the holder structure
1005  *
1006  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1007  *
1008  * Returns: 0, GLR_TRYFAILED, or errno on failure
1009  */
1010
1011 int gfs2_glock_nq(struct gfs2_holder *gh)
1012 {
1013         struct gfs2_glock *gl = gh->gh_gl;
1014         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1015         int error = 0;
1016
1017         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1018                 return -EIO;
1019
1020         if (test_bit(GLF_LRU, &gl->gl_flags))
1021                 gfs2_glock_remove_from_lru(gl);
1022
1023         spin_lock(&gl->gl_lockref.lock);
1024         add_to_queue(gh);
1025         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1026                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1027                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1028                 gl->gl_lockref.count++;
1029                 __gfs2_glock_queue_work(gl, 0);
1030         }
1031         run_queue(gl, 1);
1032         spin_unlock(&gl->gl_lockref.lock);
1033
1034         if (!(gh->gh_flags & GL_ASYNC))
1035                 error = gfs2_glock_wait(gh);
1036
1037         return error;
1038 }
1039
1040 /**
1041  * gfs2_glock_poll - poll to see if an async request has been completed
1042  * @gh: the holder
1043  *
1044  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1045  */
1046
1047 int gfs2_glock_poll(struct gfs2_holder *gh)
1048 {
1049         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1050 }
1051
1052 /**
1053  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1054  * @gh: the glock holder
1055  *
1056  */
1057
1058 void gfs2_glock_dq(struct gfs2_holder *gh)
1059 {
1060         struct gfs2_glock *gl = gh->gh_gl;
1061         const struct gfs2_glock_operations *glops = gl->gl_ops;
1062         unsigned delay = 0;
1063         int fast_path = 0;
1064
1065         spin_lock(&gl->gl_lockref.lock);
1066         if (gh->gh_flags & GL_NOCACHE)
1067                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1068
1069         list_del_init(&gh->gh_list);
1070         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1071         if (find_first_holder(gl) == NULL) {
1072                 if (glops->go_unlock) {
1073                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1074                         spin_unlock(&gl->gl_lockref.lock);
1075                         glops->go_unlock(gh);
1076                         spin_lock(&gl->gl_lockref.lock);
1077                         clear_bit(GLF_LOCK, &gl->gl_flags);
1078                 }
1079                 if (list_empty(&gl->gl_holders) &&
1080                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1081                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1082                         fast_path = 1;
1083         }
1084         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl) &&
1085             (glops->go_flags & GLOF_LRU))
1086                 gfs2_glock_add_to_lru(gl);
1087
1088         trace_gfs2_glock_queue(gh, 0);
1089         if (unlikely(!fast_path)) {
1090                 gl->gl_lockref.count++;
1091                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1092                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1093                     gl->gl_name.ln_type == LM_TYPE_INODE)
1094                         delay = gl->gl_hold_time;
1095                 __gfs2_glock_queue_work(gl, delay);
1096         }
1097         spin_unlock(&gl->gl_lockref.lock);
1098 }
1099
1100 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1101 {
1102         struct gfs2_glock *gl = gh->gh_gl;
1103         gfs2_glock_dq(gh);
1104         might_sleep();
1105         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1106 }
1107
1108 /**
1109  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1110  * @gh: the holder structure
1111  *
1112  */
1113
1114 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1115 {
1116         gfs2_glock_dq(gh);
1117         gfs2_holder_uninit(gh);
1118 }
1119
1120 /**
1121  * gfs2_glock_nq_num - acquire a glock based on lock number
1122  * @sdp: the filesystem
1123  * @number: the lock number
1124  * @glops: the glock operations for the type of glock
1125  * @state: the state to acquire the glock in
1126  * @flags: modifier flags for the acquisition
1127  * @gh: the struct gfs2_holder
1128  *
1129  * Returns: errno
1130  */
1131
1132 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1133                       const struct gfs2_glock_operations *glops,
1134                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1135 {
1136         struct gfs2_glock *gl;
1137         int error;
1138
1139         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1140         if (!error) {
1141                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1142                 gfs2_glock_put(gl);
1143         }
1144
1145         return error;
1146 }
1147
1148 /**
1149  * glock_compare - Compare two struct gfs2_glock structures for sorting
1150  * @arg_a: the first structure
1151  * @arg_b: the second structure
1152  *
1153  */
1154
1155 static int glock_compare(const void *arg_a, const void *arg_b)
