gfs2/glock.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   5  */
   6
   7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/sched.h>
  10 #include <linux/slab.h>
  11 #include <linux/spinlock.h>
  12 #include <linux/buffer_head.h>
  13 #include <linux/delay.h>
  14 #include <linux/sort.h>
  15 #include <linux/hash.h>
  16 #include <linux/jhash.h>
  17 #include <linux/kallsyms.h>
  18 #include <linux/gfs2_ondisk.h>
  19 #include <linux/list.h>
  20 #include <linux/wait.h>
  21 #include <linux/module.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/debugfs.h>
  25 #include <linux/kthread.h>
  26 #include <linux/freezer.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/jiffies.h>
  29 #include <linux/rcupdate.h>
  30 #include <linux/rculist_bl.h>
  31 #include <linux/bit_spinlock.h>
  32 #include <linux/percpu.h>
  33 #include <linux/list_sort.h>
  34 #include <linux/lockref.h>
  35 #include <linux/rhashtable.h>
  36
  37 #include "gfs2.h"
  38 #include "incore.h"
  39 #include "glock.h"
  40 #include "glops.h"
  41 #include "inode.h"
  42 #include "lops.h"
  43 #include "meta_io.h"
  44 #include "quota.h"
  45 #include "super.h"
  46 #include "util.h"
  47 #include "bmap.h"
  48 #define CREATE_TRACE_POINTS
  49 #include "trace_gfs2.h"
  50
  51 struct gfs2_glock_iter {
  52         struct gfs2_sbd *sdp;           /* incore superblock           */
  53         struct rhashtable_iter hti;     /* rhashtable iterator         */
  54         struct gfs2_glock *gl;          /* current glock struct        */
  55         loff_t last_pos;                /* last position               */
  56 };
  57
  58 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  59
  60 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  61
  62 static struct dentry *gfs2_root;
  63 static struct workqueue_struct *glock_workqueue;
  64 struct workqueue_struct *gfs2_delete_workqueue;
  65 static LIST_HEAD(lru_list);
  66 static atomic_t lru_count = ATOMIC_INIT(0);
  67 static DEFINE_SPINLOCK(lru_lock);
  68
  69 #define GFS2_GL_HASH_SHIFT      15
  70 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
  71
  72 static const struct rhashtable_params ht_parms = {
  73         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
  74         .key_len = offsetofend(struct lm_lockname, ln_type),
  75         .key_offset = offsetof(struct gfs2_glock, gl_name),
  76         .head_offset = offsetof(struct gfs2_glock, gl_node),
  77 };
  78
  79 static struct rhashtable gl_hash_table;
  80
  81 #define GLOCK_WAIT_TABLE_BITS 12
  82 #define GLOCK_WAIT_TABLE_SIZE (1 << GLOCK_WAIT_TABLE_BITS)
  83 static wait_queue_head_t glock_wait_table[GLOCK_WAIT_TABLE_SIZE] __cacheline_aligned;
  84
  85 struct wait_glock_queue {
  86         struct lm_lockname *name;
  87         wait_queue_entry_t wait;
  88 };
  89
  90 static int glock_wake_function(wait_queue_entry_t *wait, unsigned int mode,
  91                                int sync, void *key)
  92 {
  93         struct wait_glock_queue *wait_glock =
  94                 container_of(wait, struct wait_glock_queue, wait);
  95         struct lm_lockname *wait_name = wait_glock->name;
  96         struct lm_lockname *wake_name = key;
  97
  98         if (wake_name->ln_sbd != wait_name->ln_sbd ||
  99             wake_name->ln_number != wait_name->ln_number ||
 100             wake_name->ln_type != wait_name->ln_type)
 101                 return 0;
 102         return autoremove_wake_function(wait, mode, sync, key);
 103 }
 104
 105 static wait_queue_head_t *glock_waitqueue(struct lm_lockname *name)
 106 {
 107         u32 hash = jhash2((u32 *)name, ht_parms.key_len / 4, 0);
 108
 109         return glock_wait_table + hash_32(hash, GLOCK_WAIT_TABLE_BITS);
 110 }
 111
 112 /**
 113  * wake_up_glock  -  Wake up waiters on a glock
 114  * @gl: the glock
 115  */
 116 static void wake_up_glock(struct gfs2_glock *gl)
 117 {
 118         wait_queue_head_t *wq = glock_waitqueue(&gl->gl_name);
 119
 120         if (waitqueue_active(wq))
 121                 __wake_up(wq, TASK_NORMAL, 1, &gl->gl_name);
 122 }
 123
 124 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 125 {
 126         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 127
 128         kfree(gl->gl_lksb.sb_lvbptr);
 129         if (gl->gl_ops->go_flags & GLOF_ASPACE)
 130                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 131         else
 132                 kmem_cache_free(gfs2_glock_cachep, gl);
 133 }
 134
 135 /**
 136  * glock_blocked_by_withdraw - determine if we can still use a glock
 137  * @gl: the glock
 138  *
 139  * We need to allow some glocks to be enqueued, dequeued, promoted, and demoted
 140  * when we're withdrawn. For example, to maintain metadata integrity, we should
 141  * disallow the use of inode and rgrp glocks when withdrawn. Other glocks, like
 142  * iopen or the transaction glocks may be safely used because none of their
 143  * metadata goes through the journal. So in general, we should disallow all
 144  * glocks that are journaled, and allow all the others. One exception is:
 145  * we need to allow our active journal to be promoted and demoted so others
 146  * may recover it and we can reacquire it when they're done.
 147  */
 148 static bool glock_blocked_by_withdraw(struct gfs2_glock *gl)
 149 {
 150         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 151
 152         if (likely(!gfs2_withdrawn(sdp)))
 153                 return false;
 154         if (gl->gl_ops->go_flags & GLOF_NONDISK)
 155                 return false;
 156         if (!sdp->sd_jdesc ||
 157             gl->gl_name.ln_number == sdp->sd_jdesc->jd_no_addr)
 158                 return false;
 159         return true;
 160 }
 161
 162 void gfs2_glock_free(struct gfs2_glock *gl)
 163 {
 164         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 165
 166         gfs2_glock_assert_withdraw(gl, atomic_read(&gl->gl_revokes) == 0);
 167         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 168         smp_mb();
 169         wake_up_glock(gl);
 170         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 171         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 172                 wake_up(&sdp->sd_glock_wait);
 173 }
 174
 175 /**
 176  * gfs2_glock_hold() - increment reference count on glock
 177  * @gl: The glock to hold
 178  *
 179  */
 180
 181 void gfs2_glock_hold(struct gfs2_glock *gl)
 182 {
 183         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 184         lockref_get(&gl->gl_lockref);
 185 }
 186
 187 /**
 188  * demote_ok - Check to see if it's ok to unlock a glock
 189  * @gl: the glock
 190  *
 191  * Returns: 1 if it's ok
 192  */
 193
 194 static int demote_ok(const struct gfs2_glock *gl)
 195 {
 196         const struct gfs2_glock_operations *glops = gl->gl_ops;
 197
 198         if (gl->gl_state == LM_ST_UNLOCKED)
 199                 return 0;
 200         if (!list_empty(&gl->gl_holders))
 201                 return 0;
 202         if (glops->go_demote_ok)
 203                 return glops->go_demote_ok(gl);
 204         return 1;
 205 }
 206
 207
 208 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 209 {
 210         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 211                 return;
 212
 213         spin_lock(&lru_lock);
 214
 215         list_del(&gl->gl_lru);
 216         list_add_tail(&gl->gl_lru, &lru_list);
 217
 218         if (!test_bit(GLF_LRU, &gl->gl_flags)) {
 219                 set_bit(GLF_LRU, &gl->gl_flags);
 220                 atomic_inc(&lru_count);
 221         }
 222
 223         spin_unlock(&lru_lock);
 224 }
 225
 226 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 227 {
 228         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 229                 return;
 230
 231         spin_lock(&lru_lock);
 232         if (test_bit(GLF_LRU, &gl->gl_flags)) {
 233                 list_del_init(&gl->gl_lru);
 234                 atomic_dec(&lru_count);
 235                 clear_bit(GLF_LRU, &gl->gl_flags);
 236         }
 237         spin_unlock(&lru_lock);
 238 }
 239
 240 /*
 241  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 242  * work queue.
 243  */
 244 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 245         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 246                 /*
 247                  * We are holding the lockref spinlock, and the work was still
 248                  * queued above.  The queued work (glock_work_func) takes that
 249                  * spinlock before dropping its glock reference(s), so it
 250                  * cannot have dropped them in the meantime.
 251                  */
 252                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 253                 gl->gl_lockref.count--;
 254         }
 255 }
 256
 257 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 258         spin_lock(&gl->gl_lockref.lock);
 259         __gfs2_glock_queue_work(gl, delay);
 260         spin_unlock(&gl->gl_lockref.lock);
 261 }
 262
 263 static void __gfs2_glock_put(struct gfs2_glock *gl)
 264 {
 265         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 266         struct address_space *mapping = gfs2_glock2aspace(gl);
 267
 268         lockref_mark_dead(&gl->gl_lockref);
 269
 270         gfs2_glock_remove_from_lru(gl);
 271         spin_unlock(&gl->gl_lockref.lock);
 272         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 273         if (mapping) {
 274                 truncate_inode_pages_final(mapping);
 275                 if (!gfs2_withdrawn(sdp))
 276                         GLOCK_BUG_ON(gl, mapping->nrpages ||
 277                                      mapping->nrexceptional);
 278         }
 279         trace_gfs2_glock_put(gl);
 280         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 281 }
 282
 283 /*
 284  * Cause the glock to be put in work queue context.
