fs/gfs2/glock.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   5  */
   6
   7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/sched.h>
  10 #include <linux/slab.h>
  11 #include <linux/spinlock.h>
  12 #include <linux/buffer_head.h>
  13 #include <linux/delay.h>
  14 #include <linux/sort.h>
  15 #include <linux/hash.h>
  16 #include <linux/jhash.h>
  17 #include <linux/kallsyms.h>
  18 #include <linux/gfs2_ondisk.h>
  19 #include <linux/list.h>
  20 #include <linux/wait.h>
  21 #include <linux/module.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/debugfs.h>
  25 #include <linux/kthread.h>
  26 #include <linux/freezer.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/jiffies.h>
  29 #include <linux/rcupdate.h>
  30 #include <linux/rculist_bl.h>
  31 #include <linux/bit_spinlock.h>
  32 #include <linux/percpu.h>
  33 #include <linux/list_sort.h>
  34 #include <linux/lockref.h>
  35 #include <linux/rhashtable.h>
  36
  37 #include "gfs2.h"
  38 #include "incore.h"
  39 #include "glock.h"
  40 #include "glops.h"
  41 #include "inode.h"
  42 #include "lops.h"
  43 #include "meta_io.h"
  44 #include "quota.h"
  45 #include "super.h"
  46 #include "util.h"
  47 #include "bmap.h"
  48 #define CREATE_TRACE_POINTS
  49 #include "trace_gfs2.h"
  50
  51 struct gfs2_glock_iter {
  52         struct gfs2_sbd *sdp;           /* incore superblock           */
  53         struct rhashtable_iter hti;     /* rhashtable iterator         */
  54         struct gfs2_glock *gl;          /* current glock struct        */
  55         loff_t last_pos;                /* last position               */
  56 };
  57
  58 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  59
  60 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  61
  62 static struct dentry *gfs2_root;
  63 static struct workqueue_struct *glock_workqueue;
  64 struct workqueue_struct *gfs2_delete_workqueue;
  65 static LIST_HEAD(lru_list);
  66 static atomic_t lru_count = ATOMIC_INIT(0);
  67 static DEFINE_SPINLOCK(lru_lock);
  68
  69 #define GFS2_GL_HASH_SHIFT      15
  70 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
  71
  72 static const struct rhashtable_params ht_parms = {
  73         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
  74         .key_len = offsetofend(struct lm_lockname, ln_type),
  75         .key_offset = offsetof(struct gfs2_glock, gl_name),
  76         .head_offset = offsetof(struct gfs2_glock, gl_node),
  77 };
  78
  79 static struct rhashtable gl_hash_table;
  80
  81 #define GLOCK_WAIT_TABLE_BITS 12
  82 #define GLOCK_WAIT_TABLE_SIZE (1 << GLOCK_WAIT_TABLE_BITS)
  83 static wait_queue_head_t glock_wait_table[GLOCK_WAIT_TABLE_SIZE] __cacheline_aligned;
  84
  85 struct wait_glock_queue {
  86         struct lm_lockname *name;
  87         wait_queue_entry_t wait;
  88 };
  89
  90 static int glock_wake_function(wait_queue_entry_t *wait, unsigned int mode,
  91                                int sync, void *key)
  92 {
  93         struct wait_glock_queue *wait_glock =
  94                 container_of(wait, struct wait_glock_queue, wait);
  95         struct lm_lockname *wait_name = wait_glock->name;
  96         struct lm_lockname *wake_name = key;
  97
  98         if (wake_name->ln_sbd != wait_name->ln_sbd ||
  99             wake_name->ln_number != wait_name->ln_number ||
 100             wake_name->ln_type != wait_name->ln_type)
 101                 return 0;
 102         return autoremove_wake_function(wait, mode, sync, key);
 103 }
 104
 105 static wait_queue_head_t *glock_waitqueue(struct lm_lockname *name)
 106 {
 107         u32 hash = jhash2((u32 *)name, ht_parms.key_len / 4, 0);
 108
 109         return glock_wait_table + hash_32(hash, GLOCK_WAIT_TABLE_BITS);
 110 }
 111
 112 /**
 113  * wake_up_glock  -  Wake up waiters on a glock
 114  * @gl: the glock
 115  */
 116 static void wake_up_glock(struct gfs2_glock *gl)
 117 {
 118         wait_queue_head_t *wq = glock_waitqueue(&gl->gl_name);
 119
 120         if (waitqueue_active(wq))
 121                 __wake_up(wq, TASK_NORMAL, 1, &gl->gl_name);
 122 }
 123
 124 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 125 {
 126         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 127
 128         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
 129                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 130         } else {
 131                 kfree(gl->gl_lksb.sb_lvbptr);
 132                 kmem_cache_free(gfs2_glock_cachep, gl);
 133         }
 134 }
 135
 136 void gfs2_glock_free(struct gfs2_glock *gl)
 137 {
 138         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 139
 140         BUG_ON(atomic_read(&gl->gl_revokes));
 141         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 142         smp_mb();
 143         wake_up_glock(gl);
 144         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 145         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 146                 wake_up(&sdp->sd_glock_wait);
 147 }
 148
 149 /**
 150  * gfs2_glock_hold() - increment reference count on glock
 151  * @gl: The glock to hold
 152  *
 153  */
 154
 155 void gfs2_glock_hold(struct gfs2_glock *gl)
 156 {
 157         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 158         lockref_get(&gl->gl_lockref);
 159 }
 160
 161 /**
 162  * demote_ok - Check to see if it's ok to unlock a glock
 163  * @gl: the glock
 164  *
 165  * Returns: 1 if it's ok
 166  */
 167
 168 static int demote_ok(const struct gfs2_glock *gl)
 169 {
 170         const struct gfs2_glock_operations *glops = gl->gl_ops;
 171
 172         if (gl->gl_state == LM_ST_UNLOCKED)
 173                 return 0;
 174         if (!list_empty(&gl->gl_holders))
 175                 return 0;
 176         if (glops->go_demote_ok)
 177                 return glops->go_demote_ok(gl);
 178         return 1;
 179 }
 180
 181
 182 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 183 {
 184         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 185                 return;
 186
 187         spin_lock(&lru_lock);
 188
 189         list_del(&gl->gl_lru);
 190         list_add_tail(&gl->gl_lru, &lru_list);
 191
 192         if (!test_bit(GLF_LRU, &gl->gl_flags)) {
 193                 set_bit(GLF_LRU, &gl->gl_flags);
 194                 atomic_inc(&lru_count);
 195         }
 196
 197         spin_unlock(&lru_lock);
 198 }
 199
 200 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 201 {
 202         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 203                 return;
 204
 205         spin_lock(&lru_lock);
 206         if (test_bit(GLF_LRU, &gl->gl_flags)) {
 207                 list_del_init(&gl->gl_lru);
 208                 atomic_dec(&lru_count);
 209                 clear_bit(GLF_LRU, &gl->gl_flags);
 210         }
 211         spin_unlock(&lru_lock);
 212 }
 213
 214 /*
 215  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 216  * work queue.