1156 {
1157         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1158         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1159         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1160         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1161
1162         if (a->ln_number > b->ln_number)
1163                 return 1;
1164         if (a->ln_number < b->ln_number)
1165                 return -1;
1166         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1167         return 0;
1168 }
1169
1170 /**
1171  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1172  * @num_gh: the number of structures
1173  * @ghs: an array of struct gfs2_holder structures
1174  *
1175  * Returns: 0 on success (all glocks acquired),
1176  *          errno on failure (no glocks acquired)
1177  */
1178
1179 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1180                      struct gfs2_holder **p)
1181 {
1182         unsigned int x;
1183         int error = 0;
1184
1185         for (x = 0; x < num_gh; x++)
1186                 p[x] = &ghs[x];
1187
1188         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1189
1190         for (x = 0; x < num_gh; x++) {
1191                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1192
1193                 error = gfs2_glock_nq(p[x]);
1194                 if (error) {
1195                         while (x--)
1196                                 gfs2_glock_dq(p[x]);
1197                         break;
1198                 }
1199         }
1200
1201         return error;
1202 }
1203
1204 /**
1205  * gfs2_glock_nq_m - acquire multiple glocks
1206  * @num_gh: the number of structures
1207  * @ghs: an array of struct gfs2_holder structures
1208  *
1209  *
1210  * Returns: 0 on success (all glocks acquired),
1211  *          errno on failure (no glocks acquired)
1212  */
1213
1214 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1215 {
1216         struct gfs2_holder *tmp[4];
1217         struct gfs2_holder **pph = tmp;
1218         int error = 0;
1219
1220         switch(num_gh) {
1221         case 0:
1222                 return 0;
1223         case 1:
1224                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1225                 return gfs2_glock_nq(ghs);
1226         default:
1227                 if (num_gh <= 4)
1228                         break;
1229                 pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
1230                 if (!pph)
1231                         return -ENOMEM;
1232         }
1233
1234         error = nq_m_sync(num_gh, ghs, pph);
1235
1236         if (pph != tmp)
1237                 kfree(pph);
1238
1239         return error;
1240 }
1241
1242 /**
1243  * gfs2_glock_dq_m - release multiple glocks
1244  * @num_gh: the number of structures
1245  * @ghs: an array of struct gfs2_holder structures
1246  *
1247  */
1248
1249 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1250 {
1251         while (num_gh--)
1252                 gfs2_glock_dq(&ghs[num_gh]);
1253 }
1254
1255 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1256 {
1257         unsigned long delay = 0;
1258         unsigned long holdtime;
1259         unsigned long now = jiffies;
1260
1261         gfs2_glock_hold(gl);
1262         holdtime = gl->gl_tchange + gl->gl_hold_time;
1263         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1264             gl->gl_name.ln_type == LM_TYPE_INODE) {
1265                 if (time_before(now, holdtime))
1266                         delay = holdtime - now;
1267                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1268                         delay = gl->gl_hold_time;
1269         }
1270
1271         spin_lock(&gl->gl_lockref.lock);
1272         handle_callback(gl, state, delay, true);
1273         __gfs2_glock_queue_work(gl, delay);
1274         spin_unlock(&gl->gl_lockref.lock);
1275 }
1276
1277 /**
1278  * gfs2_should_freeze - Figure out if glock should be frozen
1279  * @gl: The glock in question
1280  *
1281  * Glocks are not frozen if (a) the result of the dlm operation is
1282  * an error, (b) the locking operation was an unlock operation or
1283  * (c) if there is a "noexp" flagged request anywhere in the queue
1284  *
1285  * Returns: 1 if freezing should occur, 0 otherwise
1286  */
1287
1288 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1289 {
1290         const struct gfs2_holder *gh;
1291
1292         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1293                 return 0;
1294         if (gl->gl_target == LM_ST_UNLOCKED)
1295                 return 0;
1296
1297         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1298                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1299                         continue;
1300                 if (LM_FLAG_NOEXP & gh->gh_flags)
1301                         return 0;
1302         }
1303
1304         return 1;
1305 }
1306
1307 /**
1308  * gfs2_glock_complete - Callback used by locking
1309  * @gl: Pointer to the glock
1310  * @ret: The return value from the dlm
1311  *
1312  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1313  * to use a bitfield shared with other glock state fields.