 285  */
 286 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 287 {
 288         gfs2_glock_queue_work(gl, 0);
 289 }
 290
 291 /**
 292  * gfs2_glock_put() - Decrement reference count on glock
 293  * @gl: The glock to put
 294  *
 295  */
 296
 297 void gfs2_glock_put(struct gfs2_glock *gl)
 298 {
 299         if (lockref_put_or_lock(&gl->gl_lockref))
 300                 return;
 301
 302         __gfs2_glock_put(gl);
 303 }
 304
 305 /**
 306  * may_grant - check if its ok to grant a new lock
 307  * @gl: The glock
 308  * @gh: The lock request which we wish to grant
 309  *
 310  * Returns: true if its ok to grant the lock
 311  */
 312
 313 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 314 {
 315         const struct gfs2_holder *gh_head = list_first_entry(&gl->gl_holders, const struct gfs2_holder, gh_list);
 316         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 317              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 318                 return 0;
 319         if (gl->gl_state == gh->gh_state)
 320                 return 1;
 321         if (gh->gh_flags & GL_EXACT)
 322                 return 0;
 323         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 324                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 325                         return 1;
 326                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 327                         return 1;
 328         }
 329         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 330                 return 1;
 331         return 0;
 332 }
 333
 334 static void gfs2_holder_wake(struct gfs2_holder *gh)
 335 {
 336         clear_bit(HIF_WAIT, &gh->gh_iflags);
 337         smp_mb__after_atomic();
 338         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 339         if (gh->gh_flags & GL_ASYNC) {
 340                 struct gfs2_sbd *sdp = gh->gh_gl->gl_name.ln_sbd;
 341
 342                 wake_up(&sdp->sd_async_glock_wait);
 343         }
 344 }
 345
 346 /**
 347  * do_error - Something unexpected has happened during a lock request
 348  *
 349  */
 350
 351 static void do_error(struct gfs2_glock *gl, const int ret)
 352 {
 353         struct gfs2_holder *gh, *tmp;
 354
 355         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 356                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 357                         continue;
 358                 if (ret & LM_OUT_ERROR)
 359                         gh->gh_error = -EIO;
 360                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 361                         gh->gh_error = GLR_TRYFAILED;
 362                 else
 363                         continue;
 364                 list_del_init(&gh->gh_list);
 365                 trace_gfs2_glock_queue(gh, 0);
 366                 gfs2_holder_wake(gh);
 367         }
 368 }
 369
 370 /**
 371  * do_promote - promote as many requests as possible on the current queue
 372  * @gl: The glock
 373  *
 374  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 375  *          if a type specific operation is underway.
 376  */
 377
 378 static int do_promote(struct gfs2_glock *gl)
 379 __releases(&gl->gl_lockref.lock)
 380 __acquires(&gl->gl_lockref.lock)
 381 {
 382         const struct gfs2_glock_operations *glops = gl->gl_ops;
 383         struct gfs2_holder *gh, *tmp;
 384         int ret;
 385
 386 restart:
 387         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 388                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 389                         continue;
 390                 if (may_grant(gl, gh)) {
 391                         if (gh->gh_list.prev == &gl->gl_holders &&
 392                             glops->go_lock) {
 393                                 spin_unlock(&gl->gl_lockref.lock);
 394                                 /* FIXME: eliminate this eventually */
 395                                 ret = glops->go_lock(gh);
 396                                 spin_lock(&gl->gl_lockref.lock);
 397                                 if (ret) {
 398                                         if (ret == 1)
 399                                                 return 2;
 400                                         gh->gh_error = ret;
 401                                         list_del_init(&gh->gh_list);
 402                                         trace_gfs2_glock_queue(gh, 0);
 403                                         gfs2_holder_wake(gh);
 404                                         goto restart;
 405                                 }
 406                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 407                                 trace_gfs2_promote(gh, 1);
 408                                 gfs2_holder_wake(gh);
 409                                 goto restart;
 410                         }
 411                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 412                         trace_gfs2_promote(gh, 0);
 413                         gfs2_holder_wake(gh);
 414                         continue;
 415                 }
 416                 if (gh->gh_list.prev == &gl->gl_holders)
 417                         return 1;
 418                 do_error(gl, 0);
 419                 break;
 420         }
 421         return 0;
 422 }
 423
 424 /**
 425  * find_first_waiter - find the first gh that's waiting for the glock
 426  * @gl: the glock
 427  */
 428
 429 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 430 {
 431         struct gfs2_holder *gh;
 432
 433         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 434                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 435                         return gh;
 436         }
 437         return NULL;
 438 }
 439
 440 /**
 441  * state_change - record that the glock is now in a different state
 442  * @gl: the glock
 443  * @new_state the new state
 444  *
 445  */
 446
 447 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 448 {
 449         int held1, held2;
 450
 451         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 452         held2 = (new_state != LM_ST_UNLOCKED);
 453
 454         if (held1 != held2) {
 455                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 456                 if (held2)
 457                         gl->gl_lockref.count++;
 458                 else
 459                         gl->gl_lockref.count--;
 460         }
 461         if (new_state != gl->gl_target)
 462                 /* shorten our minimum hold time */
 463                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 464                                        GL_GLOCK_MIN_HOLD);
 465         gl->gl_state = new_state;
 466         gl->gl_tchange = jiffies;
 467 }
 468
 469 static void gfs2_set_demote(struct gfs2_glock *gl)
 470 {
 471         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 472
 473         set_bit(GLF_DEMOTE, &gl->gl_flags);
 474         smp_mb();
 475         wake_up(&sdp->sd_async_glock_wait);
 476 }
 477
 478 static void gfs2_demote_wake(struct gfs2_glock *gl)
 479 {
 480         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 481         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 482         smp_mb__after_atomic();
 483         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 484 }
 485
 486 /**
 487  * finish_xmote - The DLM has replied to one of our lock requests
 488  * @gl: The glock
 489  * @ret: The status from the DLM
 490  *
 491  */
 492
 493 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 494 {
 495         const struct gfs2_glock_operations *glops = gl->gl_ops;
 496         struct gfs2_holder *gh;
 497         unsigned state = ret & LM_OUT_ST_MASK;
 498         int rv;
 499
 500         spin_lock(&gl->gl_lockref.lock);
 501         trace_gfs2_glock_state_change(gl, state);
 502         state_change(gl, state);
 503         gh = find_first_waiter(gl);
 504
 505         /* Demote to UN request arrived during demote to SH or DF */
 506         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 507             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 508                 gl->gl_target = LM_ST_UNLOCKED;
 509
 510         /* Check for state != intended state */
 511         if (unlikely(state != gl->gl_target)) {
 512                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 513                         /* move to back of queue and try next entry */
 514                         if (ret & LM_OUT_CANCELED) {
 515                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 516                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 517                                 gh = find_first_waiter(gl);
 518                                 gl->gl_target = gh->gh_state;
 519                                 goto retry;
 520                         }
 521                         /* Some error or failed "try lock" - report it */
 522                         if ((ret & LM_OUT_ERROR) ||
 523                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 524                                 gl->gl_target = gl->gl_state;
 525                                 do_error(gl, ret);
 526                                 goto out;
 527                         }
 528                 }
 529                 switch(state) {
 530                 /* Unlocked due to conversion deadlock, try again */
 531                 case LM_ST_UNLOCKED:
 532 retry:
 533                         do_xmote(gl, gh, gl->gl_target);
 534                         break;
 535                 /* Conversion fails, unlock and try again */
 536                 case LM_ST_SHARED:
 537                 case LM_ST_DEFERRED:
 538                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 539                         break;
 540                 default: /* Everything else */
 541                         fs_err(gl->gl_name.ln_sbd, "wanted %u got %u\n",
 542                                gl->gl_target, state);
 543                         GLOCK_BUG_ON(gl, 1);
 544                 }
 545                 spin_unlock(&gl->gl_lockref.lock);
 546                 return;
 547         }
 548
 549         /* Fast path - we got what we asked for */
 550         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 551                 gfs2_demote_wake(gl);
 552         if (state != LM_ST_UNLOCKED) {
 553                 if (glops->go_xmote_bh) {
 554                         spin_unlock(&gl->gl_lockref.lock);
 555                         rv = glops->go_xmote_bh(gl, gh);
 556                         spin_lock(&gl->gl_lockref.lock);
 557                         if (rv) {
 558                                 do_error(gl, rv);
 559                                 goto out;
 560                         }
 561                 }
 562                 rv = do_promote(gl);
 563                 if (rv == 2)
 564                         goto out_locked;
 565         }
 566 out:
 567         clear_bit(GLF_LOCK, &gl->gl_flags);
 568 out_locked:
 569         spin_unlock(&gl->gl_lockref.lock);
 570 }
 571
 572 static bool is_system_glock(struct gfs2_glock *gl)
 573 {
 574         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 575         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 576
 577         if (gl == m_ip->i_gl)
 578                 return true;
 579         return false;
 580 }
 581
 582 /**
 583  * do_xmote - Calls the DLM to change the state of a lock
 584  * @gl: The lock state
 585  * @gh: The holder (only for promotes)
 586  * @target: The target lock state
 587  *
 588  */
 589
 590 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 591 __releases(&gl->gl_lockref.lock)
 592 __acquires(&gl->gl_lockref.lock)
 593 {
 594         const struct gfs2_glock_operations *glops = gl->gl_ops;
 595         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 596         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 597         int ret;
 598
 599         if (target != LM_ST_UNLOCKED && glock_blocked_by_withdraw(gl) &&
 600             gh && !(gh->gh_flags & LM_FLAG_NOEXP))
 601                 return;
 602         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 603                       LM_FLAG_PRIORITY);
 604         GLOCK_BUG_ON(gl, gl->gl_state == target);
 605         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 606         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 607             glops->go_inval) {
 608                 /*
 609                  * If another process is already doing the invalidate, let that
 610                  * finish first.  The glock state machine will get back to this
 611                  * holder again later.
 612                  */
 613                 if (test_and_set_bit(GLF_INVALIDATE_IN_PROGRESS,
 614                                      &gl->gl_flags))
 615                         return;
 616                 do_error(gl, 0); /* Fail queued try locks */
 617         }
 618         gl->gl_req = target;
 619         set_bit(GLF_BLOCKING, &gl->gl_flags);
 620         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 621             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 622             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 623                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 624         spin_unlock(&gl->gl_lockref.lock);
 625         if (glops->go_sync) {
 626                 ret = glops->go_sync(gl);
 627                 /* If we had a problem syncing (due to io errors or whatever,
 628                  * we should not invalidate the metadata or tell dlm to
 629                  * release the glock to other nodes.
 630                  */
 631                 if (ret) {
 632                         if (cmpxchg(&sdp->sd_log_error, 0, ret)) {
 633                                 fs_err(sdp, "Error %d syncing glock \n", ret);
 634                                 gfs2_dump_glock(NULL, gl, true);
 635                         }
 636                         goto skip_inval;
 637                 }
 638         }
 639         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags)) {
 640                 /*
 641                  * The call to go_sync should have cleared out the ail list.