 217  */
 218 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 219         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 220                 /*
 221                  * We are holding the lockref spinlock, and the work was still
 222                  * queued above.  The queued work (glock_work_func) takes that
 223                  * spinlock before dropping its glock reference(s), so it
 224                  * cannot have dropped them in the meantime.
 225                  */
 226                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 227                 gl->gl_lockref.count--;
 228         }
 229 }
 230
 231 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 232         spin_lock(&gl->gl_lockref.lock);
 233         __gfs2_glock_queue_work(gl, delay);
 234         spin_unlock(&gl->gl_lockref.lock);
 235 }
 236
 237 static void __gfs2_glock_put(struct gfs2_glock *gl)
 238 {
 239         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 240         struct address_space *mapping = gfs2_glock2aspace(gl);
 241
 242         lockref_mark_dead(&gl->gl_lockref);
 243
 244         gfs2_glock_remove_from_lru(gl);
 245         spin_unlock(&gl->gl_lockref.lock);
 246         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 247         GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
 248         trace_gfs2_glock_put(gl);
 249         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 250 }
 251
 252 /*
 253  * Cause the glock to be put in work queue context.
 254  */
 255 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 256 {
 257         gfs2_glock_queue_work(gl, 0);
 258 }
 259
 260 /**
 261  * gfs2_glock_put() - Decrement reference count on glock
 262  * @gl: The glock to put
 263  *
 264  */
 265
 266 void gfs2_glock_put(struct gfs2_glock *gl)
 267 {
 268         if (lockref_put_or_lock(&gl->gl_lockref))
 269                 return;
 270
 271         __gfs2_glock_put(gl);
 272 }
 273
 274 /**
 275  * may_grant - check if its ok to grant a new lock
 276  * @gl: The glock
 277  * @gh: The lock request which we wish to grant
 278  *
 279  * Returns: true if its ok to grant the lock
 280  */
 281
 282 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 283 {
 284         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 285         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 286              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 287                 return 0;
 288         if (gl->gl_state == gh->gh_state)
 289                 return 1;
 290         if (gh->gh_flags & GL_EXACT)
 291                 return 0;
 292         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 293                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 294                         return 1;
 295                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 296                         return 1;
 297         }
 298         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 299                 return 1;
 300         return 0;
 301 }
 302
 303 static void gfs2_holder_wake(struct gfs2_holder *gh)
 304 {
 305         clear_bit(HIF_WAIT, &gh->gh_iflags);
 306         smp_mb__after_atomic();
 307         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 308         if (gh->gh_flags & GL_ASYNC) {
 309                 struct gfs2_sbd *sdp = gh->gh_gl->gl_name.ln_sbd;
 310
 311                 wake_up(&sdp->sd_async_glock_wait);
 312         }
 313 }
 314
 315 /**
 316  * do_error - Something unexpected has happened during a lock request
 317  *
 318  */
 319
 320 static void do_error(struct gfs2_glock *gl, const int ret)
 321 {
 322         struct gfs2_holder *gh, *tmp;
 323
 324         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 325                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 326                         continue;
 327                 if (ret & LM_OUT_ERROR)
 328                         gh->gh_error = -EIO;
 329                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 330                         gh->gh_error = GLR_TRYFAILED;
 331                 else
 332                         continue;
 333                 list_del_init(&gh->gh_list);
 334                 trace_gfs2_glock_queue(gh, 0);
 335                 gfs2_holder_wake(gh);
 336         }
 337 }
 338
 339 /**
 340  * do_promote - promote as many requests as possible on the current queue
 341  * @gl: The glock
 342  *
 343  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 344  *          if a type specific operation is underway.