1314  */
1315
1316 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1317 {
1318         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1319
1320         spin_lock(&gl->gl_lockref.lock);
1321         gl->gl_reply = ret;
1322
1323         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1324                 if (gfs2_should_freeze(gl)) {
1325                         set_bit(GLF_FROZEN, &gl->gl_flags);
1326                         spin_unlock(&gl->gl_lockref.lock);
1327                         return;
1328                 }
1329         }
1330
1331         gl->gl_lockref.count++;
1332         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1333         __gfs2_glock_queue_work(gl, 0);
1334         spin_unlock(&gl->gl_lockref.lock);
1335 }
1336
1337 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1338 {
1339         struct gfs2_glock *gla, *glb;
1340
1341         gla = list_entry(a, struct gfs2_glock, gl_lru);
1342         glb = list_entry(b, struct gfs2_glock, gl_lru);
1343
1344         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1345                 return 1;
1346         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1347                 return -1;
1348
1349         return 0;
1350 }
1351
1352 /**
1353  * gfs2_dispose_glock_lru - Demote a list of glocks
1354  * @list: The list to dispose of
1355  *
1356  * Disposing of glocks may involve disk accesses, so that here we sort
1357  * the glocks by number (i.e. disk location of the inodes) so that if
1358  * there are any such accesses, they'll be sent in order (mostly).
1359  *
1360  * Must be called under the lru_lock, but may drop and retake this
1361  * lock. While the lru_lock is dropped, entries may vanish from the
1362  * list, but no new entries will appear on the list (since it is
1363  * private)
1364  */
1365
1366 static void gfs2_dispose_glock_lru(struct list_head *list)
1367 __releases(&lru_lock)
1368 __acquires(&lru_lock)
1369 {
1370         struct gfs2_glock *gl;
1371
1372         list_sort(NULL, list, glock_cmp);
1373
1374         while(!list_empty(list)) {
1375                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1376                 list_del_init(&gl->gl_lru);
1377                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1378 add_back_to_lru:
1379                         list_add(&gl->gl_lru, &lru_list);
1380                         atomic_inc(&lru_count);
1381                         continue;
1382                 }
1383                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1384                         spin_unlock(&gl->gl_lockref.lock);
1385                         goto add_back_to_lru;
1386                 }
1387                 clear_bit(GLF_LRU, &gl->gl_flags);
1388                 gl->gl_lockref.count++;
1389                 if (demote_ok(gl))
1390                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1391                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1392                 __gfs2_glock_queue_work(gl, 0);
1393                 spin_unlock(&gl->gl_lockref.lock);
1394                 cond_resched_lock(&lru_lock);
1395         }
1396 }
1397
1398 /**
1399  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1400  * @nr: The number of entries to scan
1401  *
1402  * This function selects the entries on the LRU which are able to
1403  * be demoted, and then kicks off the process by calling
1404  * gfs2_dispose_glock_lru() above.
1405  */
1406
1407 static long gfs2_scan_glock_lru(int nr)
1408 {
1409         struct gfs2_glock *gl;
1410         LIST_HEAD(skipped);
1411         LIST_HEAD(dispose);
1412         long freed = 0;
1413
1414         spin_lock(&lru_lock);
1415         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1416                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1417
1418                 /* Test for being demotable */
1419                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1420                         list_move(&gl->gl_lru, &dispose);
1421                         atomic_dec(&lru_count);
1422                         freed++;
1423                         continue;
1424                 }
1425
1426                 list_move(&gl->gl_lru, &skipped);
1427         }
1428         list_splice(&skipped, &lru_list);
1429         if (!list_empty(&dispose))
1430                 gfs2_dispose_glock_lru(&dispose);
1431         spin_unlock(&lru_lock);
1432
1433         return freed;
1434 }
1435
1436 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1437                                             struct shrink_control *sc)
1438 {
1439         if (!(sc->gfp_mask & __GFP_FS))
1440                 return SHRINK_STOP;
1441         return gfs2_scan_glock_lru(sc->nr_to_scan);
1442 }
1443
1444 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1445                                              struct shrink_control *sc)
1446 {
1447         return vfs_pressure_ratio(atomic_read(&lru_count));
1448 }
1449
1450 static struct shrinker glock_shrinker = {
1451         .seeks = DEFAULT_SEEKS,
1452         .count_objects = gfs2_glock_shrink_count,
1453         .scan_objects = gfs2_glock_shrink_scan,
1454 };
1455
1456 /**
1457  * examine_bucket - Call a function for glock in a hash bucket
1458  * @examiner: the function
1459  * @sdp: the filesystem
1460  * @bucket: the bucket
1461  *
1462  * Note that the function can be called multiple times on the same
1463  * object.  So the user must ensure that the function can cope with
1464  * that.