 642                  * If there are still items, we have a problem. We ought to
 643                  * withdraw, but we can't because the withdraw code also uses
 644                  * glocks. Warn about the error, dump the glock, then fall
 645                  * through and wait for logd to do the withdraw for us.
 646                  */
 647                 if ((atomic_read(&gl->gl_ail_count) != 0) &&
 648                     (!cmpxchg(&sdp->sd_log_error, 0, -EIO))) {
 649                         gfs2_glock_assert_warn(gl,
 650                                                !atomic_read(&gl->gl_ail_count));
 651                         gfs2_dump_glock(NULL, gl, true);
 652                 }
 653                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 654                 clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 655         }
 656
 657 skip_inval:
 658         gfs2_glock_hold(gl);
 659         /*
 660          * Check for an error encountered since we called go_sync and go_inval.
 661          * If so, we can't withdraw from the glock code because the withdraw
 662          * code itself uses glocks (see function signal_our_withdraw) to
 663          * change the mount to read-only. Most importantly, we must not call
 664          * dlm to unlock the glock until the journal is in a known good state
 665          * (after journal replay) otherwise other nodes may use the object
 666          * (rgrp or dinode) and then later, journal replay will corrupt the
 667          * file system. The best we can do here is wait for the logd daemon
 668          * to see sd_log_error and withdraw, and in the meantime, requeue the
 669          * work for later.
 670          *
 671          * We make a special exception for some system glocks, such as the
 672          * system statfs inode glock, which needs to be granted before the
 673          * gfs2_quotad daemon can exit, and that exit needs to finish before
 674          * we can unmount the withdrawn file system.
 675          *
 676          * However, if we're just unlocking the lock (say, for unmount, when
 677          * gfs2_gl_hash_clear calls clear_glock) and recovery is complete
 678          * then it's okay to tell dlm to unlock it.
 679          */
 680         if (unlikely(sdp->sd_log_error && !gfs2_withdrawn(sdp)))
 681                 gfs2_withdraw_delayed(sdp);
 682         if (glock_blocked_by_withdraw(gl) &&
 683             (target != LM_ST_UNLOCKED ||
 684              test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags))) {
 685                 if (!is_system_glock(gl)) {
 686                         gfs2_glock_queue_work(gl, GL_GLOCK_DFT_HOLD);
 687                         goto out;
 688                 } else {
 689                         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 690                 }
 691         }
 692
 693         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 694                 /* lock_dlm */
 695                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 696                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 697                     target == LM_ST_UNLOCKED &&
 698                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 699                         finish_xmote(gl, target);
 700                         gfs2_glock_queue_work(gl, 0);
 701                 } else if (ret) {
 702                         fs_err(sdp, "lm_lock ret %d\n", ret);
 703                         GLOCK_BUG_ON(gl, !gfs2_withdrawn(sdp));
 704                 }
 705         } else { /* lock_nolock */
 706                 finish_xmote(gl, target);
 707                 gfs2_glock_queue_work(gl, 0);
 708         }
 709 out:
 710         spin_lock(&gl->gl_lockref.lock);
 711 }
 712
 713 /**
 714  * find_first_holder - find the first "holder" gh
 715  * @gl: the glock
 716  */
 717
 718 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 719 {
 720         struct gfs2_holder *gh;
 721
 722         if (!list_empty(&gl->gl_holders)) {
 723                 gh = list_first_entry(&gl->gl_holders, struct gfs2_holder, gh_list);
 724                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 725                         return gh;
 726         }
 727         return NULL;
 728 }
 729
 730 /**
 731  * run_queue - do all outstanding tasks related to a glock
 732  * @gl: The glock in question
 733  * @nonblock: True if we must not block in run_queue
 734  *
 735  */
 736
 737 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 738 __releases(&gl->gl_lockref.lock)
 739 __acquires(&gl->gl_lockref.lock)
 740 {
 741         struct gfs2_holder *gh = NULL;
 742         int ret;
 743
 744         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 745                 return;
 746
 747         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 748
 749         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 750             gl->gl_demote_state != gl->gl_state) {
 751                 if (find_first_holder(gl))
 752                         goto out_unlock;
 753                 if (nonblock)
 754                         goto out_sched;
 755                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 756                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 757                 gl->gl_target = gl->gl_demote_state;
 758         } else {
 759                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 760                         gfs2_demote_wake(gl);
 761                 ret = do_promote(gl);
 762                 if (ret == 0)
 763                         goto out_unlock;
 764                 if (ret == 2)
 765                         goto out;
 766                 gh = find_first_waiter(gl);
 767                 gl->gl_target = gh->gh_state;
 768                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 769                         do_error(gl, 0); /* Fail queued try locks */
 770         }
 771         do_xmote(gl, gh, gl->gl_target);
 772 out:
 773         return;
 774
 775 out_sched:
 776         clear_bit(GLF_LOCK, &gl->gl_flags);
 777         smp_mb__after_atomic();
 778         gl->gl_lockref.count++;
 779         __gfs2_glock_queue_work(gl, 0);
 780         return;
 781
 782 out_unlock:
 783         clear_bit(GLF_LOCK, &gl->gl_flags);
 784         smp_mb__after_atomic();
 785         return;
 786 }
 787
 788 void gfs2_inode_remember_delete(struct gfs2_glock *gl, u64 generation)
 789 {
 790         struct gfs2_inode_lvb *ri = (void *)gl->gl_lksb.sb_lvbptr;
 791
 792         if (ri->ri_magic == 0)
 793                 ri->ri_magic = cpu_to_be32(GFS2_MAGIC);
 794         if (ri->ri_magic == cpu_to_be32(GFS2_MAGIC))
 795                 ri->ri_generation_deleted = cpu_to_be64(generation);
 796 }
 797
 798 bool gfs2_inode_already_deleted(struct gfs2_glock *gl, u64 generation)
 799 {
 800         struct gfs2_inode_lvb *ri = (void *)gl->gl_lksb.sb_lvbptr;
 801
 802         if (ri->ri_magic != cpu_to_be32(GFS2_MAGIC))
 803                 return false;
 804         return generation <= be64_to_cpu(ri->ri_generation_deleted);
 805 }
 806
 807 static void gfs2_glock_poke(struct gfs2_glock *gl)
 808 {
 809         int flags = LM_FLAG_TRY_1CB | LM_FLAG_ANY | GL_SKIP;
 810         struct gfs2_holder gh;
 811         int error;
 812
 813         gfs2_holder_init(gl, LM_ST_SHARED, flags, &gh);
 814         error = gfs2_glock_nq(&gh);
 815         if (!error)
 816                 gfs2_glock_dq(&gh);
 817         gfs2_holder_uninit(&gh);
 818 }
 819
 820 static bool gfs2_try_evict(struct gfs2_glock *gl)
 821 {
 822         struct gfs2_inode *ip;
 823         bool evicted = false;
 824
 825         /*
 826          * If there is contention on the iopen glock and we have an inode, try
 827          * to grab and release the inode so that it can be evicted.  This will
 828          * allow the remote node to go ahead and delete the inode without us
 829          * having to do it, which will avoid rgrp glock thrashing.
 830          *
 831          * The remote node is likely still holding the corresponding inode
 832          * glock, so it will run before we get to verify that the delete has
 833          * happened below.
 834          */
 835         spin_lock(&gl->gl_lockref.lock);
 836         ip = gl->gl_object;
 837         if (ip && !igrab(&ip->i_inode))
 838                 ip = NULL;
 839         spin_unlock(&gl->gl_lockref.lock);
 840         if (ip) {
 841                 struct gfs2_glock *inode_gl = NULL;
 842
 843                 gl->gl_no_formal_ino = ip->i_no_formal_ino;
 844                 set_bit(GIF_DEFERRED_DELETE, &ip->i_flags);
 845                 d_prune_aliases(&ip->i_inode);
 846                 iput(&ip->i_inode);
 847
 848                 /* If the inode was evicted, gl->gl_object will now be NULL. */
 849                 spin_lock(&gl->gl_lockref.lock);
 850                 ip = gl->gl_object;
 851                 if (ip) {
 852                         inode_gl = ip->i_gl;
 853                         lockref_get(&inode_gl->gl_lockref);
 854                         clear_bit(GIF_DEFERRED_DELETE, &ip->i_flags);
 855                 }
 856                 spin_unlock(&gl->gl_lockref.lock);
 857                 if (inode_gl) {
 858                         gfs2_glock_poke(inode_gl);
 859                         gfs2_glock_put(inode_gl);
 860                 }
 861                 evicted = !ip;
 862         }
 863         return evicted;
 864 }
 865
 866 static void delete_work_func(struct work_struct *work)
 867 {
 868         struct delayed_work *dwork = to_delayed_work(work);
 869         struct gfs2_glock *gl = container_of(dwork, struct gfs2_glock, gl_delete);
 870         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 871         struct inode *inode;
 872         u64 no_addr = gl->gl_name.ln_number;
 873
 874         spin_lock(&gl->gl_lockref.lock);
 875         clear_bit(GLF_PENDING_DELETE, &gl->gl_flags);
 876         spin_unlock(&gl->gl_lockref.lock);
 877
 878         /* If someone's using this glock to create a new dinode, the block must
 879            have been freed by another node, then re-used, in which case our
 880            iopen callback is too late after the fact. Ignore it. */
 881         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 882                 goto out;
 883
 884         if (test_bit(GLF_DEMOTE, &gl->gl_flags)) {
 885                 /*
 886                  * If we can evict the inode, give the remote node trying to
 887                  * delete the inode some time before verifying that the delete
 888                  * has happened.  Otherwise, if we cause contention on the inode glock
 889                  * immediately, the remote node will think that we still have
 890                  * the inode in use, and so it will give up waiting.
 891                  *
 892                  * If we can't evict the inode, signal to the remote node that
 893                  * the inode is still in use.  We'll later try to delete the
 894                  * inode locally in gfs2_evict_inode.
 895                  *
 896                  * FIXME: We only need to verify that the remote node has
 897                  * deleted the inode because nodes before this remote delete
 898                  * rework won't cooperate.  At a later time, when we no longer
 899                  * care about compatibility with such nodes, we can skip this
 900                  * step entirely.