 345  */
 346
 347 static int do_promote(struct gfs2_glock *gl)
 348 __releases(&gl->gl_lockref.lock)
 349 __acquires(&gl->gl_lockref.lock)
 350 {
 351         const struct gfs2_glock_operations *glops = gl->gl_ops;
 352         struct gfs2_holder *gh, *tmp;
 353         int ret;
 354
 355 restart:
 356         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 357                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 358                         continue;
 359                 if (may_grant(gl, gh)) {
 360                         if (gh->gh_list.prev == &gl->gl_holders &&
 361                             glops->go_lock) {
 362                                 spin_unlock(&gl->gl_lockref.lock);
 363                                 /* FIXME: eliminate this eventually */
 364                                 ret = glops->go_lock(gh);
 365                                 spin_lock(&gl->gl_lockref.lock);
 366                                 if (ret) {
 367                                         if (ret == 1)
 368                                                 return 2;
 369                                         gh->gh_error = ret;
 370                                         list_del_init(&gh->gh_list);
 371                                         trace_gfs2_glock_queue(gh, 0);
 372                                         gfs2_holder_wake(gh);
 373                                         goto restart;
 374                                 }
 375                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 376                                 trace_gfs2_promote(gh, 1);
 377                                 gfs2_holder_wake(gh);
 378                                 goto restart;
 379                         }
 380                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 381                         trace_gfs2_promote(gh, 0);
 382                         gfs2_holder_wake(gh);
 383                         continue;
 384                 }
 385                 if (gh->gh_list.prev == &gl->gl_holders)
 386                         return 1;
 387                 do_error(gl, 0);
 388                 break;
 389         }
 390         return 0;
 391 }
 392
 393 /**
 394  * find_first_waiter - find the first gh that's waiting for the glock
 395  * @gl: the glock
 396  */
 397
 398 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 399 {
 400         struct gfs2_holder *gh;
 401
 402         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 403                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 404                         return gh;
 405         }
 406         return NULL;
 407 }
 408
 409 /**
 410  * state_change - record that the glock is now in a different state
 411  * @gl: the glock
 412  * @new_state the new state
 413  *
 414  */
 415
 416 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 417 {
 418         int held1, held2;
 419
 420         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 421         held2 = (new_state != LM_ST_UNLOCKED);
 422
 423         if (held1 != held2) {
 424                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 425                 if (held2)
 426                         gl->gl_lockref.count++;
 427                 else
 428                         gl->gl_lockref.count--;
 429         }
 430         if (held1 && held2 && list_empty(&gl->gl_holders))
 431                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 432
 433         if (new_state != gl->gl_target)
 434                 /* shorten our minimum hold time */
 435                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 436                                        GL_GLOCK_MIN_HOLD);
 437         gl->gl_state = new_state;
 438         gl->gl_tchange = jiffies;
 439 }
 440
 441 static void gfs2_demote_wake(struct gfs2_glock *gl)
 442 {
 443         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 444         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 445         smp_mb__after_atomic();
 446         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 447 }
 448
 449 /**
 450  * finish_xmote - The DLM has replied to one of our lock requests
 451  * @gl: The glock
 452  * @ret: The status from the DLM
 453  *
 454  */
 455
 456 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 457 {
 458         const struct gfs2_glock_operations *glops = gl->gl_ops;
 459         struct gfs2_holder *gh;
 460         unsigned state = ret & LM_OUT_ST_MASK;
 461         int rv;
 462
 463         spin_lock(&gl->gl_lockref.lock);
 464         trace_gfs2_glock_state_change(gl, state);
 465         state_change(gl, state);
 466         gh = find_first_waiter(gl);
 467
 468         /* Demote to UN request arrived during demote to SH or DF */
 469         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 470             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 471                 gl->gl_target = LM_ST_UNLOCKED;
 472
 473         /* Check for state != intended state */
 474         if (unlikely(state != gl->gl_target)) {
 475                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 476                         /* move to back of queue and try next entry */
 477                         if (ret & LM_OUT_CANCELED) {
 478                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 479                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 480                                 gh = find_first_waiter(gl);
 481                                 gl->gl_target = gh->gh_state;
 482                                 goto retry;
 483                         }
 484                         /* Some error or failed "try lock" - report it */
 485                         if ((ret & LM_OUT_ERROR) ||
 486                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 487                                 gl->gl_target = gl->gl_state;
 488                                 do_error(gl, ret);
 489                                 goto out;
 490                         }
 491                 }
 492                 switch(state) {
 493                 /* Unlocked due to conversion deadlock, try again */
 494                 case LM_ST_UNLOCKED:
 495 retry:
 496                         do_xmote(gl, gh, gl->gl_target);
 497                         break;
 498                 /* Conversion fails, unlock and try again */
 499                 case LM_ST_SHARED:
 500                 case LM_ST_DEFERRED:
 501                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 502                         break;
 503                 default: /* Everything else */
 504                         fs_err(gl->gl_name.ln_sbd, "wanted %u got %u\n",
 505                                gl->gl_target, state);
 506                         GLOCK_BUG_ON(gl, 1);
 507                 }
 508                 spin_unlock(&gl->gl_lockref.lock);
 509                 return;
 510         }
 511
 512         /* Fast path - we got what we asked for */
 513         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 514                 gfs2_demote_wake(gl);
 515         if (state != LM_ST_UNLOCKED) {
 516                 if (glops->go_xmote_bh) {
 517                         spin_unlock(&gl->gl_lockref.lock);
 518                         rv = glops->go_xmote_bh(gl, gh);
 519                         spin_lock(&gl->gl_lockref.lock);
 520                         if (rv) {
 521                                 do_error(gl, rv);
 522                                 goto out;
 523                         }
 524                 }
 525                 rv = do_promote(gl);
 526                 if (rv == 2)
 527                         goto out_locked;
 528         }
 529 out:
 530         clear_bit(GLF_LOCK, &gl->gl_flags);
 531 out_locked:
 532         spin_unlock(&gl->gl_lockref.lock);
 533 }
 534
 535 /**
 536  * do_xmote - Calls the DLM to change the state of a lock