1465  */
1466
1467 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1468 {
1469         struct gfs2_glock *gl;
1470         struct rhashtable_iter iter;
1471
1472         rhashtable_walk_enter(&gl_hash_table, &iter);
1473
1474         do {
1475                 gl = ERR_PTR(rhashtable_walk_start(&iter));
1476                 if (gl)
1477                         continue;
1478
1479                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1480                         if ((gl->gl_name.ln_sbd == sdp) &&
1481                             lockref_get_not_dead(&gl->gl_lockref))
1482                                 examiner(gl);
1483
1484                 rhashtable_walk_stop(&iter);
1485         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1486
1487         rhashtable_walk_exit(&iter);
1488 }
1489
1490 /**
1491  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1492  * @gl: The glock to thaw
1493  *
1494  */
1495
1496 static void thaw_glock(struct gfs2_glock *gl)
1497 {
1498         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1499                 gfs2_glock_put(gl);
1500                 return;
1501         }
1502         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1503         gfs2_glock_queue_work(gl, 0);
1504 }
1505
1506 /**
1507  * clear_glock - look at a glock and see if we can free it from glock cache
1508  * @gl: the glock to look at
1509  *
1510  */
1511
1512 static void clear_glock(struct gfs2_glock *gl)
1513 {
1514         gfs2_glock_remove_from_lru(gl);
1515
1516         spin_lock(&gl->gl_lockref.lock);
1517         if (gl->gl_state != LM_ST_UNLOCKED)
1518                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1519         __gfs2_glock_queue_work(gl, 0);
1520         spin_unlock(&gl->gl_lockref.lock);
1521 }
1522
1523 /**
1524  * gfs2_glock_thaw - Thaw any frozen glocks
1525  * @sdp: The super block
1526  *
1527  */
1528
1529 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1530 {
1531         glock_hash_walk(thaw_glock, sdp);
1532 }
1533
1534 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1535 {
1536         spin_lock(&gl->gl_lockref.lock);
1537         gfs2_dump_glock(seq, gl);
1538         spin_unlock(&gl->gl_lockref.lock);
1539 }
1540
1541 static void dump_glock_func(struct gfs2_glock *gl)
1542 {
1543         dump_glock(NULL, gl);
1544 }
1545
1546 /**
1547  * gfs2_gl_hash_clear - Empty out the glock hash table
1548  * @sdp: the filesystem
1549  * @wait: wait until it's all gone
1550  *
1551  * Called when unmounting the filesystem.