 901                  */
 902                 if (gfs2_try_evict(gl)) {
 903                         if (gfs2_queue_delete_work(gl, 5 * HZ))
 904                                 return;
 905                 }
 906                 goto out;
 907         }
 908
 909         inode = gfs2_lookup_by_inum(sdp, no_addr, gl->gl_no_formal_ino,
 910                                     GFS2_BLKST_UNLINKED);
 911         if (!IS_ERR_OR_NULL(inode)) {
 912                 d_prune_aliases(inode);
 913                 iput(inode);
 914         }
 915 out:
 916         gfs2_glock_put(gl);
 917 }
 918
 919 static void glock_work_func(struct work_struct *work)
 920 {
 921         unsigned long delay = 0;
 922         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 923         unsigned int drop_refs = 1;
 924
 925         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 926                 finish_xmote(gl, gl->gl_reply);
 927                 drop_refs++;
 928         }
 929         spin_lock(&gl->gl_lockref.lock);
 930         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 931             gl->gl_state != LM_ST_UNLOCKED &&
 932             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 933                 unsigned long holdtime, now = jiffies;
 934
 935                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 936                 if (time_before(now, holdtime))
 937                         delay = holdtime - now;
 938
 939                 if (!delay) {
 940                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 941                         gfs2_set_demote(gl);
 942                 }
 943         }
 944         run_queue(gl, 0);
 945         if (delay) {
 946                 /* Keep one glock reference for the work we requeue. */
 947                 drop_refs--;
 948                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 949                         delay = 0;
 950                 __gfs2_glock_queue_work(gl, delay);
 951         }
 952
 953         /*
 954          * Drop the remaining glock references manually here. (Mind that
 955          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 956          * here as well.)
 957          */
 958         gl->gl_lockref.count -= drop_refs;
 959         if (!gl->gl_lockref.count) {
 960                 __gfs2_glock_put(gl);
 961                 return;
 962         }
 963         spin_unlock(&gl->gl_lockref.lock);
 964 }
 965
 966 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 967                                             struct gfs2_glock *new)
 968 {
 969         struct wait_glock_queue wait;
 970         wait_queue_head_t *wq = glock_waitqueue(name);
 971         struct gfs2_glock *gl;
 972
 973         wait.name = name;
 974         init_wait(&wait.wait);
 975         wait.wait.func = glock_wake_function;
 976
 977 again:
 978         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 979         rcu_read_lock();
 980         if (new) {
 981                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 982                         &new->gl_node, ht_parms);
 983                 if (IS_ERR(gl))
 984                         goto out;
 985         } else {
 986                 gl = rhashtable_lookup_fast(&gl_hash_table,
 987                         name, ht_parms);
 988         }
 989         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 990                 rcu_read_unlock();
 991                 schedule();
 992                 goto again;
 993         }
 994 out:
 995         rcu_read_unlock();
 996         finish_wait(wq, &wait.wait);
 997         return gl;
 998 }
 999
1000 /**
1001  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
1002  * @sdp: The GFS2 superblock
1003  * @number: the lock number
1004  * @glops: The glock_operations to use
1005  * @create: If 0, don't create the glock if it doesn't exist
1006  * @glp: the glock is returned here
1007  *
1008  * This does not lock a glock, just finds/creates structures for one.
1009  *
1010  * Returns: errno
1011  */
1012
1013 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
1014                    const struct gfs2_glock_operations *glops, int create,
1015                    struct gfs2_glock **glp)
1016 {
1017         struct super_block *s = sdp->sd_vfs;
1018         struct lm_lockname name = { .ln_number = number,
1019                                     .ln_type = glops->go_type,
1020                                     .ln_sbd = sdp };
1021         struct gfs2_glock *gl, *tmp;
1022         struct address_space *mapping;
1023         struct kmem_cache *cachep;
1024         int ret = 0;
1025
1026         gl = find_insert_glock(&name, NULL);
1027         if (gl) {
1028                 *glp = gl;
1029                 return 0;
1030         }
1031         if (!create)
1032                 return -ENOENT;
1033
1034         if (glops->go_flags & GLOF_ASPACE)
1035                 cachep = gfs2_glock_aspace_cachep;
1036         else
1037                 cachep = gfs2_glock_cachep;
1038         gl = kmem_cache_alloc(cachep, GFP_NOFS);
1039         if (!gl)
1040                 return -ENOMEM;
1041
1042         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
1043
1044         if (glops->go_flags & GLOF_LVB) {
1045                 gl->gl_lksb.sb_lvbptr = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1046                 if (!gl->gl_lksb.sb_lvbptr) {
1047                         kmem_cache_free(cachep, gl);
1048                         return -ENOMEM;
1049                 }
1050         }
1051
1052         atomic_inc(&sdp->sd_glock_disposal);
1053         gl->gl_node.next = NULL;
1054         gl->gl_flags = 0;
1055         gl->gl_name = name;
1056         lockdep_set_subclass(&gl->gl_lockref.lock, glops->go_subclass);
1057         gl->gl_lockref.count = 1;
1058         gl->gl_state = LM_ST_UNLOCKED;
1059         gl->gl_target = LM_ST_UNLOCKED;
1060         gl->gl_demote_state = LM_ST_EXCLUSIVE;
1061         gl->gl_ops = glops;
1062         gl->gl_dstamp = 0;
1063         preempt_disable();
1064         /* We use the global stats to estimate the initial per-glock stats */
1065         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
1066         preempt_enable();
1067         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
1068         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
1069         gl->gl_tchange = jiffies;
1070         gl->gl_object = NULL;
1071         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
1072         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
1073         if (gl->gl_name.ln_type == LM_TYPE_IOPEN)
1074                 INIT_DELAYED_WORK(&gl->gl_delete, delete_work_func);
1075
1076         mapping = gfs2_glock2aspace(gl);
1077         if (mapping) {
1078                 mapping->a_ops = &gfs2_meta_aops;
1079                 mapping->host = s->s_bdev->bd_inode;
1080                 mapping->flags = 0;
1081                 mapping_set_gfp_mask(mapping, GFP_NOFS);
1082                 mapping->private_data = NULL;
1083                 mapping->writeback_index = 0;
1084         }
1085
1086         tmp = find_insert_glock(&name, gl);
1087         if (!tmp) {
1088                 *glp = gl;
1089                 goto out;
1090         }
1091         if (IS_ERR(tmp)) {
1092                 ret = PTR_ERR(tmp);
1093                 goto out_free;
1094         }
1095         *glp = tmp;
1096
1097 out_free:
1098         kfree(gl->gl_lksb.sb_lvbptr);
1099         kmem_cache_free(cachep, gl);
1100         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
1101                 wake_up(&sdp->sd_glock_wait);
1102
1103 out:
1104         return ret;
1105 }
1106
1107 /**
1108  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
1109  * @gl: the glock
1110  * @state: the state we're requesting
1111  * @flags: the modifier flags
1112  * @gh: the holder structure
1113  *
1114  */
1115
1116 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
1117                       struct gfs2_holder *gh)
1118 {
1119         INIT_LIST_HEAD(&gh->gh_list);
1120         gh->gh_gl = gl;
1121         gh->gh_ip = _RET_IP_;
1122         gh->gh_owner_pid = get_pid(task_pid(current));
1123         gh->gh_state = state;
1124         gh->gh_flags = flags;
1125         gh->gh_error = 0;
1126         gh->gh_iflags = 0;
1127         gfs2_glock_hold(gl);
1128 }
1129
1130 /**
1131  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
1132  * @state: the state we're requesting
1133  * @flags: the modifier flags
1134  * @gh: the holder structure
1135  *
1136  * Don't mess with the glock.
1137  *
1138  */
1139
1140 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
1141 {
1142         gh->gh_state = state;
1143         gh->gh_flags = flags;
1144         gh->gh_iflags = 0;
1145         gh->gh_ip = _RET_IP_;
1146         put_pid(gh->gh_owner_pid);
1147         gh->gh_owner_pid = get_pid(task_pid(current));
1148 }
1149
1150 /**
1151  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
1152  * @gh: the holder structure
1153  *
1154  */
1155
1156 void gfs2_holder_uninit(struct gfs2_holder *gh)
1157 {
1158         put_pid(gh->gh_owner_pid);
1159         gfs2_glock_put(gh->gh_gl);
1160         gfs2_holder_mark_uninitialized(gh);
1161         gh->gh_ip = 0;
1162 }
1163
1164 static void gfs2_glock_update_hold_time(struct gfs2_glock *gl,
1165                                         unsigned long start_time)
1166 {
1167         /* Have we waited longer that a second? */
1168         if (time_after(jiffies, start_time + HZ)) {
1169                 /* Lengthen the minimum hold time. */
1170                 gl->gl_hold_time = min(gl->gl_hold_time + GL_GLOCK_HOLD_INCR,
1171                                        GL_GLOCK_MAX_HOLD);
1172         }
1173 }
1174
1175 /**
1176  * gfs2_glock_wait - wait on a glock acquisition
1177  * @gh: the glock holder
1178  *
1179  * Returns: 0 on success
1180  */
1181
1182 int gfs2_glock_wait(struct gfs2_holder *gh)
1183 {
1184         unsigned long start_time = jiffies;
1185
1186         might_sleep();
1187         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
1188         gfs2_glock_update_hold_time(gh->gh_gl, start_time);
1189         return gh->gh_error;
1190 }
1191
1192 static int glocks_pending(unsigned int num_gh, struct gfs2_holder *ghs)
1193 {
1194         int i;
1195
1196         for (i = 0; i < num_gh; i++)
1197                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags))
1198                         return 1;
1199         return 0;
1200 }
1201
1202 /**
1203  * gfs2_glock_async_wait - wait on multiple asynchronous glock acquisitions
1204  * @num_gh: the number of holders in the array
1205  * @ghs: the glock holder array
1206  *
1207  * Returns: 0 on success, meaning all glocks have been granted and are held.
1208  *          -ESTALE if the request timed out, meaning all glocks were released,
1209  *          and the caller should retry the operation.
1210  */
1211
1212 int gfs2_glock_async_wait(unsigned int num_gh, struct gfs2_holder *ghs)
1213 {
1214         struct gfs2_sbd *sdp = ghs[0].gh_gl->gl_name.ln_sbd;
1215         int i, ret = 0, timeout = 0;
1216         unsigned long start_time = jiffies;
1217         bool keep_waiting;
1218
1219         might_sleep();
1220         /*
1221          * Total up the (minimum hold time * 2) of all glocks and use that to
1222          * determine the max amount of time we should wait.