 537  * @gl: The lock state
 538  * @gh: The holder (only for promotes)
 539  * @target: The target lock state
 540  *
 541  */
 542
 543 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 544 __releases(&gl->gl_lockref.lock)
 545 __acquires(&gl->gl_lockref.lock)
 546 {
 547         const struct gfs2_glock_operations *glops = gl->gl_ops;
 548         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 549         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 550         int ret;
 551
 552         if (unlikely(test_bit(SDF_WITHDRAWN, &sdp->sd_flags)) &&
 553             target != LM_ST_UNLOCKED)
 554                 return;
 555         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 556                       LM_FLAG_PRIORITY);
 557         GLOCK_BUG_ON(gl, gl->gl_state == target);
 558         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 559         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 560             glops->go_inval) {
 561                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 562                 do_error(gl, 0); /* Fail queued try locks */
 563         }
 564         gl->gl_req = target;
 565         set_bit(GLF_BLOCKING, &gl->gl_flags);
 566         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 567             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 568             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 569                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 570         spin_unlock(&gl->gl_lockref.lock);
 571         if (glops->go_sync)
 572                 glops->go_sync(gl);
 573         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 574                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 575         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 576
 577         gfs2_glock_hold(gl);
 578         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 579                 /* lock_dlm */
 580                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 581                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 582                     target == LM_ST_UNLOCKED &&
 583                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 584                         finish_xmote(gl, target);
 585                         gfs2_glock_queue_work(gl, 0);
 586                 }
 587                 else if (ret) {
 588                         fs_err(sdp, "lm_lock ret %d\n", ret);
 589                         GLOCK_BUG_ON(gl, !test_bit(SDF_WITHDRAWN,
 590                                                    &sdp->sd_flags));
 591                 }
 592         } else { /* lock_nolock */
 593                 finish_xmote(gl, target);
 594                 gfs2_glock_queue_work(gl, 0);
 595         }
 596
 597         spin_lock(&gl->gl_lockref.lock);
 598 }
 599
 600 /**
 601  * find_first_holder - find the first "holder" gh
 602  * @gl: the glock
 603  */
 604
 605 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 606 {
 607         struct gfs2_holder *gh;
 608
 609         if (!list_empty(&gl->gl_holders)) {
 610                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 611                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 612                         return gh;
 613         }
 614         return NULL;
 615 }
 616
 617 /**
 618  * run_queue - do all outstanding tasks related to a glock
 619  * @gl: The glock in question
 620  * @nonblock: True if we must not block in run_queue
 621  *
 622  */
 623
 624 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 625 __releases(&gl->gl_lockref.lock)
 626 __acquires(&gl->gl_lockref.lock)
 627 {
 628         struct gfs2_holder *gh = NULL;
 629         int ret;
 630
 631         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 632                 return;
 633
 634         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 635
 636         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 637             gl->gl_demote_state != gl->gl_state) {
 638                 if (find_first_holder(gl))
 639                         goto out_unlock;
 640                 if (nonblock)
 641                         goto out_sched;
 642                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 643                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 644                 gl->gl_target = gl->gl_demote_state;
 645         } else {
 646                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 647                         gfs2_demote_wake(gl);
 648                 ret = do_promote(gl);
 649                 if (ret == 0)
 650                         goto out_unlock;
 651                 if (ret == 2)
 652                         goto out;
 653                 gh = find_first_waiter(gl);
 654                 gl->gl_target = gh->gh_state;
 655                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 656                         do_error(gl, 0); /* Fail queued try locks */
 657         }
 658         do_xmote(gl, gh, gl->gl_target);
 659 out:
 660         return;
 661
 662 out_sched:
 663         clear_bit(GLF_LOCK, &gl->gl_flags);
 664         smp_mb__after_atomic();
 665         gl->gl_lockref.count++;
 666         __gfs2_glock_queue_work(gl, 0);
 667         return;
 668
 669 out_unlock:
 670         clear_bit(GLF_LOCK, &gl->gl_flags);
 671         smp_mb__after_atomic();
 672         return;
 673 }
 674
 675 static void delete_work_func(struct work_struct *work)
 676 {
 677         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 678         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 679         struct inode *inode;
 680         u64 no_addr = gl->gl_name.ln_number;
 681
 682         /* If someone's using this glock to create a new dinode, the block must
 683            have been freed by another node, then re-used, in which case our
 684            iopen callback is too late after the fact. Ignore it. */
 685         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 686                 goto out;
 687
 688         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 689         if (!IS_ERR_OR_NULL(inode)) {
 690                 d_prune_aliases(inode);
 691                 iput(inode);
 692         }
 693 out:
 694         gfs2_glock_put(gl);
 695 }
 696
 697 static void glock_work_func(struct work_struct *work)
 698 {
 699         unsigned long delay = 0;
 700         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 701         unsigned int drop_refs = 1;
 702
 703         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 704                 finish_xmote(gl, gl->gl_reply);
 705                 drop_refs++;
 706         }
 707         spin_lock(&gl->gl_lockref.lock);
 708         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 709             gl->gl_state != LM_ST_UNLOCKED &&
 710             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 711                 unsigned long holdtime, now = jiffies;
 712
 713                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 714                 if (time_before(now, holdtime))
 715                         delay = holdtime - now;
 716
 717                 if (!delay) {
 718                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 719                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 720                 }
 721         }
 722         run_queue(gl, 0);
 723         if (delay) {
 724                 /* Keep one glock reference for the work we requeue. */
 725                 drop_refs--;
 726                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 727                         delay = 0;
 728                 __gfs2_glock_queue_work(gl, delay);
 729         }
 730
 731         /*
 732          * Drop the remaining glock references manually here. (Mind that
 733          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 734          * here as well.)