1552  */
1553
1554 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1555 {
1556         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1557         flush_workqueue(glock_workqueue);
1558         glock_hash_walk(clear_glock, sdp);
1559         flush_workqueue(glock_workqueue);
1560         wait_event_timeout(sdp->sd_glock_wait,
1561                            atomic_read(&sdp->sd_glock_disposal) == 0,
1562                            HZ * 600);
1563         glock_hash_walk(dump_glock_func, sdp);
1564 }
1565
1566 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1567 {
1568         struct gfs2_glock *gl = ip->i_gl;
1569         int ret;
1570
1571         ret = gfs2_truncatei_resume(ip);
1572         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1573
1574         spin_lock(&gl->gl_lockref.lock);
1575         clear_bit(GLF_LOCK, &gl->gl_flags);
1576         run_queue(gl, 1);
1577         spin_unlock(&gl->gl_lockref.lock);
1578 }
1579
1580 static const char *state2str(unsigned state)
1581 {
1582         switch(state) {
1583         case LM_ST_UNLOCKED:
1584                 return "UN";
1585         case LM_ST_SHARED:
1586                 return "SH";
1587         case LM_ST_DEFERRED:
1588                 return "DF";
1589         case LM_ST_EXCLUSIVE:
1590                 return "EX";
1591         }
1592         return "??";
1593 }
1594
1595 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1596 {
1597         char *p = buf;
1598         if (flags & LM_FLAG_TRY)
1599                 *p++ = 't';
1600         if (flags & LM_FLAG_TRY_1CB)
1601                 *p++ = 'T';
1602         if (flags & LM_FLAG_NOEXP)
1603                 *p++ = 'e';
1604         if (flags & LM_FLAG_ANY)
1605                 *p++ = 'A';
1606         if (flags & LM_FLAG_PRIORITY)
1607                 *p++ = 'p';
1608         if (flags & GL_ASYNC)
1609                 *p++ = 'a';
1610         if (flags & GL_EXACT)
1611                 *p++ = 'E';
1612         if (flags & GL_NOCACHE)
1613                 *p++ = 'c';
1614         if (test_bit(HIF_HOLDER, &iflags))
1615                 *p++ = 'H';
1616         if (test_bit(HIF_WAIT, &iflags))
1617                 *p++ = 'W';
1618         if (test_bit(HIF_FIRST, &iflags))
1619                 *p++ = 'F';
1620         *p = 0;
1621         return buf;
1622 }
1623
1624 /**
1625  * dump_holder - print information about a glock holder
1626  * @seq: the seq_file struct
1627  * @gh: the glock holder
1628  *
1629  */
1630
1631 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1632 {
1633         struct task_struct *gh_owner = NULL;
1634         char flags_buf[32];
1635
1636         rcu_read_lock();
1637         if (gh->gh_owner_pid)
1638                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1639         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1640                        state2str(gh->gh_state),
1641                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1642                        gh->gh_error,
1643                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1644                        gh_owner ? gh_owner->comm : "(ended)",
1645                        (void *)gh->gh_ip);
1646         rcu_read_unlock();
1647 }
1648
1649 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1650 {
1651         const unsigned long *gflags = &gl->gl_flags;
1652         char *p = buf;
1653
1654         if (test_bit(GLF_LOCK, gflags))
1655                 *p++ = 'l';
1656         if (test_bit(GLF_DEMOTE, gflags))
1657                 *p++ = 'D';
1658         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1659                 *p++ = 'd';
1660         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1661                 *p++ = 'p';
1662         if (test_bit(GLF_DIRTY, gflags))
1663                 *p++ = 'y';
1664         if (test_bit(GLF_LFLUSH, gflags))
1665                 *p++ = 'f';
1666         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1667                 *p++ = 'i';
1668         if (test_bit(GLF_REPLY_PENDING, gflags))
1669                 *p++ = 'r';
1670         if (test_bit(GLF_INITIAL, gflags))
1671                 *p++ = 'I';
1672         if (test_bit(GLF_FROZEN, gflags))
1673                 *p++ = 'F';
1674         if (test_bit(GLF_QUEUED, gflags))
1675                 *p++ = 'q';
1676         if (test_bit(GLF_LRU, gflags))
1677                 *p++ = 'L';
1678         if (gl->gl_object)
1679                 *p++ = 'o';
1680         if (test_bit(GLF_BLOCKING, gflags))
1681                 *p++ = 'b';
1682         *p = 0;
1683         return buf;
1684 }
1685
1686 /**
1687  * gfs2_dump_glock - print information about a glock
1688  * @seq: The seq_file struct
1689  * @gl: the glock
1690  *
1691  * The file format is as follows:
1692  * One line per object, capital letters are used to indicate objects
1693  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1694  * other objects are indented by a single space and follow the glock to
1695  * which they are related. Fields are indicated by lower case letters
1696  * followed by a colon and the field value, except for strings which are in
1697  * [] so that its possible to see if they are composed of spaces for
1698  * example. The field's are n = number (id of the object), f = flags,
1699  * t = type, s = state, r = refcount, e = error, p = pid.