1223          */
1224         for (i = 0; i < num_gh; i++)
1225                 timeout += ghs[i].gh_gl->gl_hold_time << 1;
1226
1227 wait_for_dlm:
1228         if (!wait_event_timeout(sdp->sd_async_glock_wait,
1229                                 !glocks_pending(num_gh, ghs), timeout))
1230                 ret = -ESTALE; /* request timed out. */
1231
1232         /*
1233          * If dlm granted all our requests, we need to adjust the glock
1234          * minimum hold time values according to how long we waited.
1235          *
1236          * If our request timed out, we need to repeatedly release any held
1237          * glocks we acquired thus far to allow dlm to acquire the remaining
1238          * glocks without deadlocking.  We cannot currently cancel outstanding
1239          * glock acquisitions.
1240          *
1241          * The HIF_WAIT bit tells us which requests still need a response from
1242          * dlm.
1243          *
1244          * If dlm sent us any errors, we return the first error we find.
1245          */
1246         keep_waiting = false;
1247         for (i = 0; i < num_gh; i++) {
1248                 /* Skip holders we have already dequeued below. */
1249                 if (!gfs2_holder_queued(&ghs[i]))
1250                         continue;
1251                 /* Skip holders with a pending DLM response. */
1252                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags)) {
1253                         keep_waiting = true;
1254                         continue;
1255                 }
1256
1257                 if (test_bit(HIF_HOLDER, &ghs[i].gh_iflags)) {
1258                         if (ret == -ESTALE)
1259                                 gfs2_glock_dq(&ghs[i]);
1260                         else
1261                                 gfs2_glock_update_hold_time(ghs[i].gh_gl,
1262                                                             start_time);
1263                 }
1264                 if (!ret)
1265                         ret = ghs[i].gh_error;
1266         }
1267
1268         if (keep_waiting)
1269                 goto wait_for_dlm;
1270
1271         /*
1272          * At this point, we've either acquired all locks or released them all.
1273          */
1274         return ret;
1275 }
1276
1277 /**
1278  * handle_callback - process a demote request
1279  * @gl: the glock
1280  * @state: the state the caller wants us to change to
1281  *
1282  * There are only two requests that we are going to see in actual
1283  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
1284  */
1285
1286 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
1287                             unsigned long delay, bool remote)
1288 {
1289         if (delay)
1290                 set_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
1291         else
1292                 gfs2_set_demote(gl);
1293         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
1294                 gl->gl_demote_state = state;
1295                 gl->gl_demote_time = jiffies;
1296         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
1297                         gl->gl_demote_state != state) {
1298                 gl->gl_demote_state = LM_ST_UNLOCKED;
1299         }
1300         if (gl->gl_ops->go_callback)
1301                 gl->gl_ops->go_callback(gl, remote);
1302         trace_gfs2_demote_rq(gl, remote);
1303 }
1304
1305 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
1306 {
1307         struct va_format vaf;
1308         va_list args;
1309
1310         va_start(args, fmt);
1311
1312         if (seq) {
1313                 seq_vprintf(seq, fmt, args);
1314         } else {
1315                 vaf.fmt = fmt;
1316                 vaf.va = &args;
1317
1318                 pr_err("%pV", &vaf);
1319         }
1320
1321         va_end(args);
1322 }
1323
1324 /**
1325  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1326  * @gh: the holder structure to add
1327  *
1328  * Eventually we should move the recursive locking trap to a
1329  * debugging option or something like that. This is the fast
1330  * path and needs to have the minimum number of distractions.
1331  *
1332  */
1333
1334 static inline void add_to_queue(struct gfs2_holder *gh)
1335 __releases(&gl->gl_lockref.lock)
1336 __acquires(&gl->gl_lockref.lock)
1337 {
1338         struct gfs2_glock *gl = gh->gh_gl;
1339         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1340         struct list_head *insert_pt = NULL;
1341         struct gfs2_holder *gh2;
1342         int try_futile = 0;
1343
1344         GLOCK_BUG_ON(gl, gh->gh_owner_pid == NULL);
1345         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1346                 GLOCK_BUG_ON(gl, true);
1347
1348         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1349                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1350                         try_futile = !may_grant(gl, gh);
1351                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1352                         goto fail;
1353         }
1354
1355         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1356                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1357                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1358                         goto trap_recursive;
1359                 if (try_futile &&
1360                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1361 fail:
1362                         gh->gh_error = GLR_TRYFAILED;
1363                         gfs2_holder_wake(gh);
1364                         return;
1365                 }
1366                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1367                         continue;
1368                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1369                         insert_pt = &gh2->gh_list;
1370         }
1371         trace_gfs2_glock_queue(gh, 1);
1372         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1373         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1374         if (likely(insert_pt == NULL)) {
1375                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1376                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1377                         goto do_cancel;
1378                 return;
1379         }
1380         list_add_tail(&gh->gh_list, insert_pt);
1381 do_cancel:
1382         gh = list_first_entry(&gl->gl_holders, struct gfs2_holder, gh_list);
1383         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1384                 spin_unlock(&gl->gl_lockref.lock);
1385                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1386                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1387                 spin_lock(&gl->gl_lockref.lock);
1388         }
1389         return;
1390
1391 trap_recursive:
1392         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1393         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1394         fs_err(sdp, "lock type: %d req lock state : %d\n",
1395                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1396         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1397         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1398         fs_err(sdp, "lock type: %d req lock state : %d\n",
1399                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1400         gfs2_dump_glock(NULL, gl, true);
1401         BUG();
1402 }
1403
1404 /**
1405  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1406  * @gh: the holder structure
1407  *
1408  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1409  *
1410  * Returns: 0, GLR_TRYFAILED, or errno on failure
1411  */
1412
1413 int gfs2_glock_nq(struct gfs2_holder *gh)
1414 {
1415         struct gfs2_glock *gl = gh->gh_gl;
1416         int error = 0;
1417
1418         if (glock_blocked_by_withdraw(gl) && !(gh->gh_flags & LM_FLAG_NOEXP))
1419                 return -EIO;
1420
1421         if (test_bit(GLF_LRU, &gl->gl_flags))
1422                 gfs2_glock_remove_from_lru(gl);
1423
1424         spin_lock(&gl->gl_lockref.lock);
1425         add_to_queue(gh);
1426         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1427                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1428                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1429                 gl->gl_lockref.count++;
1430                 __gfs2_glock_queue_work(gl, 0);
1431         }
1432         run_queue(gl, 1);
1433         spin_unlock(&gl->gl_lockref.lock);
1434
1435         if (!(gh->gh_flags & GL_ASYNC))
1436                 error = gfs2_glock_wait(gh);
1437
1438         return error;
1439 }
1440
1441 /**
1442  * gfs2_glock_poll - poll to see if an async request has been completed
1443  * @gh: the holder
1444  *
1445  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1446  */
1447
1448 int gfs2_glock_poll(struct gfs2_holder *gh)
1449 {
1450         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1451 }
1452
1453 /**
1454  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1455  * @gh: the glock holder
1456  *
1457  */
1458
1459 void gfs2_glock_dq(struct gfs2_holder *gh)
1460 {
1461         struct gfs2_glock *gl = gh->gh_gl;
1462         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1463         unsigned delay = 0;
1464         int fast_path = 0;
1465
1466         spin_lock(&gl->gl_lockref.lock);
1467         /*
1468          * If we're in the process of file system withdraw, we cannot just
1469          * dequeue any glocks until our journal is recovered, lest we
1470          * introduce file system corruption. We need two exceptions to this
1471          * rule: We need to allow unlocking of nondisk glocks and the glock
1472          * for our own journal that needs recovery.