 735          */
 736         gl->gl_lockref.count -= drop_refs;
 737         if (!gl->gl_lockref.count) {
 738                 __gfs2_glock_put(gl);
 739                 return;
 740         }
 741         spin_unlock(&gl->gl_lockref.lock);
 742 }
 743
 744 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 745                                             struct gfs2_glock *new)
 746 {
 747         struct wait_glock_queue wait;
 748         wait_queue_head_t *wq = glock_waitqueue(name);
 749         struct gfs2_glock *gl;
 750
 751         wait.name = name;
 752         init_wait(&wait.wait);
 753         wait.wait.func = glock_wake_function;
 754
 755 again:
 756         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 757         rcu_read_lock();
 758         if (new) {
 759                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 760                         &new->gl_node, ht_parms);
 761                 if (IS_ERR(gl))
 762                         goto out;
 763         } else {
 764                 gl = rhashtable_lookup_fast(&gl_hash_table,
 765                         name, ht_parms);
 766         }
 767         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 768                 rcu_read_unlock();
 769                 schedule();
 770                 goto again;
 771         }
 772 out:
 773         rcu_read_unlock();
 774         finish_wait(wq, &wait.wait);
 775         return gl;
 776 }
 777
 778 /**
 779  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 780  * @sdp: The GFS2 superblock
 781  * @number: the lock number
 782  * @glops: The glock_operations to use
 783  * @create: If 0, don't create the glock if it doesn't exist
 784  * @glp: the glock is returned here
 785  *
 786  * This does not lock a glock, just finds/creates structures for one.
 787  *
 788  * Returns: errno
 789  */
 790
 791 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 792                    const struct gfs2_glock_operations *glops, int create,
 793                    struct gfs2_glock **glp)
 794 {
 795         struct super_block *s = sdp->sd_vfs;
 796         struct lm_lockname name = { .ln_number = number,
 797                                     .ln_type = glops->go_type,
 798                                     .ln_sbd = sdp };
 799         struct gfs2_glock *gl, *tmp;
 800         struct address_space *mapping;
 801         struct kmem_cache *cachep;
 802         int ret = 0;
 803
 804         gl = find_insert_glock(&name, NULL);
 805         if (gl) {
 806                 *glp = gl;
 807                 return 0;
 808         }
 809         if (!create)
 810                 return -ENOENT;
 811
 812         if (glops->go_flags & GLOF_ASPACE)
 813                 cachep = gfs2_glock_aspace_cachep;
 814         else
 815                 cachep = gfs2_glock_cachep;
 816         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 817         if (!gl)
 818                 return -ENOMEM;
 819
 820         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 821
 822         if (glops->go_flags & GLOF_LVB) {
 823                 gl->gl_lksb.sb_lvbptr = kzalloc(GFS2_MIN_LVB_SIZE, GFP_NOFS);
 824                 if (!gl->gl_lksb.sb_lvbptr) {
 825                         kmem_cache_free(cachep, gl);
 826                         return -ENOMEM;
 827                 }
 828         }
 829
 830         atomic_inc(&sdp->sd_glock_disposal);
 831         gl->gl_node.next = NULL;
 832         gl->gl_flags = 0;
 833         gl->gl_name = name;
 834         gl->gl_lockref.count = 1;
 835         gl->gl_state = LM_ST_UNLOCKED;
 836         gl->gl_target = LM_ST_UNLOCKED;
 837         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 838         gl->gl_ops = glops;
 839         gl->gl_dstamp = 0;
 840         preempt_disable();
 841         /* We use the global stats to estimate the initial per-glock stats */
 842         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 843         preempt_enable();
 844         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 845         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 846         gl->gl_tchange = jiffies;
 847         gl->gl_object = NULL;
 848         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 849         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 850         INIT_WORK(&gl->gl_delete, delete_work_func);
 851
 852         mapping = gfs2_glock2aspace(gl);
 853         if (mapping) {
 854                 mapping->a_ops = &gfs2_meta_aops;
 855                 mapping->host = s->s_bdev->bd_inode;
 856                 mapping->flags = 0;
 857                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 858                 mapping->private_data = NULL;
 859                 mapping->writeback_index = 0;
 860         }
 861
 862         tmp = find_insert_glock(&name, gl);
 863         if (!tmp) {
 864                 *glp = gl;
 865                 goto out;
 866         }
 867         if (IS_ERR(tmp)) {
 868                 ret = PTR_ERR(tmp);
 869                 goto out_free;
 870         }
 871         *glp = tmp;
 872
 873 out_free:
 874         kfree(gl->gl_lksb.sb_lvbptr);
 875         kmem_cache_free(cachep, gl);
 876         atomic_dec(&sdp->sd_glock_disposal);
 877
 878 out:
 879         return ret;
 880 }
 881
 882 /**
 883  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 884  * @gl: the glock
 885  * @state: the state we're requesting
 886  * @flags: the modifier flags
 887  * @gh: the holder structure
 888  *
 889  */
 890
 891 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 892                       struct gfs2_holder *gh)
 893 {
 894         INIT_LIST_HEAD(&gh->gh_list);
 895         gh->gh_gl = gl;
 896         gh->gh_ip = _RET_IP_;
 897         gh->gh_owner_pid = get_pid(task_pid(current));
 898         gh->gh_state = state;
 899         gh->gh_flags = flags;
 900         gh->gh_error = 0;
 901         gh->gh_iflags = 0;
 902         gfs2_glock_hold(gl);
 903 }
 904
 905 /**
 906  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 907  * @state: the state we're requesting
 908  * @flags: the modifier flags
 909  * @gh: the holder structure
 910  *
 911  * Don't mess with the glock.