1700  *
1701  */
1702
1703 void gfs2_dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1704 {
1705         const struct gfs2_glock_operations *glops = gl->gl_ops;
1706         unsigned long long dtime;
1707         const struct gfs2_holder *gh;
1708         char gflags_buf[32];
1709
1710         dtime = jiffies - gl->gl_demote_time;
1711         dtime *= 1000000/HZ; /* demote time in uSec */
1712         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1713                 dtime = 0;
1714         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d v:%d r:%d m:%ld\n",
1715                   state2str(gl->gl_state),
1716                   gl->gl_name.ln_type,
1717                   (unsigned long long)gl->gl_name.ln_number,
1718                   gflags2str(gflags_buf, gl),
1719                   state2str(gl->gl_target),
1720                   state2str(gl->gl_demote_state), dtime,
1721                   atomic_read(&gl->gl_ail_count),
1722                   atomic_read(&gl->gl_revokes),
1723                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1724
1725         list_for_each_entry(gh, &gl->gl_holders, gh_list)
1726                 dump_holder(seq, gh);
1727
1728         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1729                 glops->go_dump(seq, gl);
1730 }
1731
1732 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
1733 {
1734         struct gfs2_glock *gl = iter_ptr;
1735
1736         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
1737                    gl->gl_name.ln_type,
1738                    (unsigned long long)gl->gl_name.ln_number,
1739                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
1740                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
1741                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
1742                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
1743                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
1744                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
1745                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
1746                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
1747         return 0;
1748 }
1749
1750 static const char *gfs2_gltype[] = {
1751         "type",
1752         "reserved",
1753         "nondisk",
1754         "inode",
1755         "rgrp",
1756         "meta",
1757         "iopen",
1758         "flock",
1759         "plock",
1760         "quota",
1761         "journal",
1762 };
1763
1764 static const char *gfs2_stype[] = {
1765         [GFS2_LKS_SRTT]         = "srtt",
1766         [GFS2_LKS_SRTTVAR]      = "srttvar",
1767         [GFS2_LKS_SRTTB]        = "srttb",
1768         [GFS2_LKS_SRTTVARB]     = "srttvarb",
1769         [GFS2_LKS_SIRT]         = "sirt",
1770         [GFS2_LKS_SIRTVAR]      = "sirtvar",
1771         [GFS2_LKS_DCOUNT]       = "dlm",
1772         [GFS2_LKS_QCOUNT]       = "queue",
1773 };
1774
1775 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
1776
1777 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
1778 {
1779         struct gfs2_sbd *sdp = seq->private;
1780         loff_t pos = *(loff_t *)iter_ptr;
1781         unsigned index = pos >> 3;
1782         unsigned subindex = pos & 0x07;
1783         int i;
1784
1785         if (index == 0 && subindex != 0)
1786                 return 0;
1787
1788         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
1789                    (index == 0) ? "cpu": gfs2_stype[subindex]);
1790
1791         for_each_possible_cpu(i) {
1792                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
1793
1794                 if (index == 0)
1795                         seq_printf(seq, " %15u", i);
1796                 else
1797                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
1798                                    lkstats[index - 1].stats[subindex]);
1799         }
1800         seq_putc(seq, '\n');
1801         return 0;
1802 }
1803
1804 int __init gfs2_glock_init(void)
1805 {
1806         int ret;
1807
1808         ret = rhashtable_init(&gl_hash_table, &ht_parms);
1809         if (ret < 0)
1810                 return ret;
1811
1812         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1813                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1814         if (!glock_workqueue) {
1815                 rhashtable_destroy(&gl_hash_table);
1816                 return -ENOMEM;
1817         }
1818         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1819                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1820                                                 0);