1473          */
1474         if (test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags) &&
1475             glock_blocked_by_withdraw(gl) &&
1476             gh->gh_gl != sdp->sd_jinode_gl) {
1477                 sdp->sd_glock_dqs_held++;
1478                 spin_unlock(&gl->gl_lockref.lock);
1479                 might_sleep();
1480                 wait_on_bit(&sdp->sd_flags, SDF_WITHDRAW_RECOVERY,
1481                             TASK_UNINTERRUPTIBLE);
1482                 spin_lock(&gl->gl_lockref.lock);
1483         }
1484         if (gh->gh_flags & GL_NOCACHE)
1485                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1486
1487         list_del_init(&gh->gh_list);
1488         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1489         if (find_first_holder(gl) == NULL) {
1490                 if (list_empty(&gl->gl_holders) &&
1491                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1492                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1493                         fast_path = 1;
1494         }
1495         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1496                 gfs2_glock_add_to_lru(gl);
1497
1498         trace_gfs2_glock_queue(gh, 0);
1499         if (unlikely(!fast_path)) {
1500                 gl->gl_lockref.count++;
1501                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1502                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1503                     gl->gl_name.ln_type == LM_TYPE_INODE)
1504                         delay = gl->gl_hold_time;
1505                 __gfs2_glock_queue_work(gl, delay);
1506         }
1507         spin_unlock(&gl->gl_lockref.lock);
1508 }
1509
1510 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1511 {
1512         struct gfs2_glock *gl = gh->gh_gl;
1513         gfs2_glock_dq(gh);
1514         might_sleep();
1515         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1516 }
1517
1518 /**
1519  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1520  * @gh: the holder structure
1521  *
1522  */
1523
1524 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1525 {
1526         gfs2_glock_dq(gh);
1527         gfs2_holder_uninit(gh);
1528 }
1529
1530 /**
1531  * gfs2_glock_nq_num - acquire a glock based on lock number
1532  * @sdp: the filesystem
1533  * @number: the lock number
1534  * @glops: the glock operations for the type of glock
1535  * @state: the state to acquire the glock in
1536  * @flags: modifier flags for the acquisition
1537  * @gh: the struct gfs2_holder
1538  *
1539  * Returns: errno
1540  */
1541
1542 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1543                       const struct gfs2_glock_operations *glops,
1544                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1545 {
1546         struct gfs2_glock *gl;
1547         int error;
1548
1549         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1550         if (!error) {
1551                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1552                 gfs2_glock_put(gl);
1553         }
1554
1555         return error;
1556 }
1557
1558 /**
1559  * glock_compare - Compare two struct gfs2_glock structures for sorting
1560  * @arg_a: the first structure
1561  * @arg_b: the second structure
1562  *
1563  */
1564
1565 static int glock_compare(const void *arg_a, const void *arg_b)
1566 {
1567         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1568         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1569         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1570         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1571
1572         if (a->ln_number > b->ln_number)
1573                 return 1;
1574         if (a->ln_number < b->ln_number)
1575                 return -1;
1576         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1577         return 0;
1578 }
1579
1580 /**
1581  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1582  * @num_gh: the number of structures
1583  * @ghs: an array of struct gfs2_holder structures
1584  *
1585  * Returns: 0 on success (all glocks acquired),
1586  *          errno on failure (no glocks acquired)
1587  */
1588
1589 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1590                      struct gfs2_holder **p)
1591 {
1592         unsigned int x;
1593         int error = 0;
1594
1595         for (x = 0; x < num_gh; x++)
1596                 p[x] = &ghs[x];
1597
1598         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1599
1600         for (x = 0; x < num_gh; x++) {
1601                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1602
1603                 error = gfs2_glock_nq(p[x]);
1604                 if (error) {
1605                         while (x--)
1606                                 gfs2_glock_dq(p[x]);
1607                         break;
1608                 }
1609         }
1610
1611         return error;
1612 }
1613
1614 /**
1615  * gfs2_glock_nq_m - acquire multiple glocks
1616  * @num_gh: the number of structures
1617  * @ghs: an array of struct gfs2_holder structures
1618  *
1619  *
1620  * Returns: 0 on success (all glocks acquired),
1621  *          errno on failure (no glocks acquired)
1622  */
1623
1624 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1625 {
1626         struct gfs2_holder *tmp[4];
1627         struct gfs2_holder **pph = tmp;
1628         int error = 0;
1629
1630         switch(num_gh) {
1631         case 0:
1632                 return 0;
1633         case 1:
1634                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1635                 return gfs2_glock_nq(ghs);
1636         default:
1637                 if (num_gh <= 4)
1638                         break;
1639                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1640                                     GFP_NOFS);
1641                 if (!pph)
1642                         return -ENOMEM;
1643         }
1644
1645         error = nq_m_sync(num_gh, ghs, pph);
1646
1647         if (pph != tmp)
1648                 kfree(pph);
1649
1650         return error;
1651 }
1652
1653 /**
1654  * gfs2_glock_dq_m - release multiple glocks
1655  * @num_gh: the number of structures
1656  * @ghs: an array of struct gfs2_holder structures
1657  *
1658  */
1659
1660 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1661 {
1662         while (num_gh--)
1663                 gfs2_glock_dq(&ghs[num_gh]);
1664 }
1665
1666 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1667 {
1668         unsigned long delay = 0;
1669         unsigned long holdtime;
1670         unsigned long now = jiffies;
1671
1672         gfs2_glock_hold(gl);
1673         spin_lock(&gl->gl_lockref.lock);
1674         holdtime = gl->gl_tchange + gl->gl_hold_time;
1675         if (!list_empty(&gl->gl_holders) &&
1676             gl->gl_name.ln_type == LM_TYPE_INODE) {
1677                 if (time_before(now, holdtime))
1678                         delay = holdtime - now;
1679                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1680                         delay = gl->gl_hold_time;
1681         }
1682         handle_callback(gl, state, delay, true);
1683         __gfs2_glock_queue_work(gl, delay);
1684         spin_unlock(&gl->gl_lockref.lock);
1685 }
1686
1687 /**
1688  * gfs2_should_freeze - Figure out if glock should be frozen
1689  * @gl: The glock in question
1690  *
1691  * Glocks are not frozen if (a) the result of the dlm operation is
1692  * an error, (b) the locking operation was an unlock operation or
1693  * (c) if there is a "noexp" flagged request anywhere in the queue
1694  *
1695  * Returns: 1 if freezing should occur, 0 otherwise
1696  */
1697
1698 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1699 {
1700         const struct gfs2_holder *gh;
1701
1702         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1703                 return 0;
1704         if (gl->gl_target == LM_ST_UNLOCKED)
1705                 return 0;
1706
1707         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1708                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1709                         continue;
1710                 if (LM_FLAG_NOEXP & gh->gh_flags)
1711                         return 0;
1712         }
1713
1714         return 1;
1715 }
1716
1717 /**
1718  * gfs2_glock_complete - Callback used by locking
1719  * @gl: Pointer to the glock
1720  * @ret: The return value from the dlm
1721  *
1722  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1723  * to use a bitfield shared with other glock state fields.
1724  */
1725
1726 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1727 {
1728         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1729
1730         spin_lock(&gl->gl_lockref.lock);
1731         gl->gl_reply = ret;
1732
1733         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1734                 if (gfs2_should_freeze(gl)) {
1735                         set_bit(GLF_FROZEN, &gl->gl_flags);
1736                         spin_unlock(&gl->gl_lockref.lock);
1737                         return;
1738                 }
1739         }
1740
1741         gl->gl_lockref.count++;
1742         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1743         __gfs2_glock_queue_work(gl, 0);
1744         spin_unlock(&gl->gl_lockref.lock);
1745 }
1746
1747 static int glock_cmp(void *priv, const struct list_head *a,
1748                      const struct list_head *b)
1749 {
1750         struct gfs2_glock *gla, *glb;
1751
1752         gla = list_entry(a, struct gfs2_glock, gl_lru);
1753         glb = list_entry(b, struct gfs2_glock, gl_lru);
1754
1755         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1756                 return 1;
1757         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1758                 return -1;
1759
1760         return 0;
1761 }
1762
1763 /**
1764  * gfs2_dispose_glock_lru - Demote a list of glocks
1765  * @list: The list to dispose of
1766  *
1767  * Disposing of glocks may involve disk accesses, so that here we sort
1768  * the glocks by number (i.e. disk location of the inodes) so that if
1769  * there are any such accesses, they'll be sent in order (mostly).
1770  *
1771  * Must be called under the lru_lock, but may drop and retake this
1772  * lock. While the lru_lock is dropped, entries may vanish from the
1773  * list, but no new entries will appear on the list (since it is
1774  * private)
1775  */
1776
1777 static void gfs2_dispose_glock_lru(struct list_head *list)
1778 __releases(&lru_lock)
1779 __acquires(&lru_lock)
1780 {
1781         struct gfs2_glock *gl;
1782
1783         list_sort(NULL, list, glock_cmp);
1784
1785         while(!list_empty(list)) {
1786                 gl = list_first_entry(list, struct gfs2_glock, gl_lru);
1787                 list_del_init(&gl->gl_lru);
1788                 clear_bit(GLF_LRU, &gl->gl_flags);
1789                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1790 add_back_to_lru:
1791                         list_add(&gl->gl_lru, &lru_list);
1792                         set_bit(GLF_LRU, &gl->gl_flags);
1793                         atomic_inc(&lru_count);
1794                         continue;
1795                 }
1796                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1797                         spin_unlock(&gl->gl_lockref.lock);
1798                         goto add_back_to_lru;
1799                 }
1800                 gl->gl_lockref.count++;
1801                 if (demote_ok(gl))
1802                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1803                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1804                 __gfs2_glock_queue_work(gl, 0);
1805                 spin_unlock(&gl->gl_lockref.lock);
1806                 cond_resched_lock(&lru_lock);
1807         }
1808 }
1809
1810 /**
1811  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1812  * @nr: The number of entries to scan
1813  *
1814  * This function selects the entries on the LRU which are able to
1815  * be demoted, and then kicks off the process by calling
1816  * gfs2_dispose_glock_lru() above.
1817  */
1818
1819 static long gfs2_scan_glock_lru(int nr)
1820 {
1821         struct gfs2_glock *gl;
1822         LIST_HEAD(skipped);
1823         LIST_HEAD(dispose);
1824         long freed = 0;
1825
1826         spin_lock(&lru_lock);
1827         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1828                 gl = list_first_entry(&lru_list, struct gfs2_glock, gl_lru);
1829
1830                 /* Test for being demotable */
1831                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1832                         list_move(&gl->gl_lru, &dispose);
1833                         atomic_dec(&lru_count);
1834                         freed++;
1835                         continue;
1836                 }
1837
1838                 list_move(&gl->gl_lru, &skipped);
1839         }
1840         list_splice(&skipped, &lru_list);
1841         if (!list_empty(&dispose))
1842                 gfs2_dispose_glock_lru(&dispose);
1843         spin_unlock(&lru_lock);
1844
1845         return freed;
1846 }
1847
1848 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1849                                             struct shrink_control *sc)
1850 {
1851         if (!(sc->gfp_mask & __GFP_FS))
1852                 return SHRINK_STOP;
1853         return gfs2_scan_glock_lru(sc->nr_to_scan);
1854 }
1855
1856 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1857                                              struct shrink_control *sc)
1858 {
1859         return vfs_pressure_ratio(atomic_read(&lru_count));
1860 }
1861
1862 static struct shrinker glock_shrinker = {
1863         .seeks = DEFAULT_SEEKS,
1864         .count_objects = gfs2_glock_shrink_count,
1865         .scan_objects = gfs2_glock_shrink_scan,
1866 };
1867
1868 /**
1869  * glock_hash_walk - Call a function for glock in a hash bucket
1870  * @examiner: the function
1871  * @sdp: the filesystem
1872  *
1873  * Note that the function can be called multiple times on the same
1874  * object.  So the user must ensure that the function can cope with
1875  * that.