 912  *
 913  */
 914
 915 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
 916 {
 917         gh->gh_state = state;
 918         gh->gh_flags = flags;
 919         gh->gh_iflags = 0;
 920         gh->gh_ip = _RET_IP_;
 921         put_pid(gh->gh_owner_pid);
 922         gh->gh_owner_pid = get_pid(task_pid(current));
 923 }
 924
 925 /**
 926  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 927  * @gh: the holder structure
 928  *
 929  */
 930
 931 void gfs2_holder_uninit(struct gfs2_holder *gh)
 932 {
 933         put_pid(gh->gh_owner_pid);
 934         gfs2_glock_put(gh->gh_gl);
 935         gfs2_holder_mark_uninitialized(gh);
 936         gh->gh_ip = 0;
 937 }
 938
 939 static void gfs2_glock_update_hold_time(struct gfs2_glock *gl,
 940                                         unsigned long start_time)
 941 {
 942         /* Have we waited longer that a second? */
 943         if (time_after(jiffies, start_time + HZ)) {
 944                 /* Lengthen the minimum hold time. */
 945                 gl->gl_hold_time = min(gl->gl_hold_time + GL_GLOCK_HOLD_INCR,
 946                                        GL_GLOCK_MAX_HOLD);
 947         }
 948 }
 949
 950 /**
 951  * gfs2_glock_wait - wait on a glock acquisition
 952  * @gh: the glock holder
 953  *
 954  * Returns: 0 on success
 955  */
 956
 957 int gfs2_glock_wait(struct gfs2_holder *gh)
 958 {
 959         unsigned long start_time = jiffies;
 960
 961         might_sleep();
 962         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
 963         gfs2_glock_update_hold_time(gh->gh_gl, start_time);
 964         return gh->gh_error;
 965 }
 966
 967 static int glocks_pending(unsigned int num_gh, struct gfs2_holder *ghs)
 968 {
 969         int i;
 970
 971         for (i = 0; i < num_gh; i++)
 972                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags))
 973                         return 1;
 974         return 0;
 975 }
 976
 977 /**
 978  * gfs2_glock_async_wait - wait on multiple asynchronous glock acquisitions
 979  * @num_gh: the number of holders in the array
 980  * @ghs: the glock holder array
 981  *
 982  * Returns: 0 on success, meaning all glocks have been granted and are held.
 983  *          -ESTALE if the request timed out, meaning all glocks were released,
 984  *          and the caller should retry the operation.
 985  */
 986
 987 int gfs2_glock_async_wait(unsigned int num_gh, struct gfs2_holder *ghs)
 988 {
 989         struct gfs2_sbd *sdp = ghs[0].gh_gl->gl_name.ln_sbd;
 990         int i, ret = 0, timeout = 0;
 991         unsigned long start_time = jiffies;
 992         bool keep_waiting;
 993
 994         might_sleep();
 995         /*
 996          * Total up the (minimum hold time * 2) of all glocks and use that to
 997          * determine the max amount of time we should wait.
 998          */
 999         for (i = 0; i < num_gh; i++)
1000                 timeout += ghs[i].gh_gl->gl_hold_time << 1;
1001
1002 wait_for_dlm:
1003         if (!wait_event_timeout(sdp->sd_async_glock_wait,
1004                                 !glocks_pending(num_gh, ghs), timeout))
1005                 ret = -ESTALE; /* request timed out. */
1006
1007         /*
1008          * If dlm granted all our requests, we need to adjust the glock
1009          * minimum hold time values according to how long we waited.
1010          *
1011          * If our request timed out, we need to repeatedly release any held
1012          * glocks we acquired thus far to allow dlm to acquire the remaining
1013          * glocks without deadlocking.  We cannot currently cancel outstanding
1014          * glock acquisitions.
1015          *
1016          * The HIF_WAIT bit tells us which requests still need a response from
1017          * dlm.
1018          *
1019          * If dlm sent us any errors, we return the first error we find.
1020          */
1021         keep_waiting = false;
1022         for (i = 0; i < num_gh; i++) {
1023                 /* Skip holders we have already dequeued below. */
1024                 if (!gfs2_holder_queued(&ghs[i]))
1025                         continue;
1026                 /* Skip holders with a pending DLM response. */
1027                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags)) {
1028                         keep_waiting = true;
1029                         continue;
1030                 }
1031
1032                 if (test_bit(HIF_HOLDER, &ghs[i].gh_iflags)) {
1033                         if (ret == -ESTALE)
1034                                 gfs2_glock_dq(&ghs[i]);
1035                         else
1036                                 gfs2_glock_update_hold_time(ghs[i].gh_gl,
1037                                                             start_time);
1038                 }
1039                 if (!ret)
1040                         ret = ghs[i].gh_error;
1041         }
1042
1043         if (keep_waiting)
1044                 goto wait_for_dlm;
1045
1046         /*
1047          * At this point, we've either acquired all locks or released them all.
1048          */
1049         return ret;
1050 }
1051
1052 /**
1053  * handle_callback - process a demote request
1054  * @gl: the glock
1055  * @state: the state the caller wants us to change to
1056  *
1057  * There are only two requests that we are going to see in actual
1058  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
1059  */
1060
1061 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
1062                             unsigned long delay, bool remote)
1063 {
1064         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
1065
1066         set_bit(bit, &gl->gl_flags);
1067         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
1068                 gl->gl_demote_state = state;
1069                 gl->gl_demote_time = jiffies;
1070         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
1071                         gl->gl_demote_state != state) {
1072                 gl->gl_demote_state = LM_ST_UNLOCKED;
1073         }
1074         if (gl->gl_ops->go_callback)
1075                 gl->gl_ops->go_callback(gl, remote);
1076         trace_gfs2_demote_rq(gl, remote);
1077 }
1078
1079 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
1080 {
1081         struct va_format vaf;
1082         va_list args;
1083
1084         va_start(args, fmt);
1085
1086         if (seq) {
1087                 seq_vprintf(seq, fmt, args);
1088         } else {
1089                 vaf.fmt = fmt;
1090                 vaf.va = &args;
1091
1092                 pr_err("%pV", &vaf);
1093         }
1094
1095         va_end(args);
1096 }
1097
1098 /**
1099  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1100  * @gh: the holder structure to add
1101  *
1102  * Eventually we should move the recursive locking trap to a
1103  * debugging option or something like that. This is the fast
1104  * path and needs to have the minimum number of distractions.