1821         if (!gfs2_delete_workqueue) {
1822                 destroy_workqueue(glock_workqueue);
1823                 rhashtable_destroy(&gl_hash_table);
1824                 return -ENOMEM;
1825         }
1826
1827         ret = register_shrinker(&glock_shrinker);
1828         if (ret) {
1829                 destroy_workqueue(gfs2_delete_workqueue);
1830                 destroy_workqueue(glock_workqueue);
1831                 rhashtable_destroy(&gl_hash_table);
1832                 return ret;
1833         }
1834
1835         return 0;
1836 }
1837
1838 void gfs2_glock_exit(void)
1839 {
1840         unregister_shrinker(&glock_shrinker);
1841         rhashtable_destroy(&gl_hash_table);
1842         destroy_workqueue(glock_workqueue);
1843         destroy_workqueue(gfs2_delete_workqueue);
1844 }
1845
1846 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi)
1847 {
1848         while ((gi->gl = rhashtable_walk_next(&gi->hti))) {
1849                 if (IS_ERR(gi->gl)) {
1850                         if (PTR_ERR(gi->gl) == -EAGAIN)
1851                                 continue;
1852                         gi->gl = NULL;
1853                         return;
1854                 }
1855                 /* Skip entries for other sb and dead entries */
1856                 if (gi->sdp == gi->gl->gl_name.ln_sbd &&
1857                     !__lockref_is_dead(&gi->gl->gl_lockref))
1858                         return;
1859         }
1860 }
1861
1862 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1863 {
1864         struct gfs2_glock_iter *gi = seq->private;
1865         loff_t n = *pos;
1866         int ret;
1867
1868         if (gi->last_pos <= *pos)
1869                 n = (*pos - gi->last_pos);
1870
1871         ret = rhashtable_walk_start(&gi->hti);
1872         if (ret)
1873                 return NULL;
1874
1875         do {
1876                 gfs2_glock_iter_next(gi);
1877         } while (gi->gl && n--);
1878
1879         gi->last_pos = *pos;
1880         return gi->gl;
1881 }
1882
1883 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1884                                  loff_t *pos)
1885 {
1886         struct gfs2_glock_iter *gi = seq->private;
1887
1888         (*pos)++;
1889         gi->last_pos = *pos;
1890         gfs2_glock_iter_next(gi);
1891         return gi->gl;
1892 }
1893
1894 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1895 {
1896         struct gfs2_glock_iter *gi = seq->private;
1897
1898         gi->gl = NULL;
1899         rhashtable_walk_stop(&gi->hti);
1900 }
1901
1902 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
1903 {
1904         dump_glock(seq, iter_ptr);
1905         return 0;
1906 }
1907
1908 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
1909 {
1910         preempt_disable();
1911         if (*pos >= GFS2_NR_SBSTATS)
1912                 return NULL;
1913         return pos;
1914 }
1915
1916 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
1917                                    loff_t *pos)
1918 {
1919         (*pos)++;
1920         if (*pos >= GFS2_NR_SBSTATS)
1921                 return NULL;
1922         return pos;
1923 }
1924
1925 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
1926 {
1927         preempt_enable();
1928 }
1929
1930 static const struct seq_operations gfs2_glock_seq_ops = {
1931         .start = gfs2_glock_seq_start,
1932         .next  = gfs2_glock_seq_next,
1933         .stop  = gfs2_glock_seq_stop,
1934         .show  = gfs2_glock_seq_show,
1935 };
1936
1937 static const struct seq_operations gfs2_glstats_seq_ops = {
1938         .start = gfs2_glock_seq_start,
1939         .next  = gfs2_glock_seq_next,
1940         .stop  = gfs2_glock_seq_stop,
1941         .show  = gfs2_glstats_seq_show,
1942 };
1943
1944 static const struct seq_operations gfs2_sbstats_seq_ops = {
1945         .start = gfs2_sbstats_seq_start,
1946         .next  = gfs2_sbstats_seq_next,
1947         .stop  = gfs2_sbstats_seq_stop,
1948         .show  = gfs2_sbstats_seq_show,
1949 };
1950
1951 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
1952
1953 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
1954                               const struct seq_operations *ops)
1955 {
1956         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
1957         if (ret == 0) {
1958                 struct seq_file *seq = file->private_data;
1959                 struct gfs2_glock_iter *gi = seq->private;
1960
1961                 gi->sdp = inode->i_private;
1962                 gi->last_pos = 0;
1963                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
1964                 if (seq->buf)
1965                         seq->size = GFS2_SEQ_GOODSIZE;