1876  */
1877
1878 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1879 {
1880         struct gfs2_glock *gl;
1881         struct rhashtable_iter iter;
1882
1883         rhashtable_walk_enter(&gl_hash_table, &iter);
1884
1885         do {
1886                 rhashtable_walk_start(&iter);
1887
1888                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl)) {
1889                         if (gl->gl_name.ln_sbd == sdp)
1890                                 examiner(gl);
1891                 }
1892
1893                 rhashtable_walk_stop(&iter);
1894         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1895
1896         rhashtable_walk_exit(&iter);
1897 }
1898
1899 bool gfs2_queue_delete_work(struct gfs2_glock *gl, unsigned long delay)
1900 {
1901         bool queued;
1902
1903         spin_lock(&gl->gl_lockref.lock);
1904         queued = queue_delayed_work(gfs2_delete_workqueue,
1905                                     &gl->gl_delete, delay);
1906         if (queued)
1907                 set_bit(GLF_PENDING_DELETE, &gl->gl_flags);
1908         spin_unlock(&gl->gl_lockref.lock);
1909         return queued;
1910 }
1911
1912 void gfs2_cancel_delete_work(struct gfs2_glock *gl)
1913 {
1914         if (cancel_delayed_work(&gl->gl_delete)) {
1915                 clear_bit(GLF_PENDING_DELETE, &gl->gl_flags);
1916                 gfs2_glock_put(gl);
1917         }
1918 }
1919
1920 bool gfs2_delete_work_queued(const struct gfs2_glock *gl)
1921 {
1922         return test_bit(GLF_PENDING_DELETE, &gl->gl_flags);
1923 }
1924
1925 static void flush_delete_work(struct gfs2_glock *gl)
1926 {
1927         if (gl->gl_name.ln_type == LM_TYPE_IOPEN) {
1928                 if (cancel_delayed_work(&gl->gl_delete)) {
1929                         queue_delayed_work(gfs2_delete_workqueue,
1930                                            &gl->gl_delete, 0);
1931                 }
1932         }
1933 }
1934
1935 void gfs2_flush_delete_work(struct gfs2_sbd *sdp)
1936 {
1937         glock_hash_walk(flush_delete_work, sdp);
1938         flush_workqueue(gfs2_delete_workqueue);
1939 }
1940
1941 /**
1942  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1943  * @gl: The glock to thaw
1944  *
1945  */
1946
1947 static void thaw_glock(struct gfs2_glock *gl)
1948 {
1949         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1950                 return;
1951         if (!lockref_get_not_dead(&gl->gl_lockref))
1952                 return;
1953         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1954         gfs2_glock_queue_work(gl, 0);
1955 }
1956
1957 /**
1958  * clear_glock - look at a glock and see if we can free it from glock cache
1959  * @gl: the glock to look at
1960  *
1961  */
1962
1963 static void clear_glock(struct gfs2_glock *gl)
1964 {
1965         gfs2_glock_remove_from_lru(gl);
1966
1967         spin_lock(&gl->gl_lockref.lock);
1968         if (!__lockref_is_dead(&gl->gl_lockref)) {
1969                 gl->gl_lockref.count++;
1970                 if (gl->gl_state != LM_ST_UNLOCKED)
1971                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1972                 __gfs2_glock_queue_work(gl, 0);
1973         }
1974         spin_unlock(&gl->gl_lockref.lock);
1975 }
1976
1977 /**
1978  * gfs2_glock_thaw - Thaw any frozen glocks
1979  * @sdp: The super block
1980  *
1981  */
1982
1983 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1984 {
1985         glock_hash_walk(thaw_glock, sdp);
1986 }
1987
1988 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1989 {
1990         spin_lock(&gl->gl_lockref.lock);
1991         gfs2_dump_glock(seq, gl, fsid);
1992         spin_unlock(&gl->gl_lockref.lock);
1993 }
1994
1995 static void dump_glock_func(struct gfs2_glock *gl)
1996 {
1997         dump_glock(NULL, gl, true);
1998 }
1999
2000 /**
2001  * gfs2_gl_hash_clear - Empty out the glock hash table
2002  * @sdp: the filesystem
2003  * @wait: wait until it's all gone
2004  *
2005  * Called when unmounting the filesystem.
2006  */
2007
2008 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
2009 {
2010         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
2011         flush_workqueue(glock_workqueue);
2012         glock_hash_walk(clear_glock, sdp);
2013         flush_workqueue(glock_workqueue);
2014         wait_event_timeout(sdp->sd_glock_wait,
2015                            atomic_read(&sdp->sd_glock_disposal) == 0,
2016                            HZ * 600);
2017         glock_hash_walk(dump_glock_func, sdp);
2018 }
2019
2020 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
2021 {
2022         struct gfs2_glock *gl = ip->i_gl;
2023         int ret;
2024
2025         ret = gfs2_truncatei_resume(ip);
2026         gfs2_glock_assert_withdraw(gl, ret == 0);
2027
2028         spin_lock(&gl->gl_lockref.lock);
2029         clear_bit(GLF_LOCK, &gl->gl_flags);
2030         run_queue(gl, 1);
2031         spin_unlock(&gl->gl_lockref.lock);
2032 }
2033
2034 static const char *state2str(unsigned state)
2035 {
2036         switch(state) {
2037         case LM_ST_UNLOCKED:
2038                 return "UN";
2039         case LM_ST_SHARED:
2040                 return "SH";
2041         case LM_ST_DEFERRED:
2042                 return "DF";
2043         case LM_ST_EXCLUSIVE:
2044                 return "EX";
2045         }
2046         return "??";
2047 }
2048
2049 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
2050 {
2051         char *p = buf;
2052         if (flags & LM_FLAG_TRY)
2053                 *p++ = 't';
2054         if (flags & LM_FLAG_TRY_1CB)
2055                 *p++ = 'T';
2056         if (flags & LM_FLAG_NOEXP)
2057                 *p++ = 'e';
2058         if (flags & LM_FLAG_ANY)
2059                 *p++ = 'A';
2060         if (flags & LM_FLAG_PRIORITY)
2061                 *p++ = 'p';
2062         if (flags & GL_ASYNC)
2063                 *p++ = 'a';
2064         if (flags & GL_EXACT)
2065                 *p++ = 'E';
2066         if (flags & GL_NOCACHE)
2067                 *p++ = 'c';
2068         if (test_bit(HIF_HOLDER, &iflags))
2069                 *p++ = 'H';
2070         if (test_bit(HIF_WAIT, &iflags))
2071                 *p++ = 'W';
2072         if (test_bit(HIF_FIRST, &iflags))
2073                 *p++ = 'F';
2074         *p = 0;
2075         return buf;
2076 }
2077
2078 /**
2079  * dump_holder - print information about a glock holder
2080  * @seq: the seq_file struct
2081  * @gh: the glock holder
2082  * @fs_id_buf: pointer to file system id (if requested)
2083  *
2084  */
2085
2086 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh,
2087                         const char *fs_id_buf)
2088 {
2089         struct task_struct *gh_owner = NULL;
2090         char flags_buf[32];
2091
2092         rcu_read_lock();
2093         if (gh->gh_owner_pid)
2094                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
2095         gfs2_print_dbg(seq, "%s H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
2096                        fs_id_buf, state2str(gh->gh_state),
2097                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
2098                        gh->gh_error,
2099                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
2100                        gh_owner ? gh_owner->comm : "(ended)",
2101                        (void *)gh->gh_ip);
2102         rcu_read_unlock();
2103 }
2104
2105 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
2106 {
2107         const unsigned long *gflags = &gl->gl_flags;
2108         char *p = buf;
2109
2110         if (test_bit(GLF_LOCK, gflags))
2111                 *p++ = 'l';
2112         if (test_bit(GLF_DEMOTE, gflags))
2113                 *p++ = 'D';
2114         if (test_bit(GLF_PENDING_DEMOTE, gflags))
2115                 *p++ = 'd';
2116         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
2117                 *p++ = 'p';
2118         if (test_bit(GLF_DIRTY, gflags))
2119                 *p++ = 'y';
2120         if (test_bit(GLF_LFLUSH, gflags))
2121                 *p++ = 'f';
2122         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
2123                 *p++ = 'i';
2124         if (test_bit(GLF_REPLY_PENDING, gflags))
2125                 *p++ = 'r';
2126         if (test_bit(GLF_INITIAL, gflags))
2127                 *p++ = 'I';
2128         if (test_bit(GLF_FROZEN, gflags))
2129                 *p++ = 'F';
2130         if (!list_empty(&gl->gl_holders))
2131                 *p++ = 'q';
2132         if (test_bit(GLF_LRU, gflags))
2133                 *p++ = 'L';
2134         if (gl->gl_object)
2135                 *p++ = 'o';
2136         if (test_bit(GLF_BLOCKING, gflags))
2137                 *p++ = 'b';
2138         if (test_bit(GLF_INODE_CREATING, gflags))
2139                 *p++ = 'c';
2140         if (test_bit(GLF_PENDING_DELETE, gflags))
2141                 *p++ = 'P';
2142         if (test_bit(GLF_FREEING, gflags))
2143                 *p++ = 'x';
2144         *p = 0;
2145         return buf;
2146 }
2147
2148 /**
2149  * gfs2_dump_glock - print information about a glock
2150  * @seq: The seq_file struct
2151  * @gl: the glock
2152  * @fsid: If true, also dump the file system id
2153  *
2154  * The file format is as follows:
2155  * One line per object, capital letters are used to indicate objects
2156  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
2157  * other objects are indented by a single space and follow the glock to
2158  * which they are related. Fields are indicated by lower case letters
2159  * followed by a colon and the field value, except for strings which are in
2160  * [] so that its possible to see if they are composed of spaces for
2161  * example. The field's are n = number (id of the object), f = flags,
2162  * t = type, s = state, r = refcount, e = error, p = pid.