1105  *
1106  */
1107
1108 static inline void add_to_queue(struct gfs2_holder *gh)
1109 __releases(&gl->gl_lockref.lock)
1110 __acquires(&gl->gl_lockref.lock)
1111 {
1112         struct gfs2_glock *gl = gh->gh_gl;
1113         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1114         struct list_head *insert_pt = NULL;
1115         struct gfs2_holder *gh2;
1116         int try_futile = 0;
1117
1118         GLOCK_BUG_ON(gl, gh->gh_owner_pid == NULL);
1119         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1120                 GLOCK_BUG_ON(gl, true);
1121
1122         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1123                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1124                         try_futile = !may_grant(gl, gh);
1125                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1126                         goto fail;
1127         }
1128
1129         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1130                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1131                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1132                         goto trap_recursive;
1133                 if (try_futile &&
1134                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1135 fail:
1136                         gh->gh_error = GLR_TRYFAILED;
1137                         gfs2_holder_wake(gh);
1138                         return;
1139                 }
1140                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1141                         continue;
1142                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1143                         insert_pt = &gh2->gh_list;
1144         }
1145         set_bit(GLF_QUEUED, &gl->gl_flags);
1146         trace_gfs2_glock_queue(gh, 1);
1147         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1148         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1149         if (likely(insert_pt == NULL)) {
1150                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1151                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1152                         goto do_cancel;
1153                 return;
1154         }
1155         list_add_tail(&gh->gh_list, insert_pt);
1156 do_cancel:
1157         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1158         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1159                 spin_unlock(&gl->gl_lockref.lock);
1160                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1161                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1162                 spin_lock(&gl->gl_lockref.lock);
1163         }
1164         return;
1165
1166 trap_recursive:
1167         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1168         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1169         fs_err(sdp, "lock type: %d req lock state : %d\n",
1170                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1171         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1172         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1173         fs_err(sdp, "lock type: %d req lock state : %d\n",
1174                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1175         gfs2_dump_glock(NULL, gl, true);
1176         BUG();
1177 }
1178
1179 /**
1180  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1181  * @gh: the holder structure
1182  *
1183  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1184  *
1185  * Returns: 0, GLR_TRYFAILED, or errno on failure
1186  */
1187
1188 int gfs2_glock_nq(struct gfs2_holder *gh)
1189 {
1190         struct gfs2_glock *gl = gh->gh_gl;
1191         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1192         int error = 0;
1193
1194         if (unlikely(test_bit(SDF_WITHDRAWN, &sdp->sd_flags)))
1195                 return -EIO;
1196
1197         if (test_bit(GLF_LRU, &gl->gl_flags))
1198                 gfs2_glock_remove_from_lru(gl);
1199
1200         spin_lock(&gl->gl_lockref.lock);
1201         add_to_queue(gh);
1202         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1203                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1204                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1205                 gl->gl_lockref.count++;
1206                 __gfs2_glock_queue_work(gl, 0);
1207         }
1208         run_queue(gl, 1);
1209         spin_unlock(&gl->gl_lockref.lock);
1210
1211         if (!(gh->gh_flags & GL_ASYNC))
1212                 error = gfs2_glock_wait(gh);
1213
1214         return error;
1215 }
1216
1217 /**
1218  * gfs2_glock_poll - poll to see if an async request has been completed
1219  * @gh: the holder
1220  *
1221  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1222  */
1223
1224 int gfs2_glock_poll(struct gfs2_holder *gh)
1225 {
1226         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1227 }
1228
1229 /**
1230  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1231  * @gh: the glock holder
1232  *
1233  */
1234
1235 void gfs2_glock_dq(struct gfs2_holder *gh)
1236 {
1237         struct gfs2_glock *gl = gh->gh_gl;
1238         const struct gfs2_glock_operations *glops = gl->gl_ops;
1239         unsigned delay = 0;
1240         int fast_path = 0;
1241
1242         spin_lock(&gl->gl_lockref.lock);
1243         if (gh->gh_flags & GL_NOCACHE)
1244                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1245
1246         list_del_init(&gh->gh_list);
1247         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1248         if (find_first_holder(gl) == NULL) {
1249                 if (glops->go_unlock) {
1250                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1251                         spin_unlock(&gl->gl_lockref.lock);
1252                         glops->go_unlock(gh);
1253                         spin_lock(&gl->gl_lockref.lock);
1254                         clear_bit(GLF_LOCK, &gl->gl_flags);
1255                 }
1256                 if (list_empty(&gl->gl_holders) &&
1257                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1258                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1259                         fast_path = 1;
1260         }
1261         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1262                 gfs2_glock_add_to_lru(gl);
1263
1264         trace_gfs2_glock_queue(gh, 0);
1265         if (unlikely(!fast_path)) {
1266                 gl->gl_lockref.count++;
1267                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1268                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1269                     gl->gl_name.ln_type == LM_TYPE_INODE)
1270                         delay = gl->gl_hold_time;
1271                 __gfs2_glock_queue_work(gl, delay);
1272         }
1273         spin_unlock(&gl->gl_lockref.lock);
1274 }
1275
1276 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1277 {
1278         struct gfs2_glock *gl = gh->gh_gl;
1279         gfs2_glock_dq(gh);
1280         might_sleep();
1281         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1282 }
1283
1284 /**
1285  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1286  * @gh: the holder structure
1287  *
1288  */
1289
1290 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1291 {
1292         gfs2_glock_dq(gh);
1293         gfs2_holder_uninit(gh);