1966                 gi->gl = NULL;
1967                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
1968         }
1969         return ret;
1970 }
1971
1972 static int gfs2_glocks_open(struct inode *inode, struct file *file)
1973 {
1974         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
1975 }
1976
1977 static int gfs2_glocks_release(struct inode *inode, struct file *file)
1978 {
1979         struct seq_file *seq = file->private_data;
1980         struct gfs2_glock_iter *gi = seq->private;
1981
1982         gi->gl = NULL;
1983         rhashtable_walk_exit(&gi->hti);
1984         return seq_release_private(inode, file);
1985 }
1986
1987 static int gfs2_glstats_open(struct inode *inode, struct file *file)
1988 {
1989         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
1990 }
1991
1992 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
1993 {
1994         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
1995         if (ret == 0) {
1996                 struct seq_file *seq = file->private_data;
1997                 seq->private = inode->i_private;  /* sdp */
1998         }
1999         return ret;
2000 }
2001
2002 static const struct file_operations gfs2_glocks_fops = {
2003         .owner   = THIS_MODULE,
2004         .open    = gfs2_glocks_open,
2005         .read    = seq_read,
2006         .llseek  = seq_lseek,
2007         .release = gfs2_glocks_release,
2008 };
2009
2010 static const struct file_operations gfs2_glstats_fops = {
2011         .owner   = THIS_MODULE,
2012         .open    = gfs2_glstats_open,
2013         .read    = seq_read,
2014         .llseek  = seq_lseek,
2015         .release = gfs2_glocks_release,
2016 };
2017
2018 static const struct file_operations gfs2_sbstats_fops = {
2019         .owner   = THIS_MODULE,
2020         .open    = gfs2_sbstats_open,
2021         .read    = seq_read,
2022         .llseek  = seq_lseek,
2023         .release = seq_release,
2024 };
2025
2026 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2027 {
2028         struct dentry *dent;
2029
2030         dent = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2031         if (IS_ERR_OR_NULL(dent))
2032                 goto fail;
2033         sdp->debugfs_dir = dent;
2034
2035         dent = debugfs_create_file("glocks",
2036                                    S_IFREG | S_IRUGO,
2037                                    sdp->debugfs_dir, sdp,
2038                                    &gfs2_glocks_fops);
2039         if (IS_ERR_OR_NULL(dent))
2040                 goto fail;
2041         sdp->debugfs_dentry_glocks = dent;
2042
2043         dent = debugfs_create_file("glstats",
2044                                    S_IFREG | S_IRUGO,
2045                                    sdp->debugfs_dir, sdp,
2046                                    &gfs2_glstats_fops);
2047         if (IS_ERR_OR_NULL(dent))
2048                 goto fail;
2049         sdp->debugfs_dentry_glstats = dent;
2050
2051         dent = debugfs_create_file("sbstats",
2052                                    S_IFREG | S_IRUGO,
2053                                    sdp->debugfs_dir, sdp,
2054                                    &gfs2_sbstats_fops);
2055         if (IS_ERR_OR_NULL(dent))
2056                 goto fail;
2057         sdp->debugfs_dentry_sbstats = dent;
2058
2059         return 0;
2060 fail:
2061         gfs2_delete_debugfs_file(sdp);
2062         return dent ? PTR_ERR(dent) : -ENOMEM;
2063 }
2064
2065 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2066 {
2067         if (sdp->debugfs_dir) {
2068                 if (sdp->debugfs_dentry_glocks) {
2069                         debugfs_remove(sdp->debugfs_dentry_glocks);
2070                         sdp->debugfs_dentry_glocks = NULL;
2071                 }
2072                 if (sdp->debugfs_dentry_glstats) {
2073                         debugfs_remove(sdp->debugfs_dentry_glstats);
2074                         sdp->debugfs_dentry_glstats = NULL;
2075                 }
2076                 if (sdp->debugfs_dentry_sbstats) {
2077                         debugfs_remove(sdp->debugfs_dentry_sbstats);
2078                         sdp->debugfs_dentry_sbstats = NULL;
2079                 }
2080                 debugfs_remove(sdp->debugfs_dir);
2081                 sdp->debugfs_dir = NULL;
2082         }
2083 }
2084
2085 int gfs2_register_debugfs(void)
2086 {
2087         gfs2_root = debugfs_create_dir("gfs2", NULL);
2088         if (IS_ERR(gfs2_root))
2089                 return PTR_ERR(gfs2_root);
2090         return gfs2_root ? 0 : -ENOMEM;
2091 }
2092
2093 void gfs2_unregister_debugfs(void)
2094 {
2095         debugfs_remove(gfs2_root);
2096         gfs2_root = NULL;
2097 }