2163  *
2164  */
2165
2166 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
2167 {
2168         const struct gfs2_glock_operations *glops = gl->gl_ops;
2169         unsigned long long dtime;
2170         const struct gfs2_holder *gh;
2171         char gflags_buf[32];
2172         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
2173         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
2174         unsigned long nrpages = 0;
2175
2176         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
2177                 struct address_space *mapping = gfs2_glock2aspace(gl);
2178
2179                 nrpages = mapping->nrpages;
2180         }
2181         memset(fs_id_buf, 0, sizeof(fs_id_buf));
2182         if (fsid && sdp) /* safety precaution */
2183                 sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
2184         dtime = jiffies - gl->gl_demote_time;
2185         dtime *= 1000000/HZ; /* demote time in uSec */
2186         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
2187                 dtime = 0;
2188         gfs2_print_dbg(seq, "%sG:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d "
2189                        "v:%d r:%d m:%ld p:%lu\n",
2190                        fs_id_buf, state2str(gl->gl_state),
2191                        gl->gl_name.ln_type,
2192                        (unsigned long long)gl->gl_name.ln_number,
2193                        gflags2str(gflags_buf, gl),
2194                        state2str(gl->gl_target),
2195                        state2str(gl->gl_demote_state), dtime,
2196                        atomic_read(&gl->gl_ail_count),
2197                        atomic_read(&gl->gl_revokes),
2198                        (int)gl->gl_lockref.count, gl->gl_hold_time, nrpages);
2199
2200         list_for_each_entry(gh, &gl->gl_holders, gh_list)
2201                 dump_holder(seq, gh, fs_id_buf);
2202
2203         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
2204                 glops->go_dump(seq, gl, fs_id_buf);
2205 }
2206
2207 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
2208 {
2209         struct gfs2_glock *gl = iter_ptr;
2210
2211         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
2212                    gl->gl_name.ln_type,
2213                    (unsigned long long)gl->gl_name.ln_number,
2214                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
2215                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
2216                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
2217                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
2218                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
2219                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
2220                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
2221                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
2222         return 0;
2223 }
2224
2225 static const char *gfs2_gltype[] = {
2226         "type",
2227         "reserved",
2228         "nondisk",
2229         "inode",
2230         "rgrp",
2231         "meta",
2232         "iopen",
2233         "flock",
2234         "plock",
2235         "quota",
2236         "journal",
2237 };
2238
2239 static const char *gfs2_stype[] = {
2240         [GFS2_LKS_SRTT]         = "srtt",
2241         [GFS2_LKS_SRTTVAR]      = "srttvar",
2242         [GFS2_LKS_SRTTB]        = "srttb",
2243         [GFS2_LKS_SRTTVARB]     = "srttvarb",
2244         [GFS2_LKS_SIRT]         = "sirt",
2245         [GFS2_LKS_SIRTVAR]      = "sirtvar",
2246         [GFS2_LKS_DCOUNT]       = "dlm",
2247         [GFS2_LKS_QCOUNT]       = "queue",
2248 };
2249
2250 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
2251
2252 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
2253 {
2254         struct gfs2_sbd *sdp = seq->private;
2255         loff_t pos = *(loff_t *)iter_ptr;
2256         unsigned index = pos >> 3;
2257         unsigned subindex = pos & 0x07;
2258         int i;
2259
2260         if (index == 0 && subindex != 0)
2261                 return 0;
2262
2263         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
2264                    (index == 0) ? "cpu": gfs2_stype[subindex]);
2265
2266         for_each_possible_cpu(i) {
2267                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
2268
2269                 if (index == 0)
2270                         seq_printf(seq, " %15u", i);
2271                 else
2272                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
2273                                    lkstats[index - 1].stats[subindex]);
2274         }
2275         seq_putc(seq, '\n');
2276         return 0;
2277 }
2278
2279 int __init gfs2_glock_init(void)
2280 {
2281         int i, ret;
2282
2283         ret = rhashtable_init(&gl_hash_table, &ht_parms);
2284         if (ret < 0)
2285                 return ret;
2286
2287         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
2288                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
2289         if (!glock_workqueue) {
2290                 rhashtable_destroy(&gl_hash_table);
2291                 return -ENOMEM;
2292         }
2293         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
2294                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
2295                                                 0);
2296         if (!gfs2_delete_workqueue) {
2297                 destroy_workqueue(glock_workqueue);
2298                 rhashtable_destroy(&gl_hash_table);
2299                 return -ENOMEM;
2300         }
2301
2302         ret = register_shrinker(&glock_shrinker);
2303         if (ret) {
2304                 destroy_workqueue(gfs2_delete_workqueue);
2305                 destroy_workqueue(glock_workqueue);
2306                 rhashtable_destroy(&gl_hash_table);
2307                 return ret;
2308         }
2309
2310         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
2311                 init_waitqueue_head(glock_wait_table + i);
2312
2313         return 0;
2314 }
2315
2316 void gfs2_glock_exit(void)
2317 {
2318         unregister_shrinker(&glock_shrinker);
2319         rhashtable_destroy(&gl_hash_table);
2320         destroy_workqueue(glock_workqueue);
2321         destroy_workqueue(gfs2_delete_workqueue);
2322 }
2323
2324 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
2325 {
2326         struct gfs2_glock *gl = gi->gl;
2327
2328         if (gl) {
2329                 if (n == 0)
2330                         return;
2331                 if (!lockref_put_not_zero(&gl->gl_lockref))
2332                         gfs2_glock_queue_put(gl);
2333         }
2334         for (;;) {
2335                 gl = rhashtable_walk_next(&gi->hti);
2336                 if (IS_ERR_OR_NULL(gl)) {
2337                         if (gl == ERR_PTR(-EAGAIN)) {
2338                                 n = 1;
2339                                 continue;
2340                         }
2341                         gl = NULL;
2342                         break;
2343                 }
2344                 if (gl->gl_name.ln_sbd != gi->sdp)
2345                         continue;
2346                 if (n <= 1) {
2347                         if (!lockref_get_not_dead(&gl->gl_lockref))
2348                                 continue;
2349                         break;
2350                 } else {
2351                         if (__lockref_is_dead(&gl->gl_lockref))
2352                                 continue;
2353                         n--;
2354                 }
2355         }
2356         gi->gl = gl;
2357 }
2358
2359 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
2360         __acquires(RCU)
2361 {
2362         struct gfs2_glock_iter *gi = seq->private;
2363         loff_t n;
2364
2365         /*
2366          * We can either stay where we are, skip to the next hash table
2367          * entry, or start from the beginning.
2368          */
2369         if (*pos < gi->last_pos) {
2370                 rhashtable_walk_exit(&gi->hti);
2371                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2372                 n = *pos + 1;
2373         } else {
2374                 n = *pos - gi->last_pos;
2375         }
2376
2377         rhashtable_walk_start(&gi->hti);
2378
2379         gfs2_glock_iter_next(gi, n);
2380         gi->last_pos = *pos;
2381         return gi->gl;
2382 }
2383
2384 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
2385                                  loff_t *pos)
2386 {
2387         struct gfs2_glock_iter *gi = seq->private;
2388
2389         (*pos)++;
2390         gi->last_pos = *pos;
2391         gfs2_glock_iter_next(gi, 1);
2392         return gi->gl;
2393 }
2394
2395 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2396         __releases(RCU)
2397 {
2398         struct gfs2_glock_iter *gi = seq->private;
2399
2400         rhashtable_walk_stop(&gi->hti);
2401 }
2402
2403 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2404 {
2405         dump_glock(seq, iter_ptr, false);
2406         return 0;
2407 }
2408
2409 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2410 {
2411         preempt_disable();
2412         if (*pos >= GFS2_NR_SBSTATS)
2413                 return NULL;
2414         return pos;
2415 }
2416
2417 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2418                                    loff_t *pos)
2419 {
2420         (*pos)++;
2421         if (*pos >= GFS2_NR_SBSTATS)
2422                 return NULL;
2423         return pos;
2424 }
2425
2426 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2427 {
2428         preempt_enable();
2429 }
2430
2431 static const struct seq_operations gfs2_glock_seq_ops = {
2432         .start = gfs2_glock_seq_start,
2433         .next  = gfs2_glock_seq_next,
2434         .stop  = gfs2_glock_seq_stop,
2435         .show  = gfs2_glock_seq_show,
2436 };
2437
2438 static const struct seq_operations gfs2_glstats_seq_ops = {
2439         .start = gfs2_glock_seq_start,
2440         .next  = gfs2_glock_seq_next,
2441         .stop  = gfs2_glock_seq_stop,
2442         .show  = gfs2_glstats_seq_show,
2443 };
2444
2445 static const struct seq_operations gfs2_sbstats_sops = {
2446         .start = gfs2_sbstats_seq_start,
2447         .next  = gfs2_sbstats_seq_next,
2448         .stop  = gfs2_sbstats_seq_stop,
2449         .show  = gfs2_sbstats_seq_show,
2450 };
2451
2452 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2453
2454 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2455                               const struct seq_operations *ops)
2456 {
2457         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2458         if (ret == 0) {
2459                 struct seq_file *seq = file->private_data;
2460                 struct gfs2_glock_iter *gi = seq->private;
2461
2462                 gi->sdp = inode->i_private;
2463                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2464                 if (seq->buf)
2465                         seq->size = GFS2_SEQ_GOODSIZE;
2466                 /*
2467                  * Initially, we are "before" the first hash table entry; the
2468                  * first call to rhashtable_walk_next gets us the first entry.
2469                  */
2470                 gi->last_pos = -1;
2471                 gi->gl = NULL;
2472                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2473         }
2474         return ret;
2475 }
2476
2477 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2478 {
2479         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2480 }
2481
2482 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2483 {
2484         struct seq_file *seq = file->private_data;
2485         struct gfs2_glock_iter *gi = seq->private;
2486
2487         if (gi->gl)
2488                 gfs2_glock_put(gi->gl);
2489         rhashtable_walk_exit(&gi->hti);
2490         return seq_release_private(inode, file);
2491 }
2492
2493 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2494 {
2495         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2496 }
2497
2498 static const struct file_operations gfs2_glocks_fops = {
2499         .owner   = THIS_MODULE,
2500         .open    = gfs2_glocks_open,
2501         .read    = seq_read,
2502         .llseek  = seq_lseek,
2503         .release = gfs2_glocks_release,
2504 };
2505
2506 static const struct file_operations gfs2_glstats_fops = {
2507         .owner   = THIS_MODULE,
2508         .open    = gfs2_glstats_open,
2509         .read    = seq_read,
2510         .llseek  = seq_lseek,
2511         .release = gfs2_glocks_release,
2512 };
2513
2514 DEFINE_SEQ_ATTRIBUTE(gfs2_sbstats);
2515
2516 void gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2517 {
2518         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2519
2520         debugfs_create_file("glocks", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2521                             &gfs2_glocks_fops);
2522
2523         debugfs_create_file("glstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2524                             &gfs2_glstats_fops);
2525
2526         debugfs_create_file("sbstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2527                             &gfs2_sbstats_fops);
2528 }
2529
2530 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2531 {
2532         debugfs_remove_recursive(sdp->debugfs_dir);
2533         sdp->debugfs_dir = NULL;
2534 }
2535
2536 void gfs2_register_debugfs(void)
2537 {
2538         gfs2_root = debugfs_create_dir("gfs2", NULL);
2539 }
2540
2541 void gfs2_unregister_debugfs(void)
2542 {
2543         debugfs_remove(gfs2_root);
2544         gfs2_root = NULL;
2545 }