1294 }
1295
1296 /**
1297  * gfs2_glock_nq_num - acquire a glock based on lock number
1298  * @sdp: the filesystem
1299  * @number: the lock number
1300  * @glops: the glock operations for the type of glock
1301  * @state: the state to acquire the glock in
1302  * @flags: modifier flags for the acquisition
1303  * @gh: the struct gfs2_holder
1304  *
1305  * Returns: errno
1306  */
1307
1308 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1309                       const struct gfs2_glock_operations *glops,
1310                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1311 {
1312         struct gfs2_glock *gl;
1313         int error;
1314
1315         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1316         if (!error) {
1317                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1318                 gfs2_glock_put(gl);
1319         }
1320
1321         return error;
1322 }
1323
1324 /**
1325  * glock_compare - Compare two struct gfs2_glock structures for sorting
1326  * @arg_a: the first structure
1327  * @arg_b: the second structure
1328  *
1329  */
1330
1331 static int glock_compare(const void *arg_a, const void *arg_b)
1332 {
1333         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1334         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1335         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1336         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1337
1338         if (a->ln_number > b->ln_number)
1339                 return 1;
1340         if (a->ln_number < b->ln_number)
1341                 return -1;
1342         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1343         return 0;
1344 }
1345
1346 /**
1347  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1348  * @num_gh: the number of structures
1349  * @ghs: an array of struct gfs2_holder structures
1350  *
1351  * Returns: 0 on success (all glocks acquired),
1352  *          errno on failure (no glocks acquired)
1353  */
1354
1355 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1356                      struct gfs2_holder **p)
1357 {
1358         unsigned int x;
1359         int error = 0;
1360
1361         for (x = 0; x < num_gh; x++)
1362                 p[x] = &ghs[x];
1363
1364         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1365
1366         for (x = 0; x < num_gh; x++) {
1367                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1368
1369                 error = gfs2_glock_nq(p[x]);
1370                 if (error) {
1371                         while (x--)
1372                                 gfs2_glock_dq(p[x]);
1373                         break;
1374                 }
1375         }
1376
1377         return error;
1378 }
1379
1380 /**
1381  * gfs2_glock_nq_m - acquire multiple glocks
1382  * @num_gh: the number of structures
1383  * @ghs: an array of struct gfs2_holder structures
1384  *
1385  *
1386  * Returns: 0 on success (all glocks acquired),
1387  *          errno on failure (no glocks acquired)
1388  */
1389
1390 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1391 {
1392         struct gfs2_holder *tmp[4];
1393         struct gfs2_holder **pph = tmp;
1394         int error = 0;
1395
1396         switch(num_gh) {
1397         case 0:
1398                 return 0;
1399         case 1:
1400                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1401                 return gfs2_glock_nq(ghs);
1402         default:
1403                 if (num_gh <= 4)
1404                         break;
1405                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1406                                     GFP_NOFS);
1407                 if (!pph)
1408                         return -ENOMEM;
1409         }
1410
1411         error = nq_m_sync(num_gh, ghs, pph);
1412
1413         if (pph != tmp)
1414                 kfree(pph);
1415
1416         return error;
1417 }
1418
1419 /**
1420  * gfs2_glock_dq_m - release multiple glocks
1421  * @num_gh: the number of structures
1422  * @ghs: an array of struct gfs2_holder structures
1423  *
1424  */
1425
1426 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1427 {
1428         while (num_gh--)
1429                 gfs2_glock_dq(&ghs[num_gh]);
1430 }
1431
1432 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1433 {
1434         unsigned long delay = 0;
1435         unsigned long holdtime;
1436         unsigned long now = jiffies;
1437
1438         gfs2_glock_hold(gl);
1439         holdtime = gl->gl_tchange + gl->gl_hold_time;
1440         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1441             gl->gl_name.ln_type == LM_TYPE_INODE) {
1442                 if (time_before(now, holdtime))
1443                         delay = holdtime - now;
1444                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1445                         delay = gl->gl_hold_time;
1446         }
1447
1448         spin_lock(&gl->gl_lockref.lock);
1449         handle_callback(gl, state, delay, true);
1450         __gfs2_glock_queue_work(gl, delay);
1451         spin_unlock(&gl->gl_lockref.lock);
1452 }
1453
1454 /**
1455  * gfs2_should_freeze - Figure out if glock should be frozen
1456  * @gl: The glock in question
1457  *
1458  * Glocks are not frozen if (a) the result of the dlm operation is
1459  * an error, (b) the locking operation was an unlock operation or
1460  * (c) if there is a "noexp" flagged request anywhere in the queue
1461  *
1462  * Returns: 1 if freezing should occur, 0 otherwise
1463  */
1464
1465 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1466 {
1467         const struct gfs2_holder *gh;
1468
1469         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1470                 return 0;
1471         if (gl->gl_target == LM_ST_UNLOCKED)
1472                 return 0;
1473
1474         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1475                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1476                         continue;
1477                 if (LM_FLAG_NOEXP & gh->gh_flags)
1478                         return 0;
1479         }
1480
1481         return 1;
1482 }
1483
1484 /**
1485  * gfs2_glock_complete - Callback used by locking
1486  * @gl: Pointer to the glock
1487  * @ret: The return value from the dlm
1488  *
1489  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1490  * to use a bitfield shared with other glock state fields.
1491  */
1492
1493 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1494 {
1495         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1496
1497         spin_lock(&gl->gl_lockref.lock);
1498         gl->gl_reply = ret;
1499
1500         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1501                 if (gfs2_should_freeze(gl)) {
1502                         set_bit(GLF_FROZEN, &gl->gl_flags);
1503                         spin_unlock(&gl->gl_lockref.lock);
1504                         return;
1505                 }
1506         }
1507
1508         gl->gl_lockref.count++;
1509         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1510         __gfs2_glock_queue_work(gl, 0);
1511         spin_unlock(&gl->gl_lockref.lock);
1512 }
1513
1514 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1515 {
1516         struct gfs2_glock *gla, *glb;
1517
1518         gla = list_entry(a, struct gfs2_glock, gl_lru);
1519         glb = list_entry(b, struct gfs2_glock, gl_lru);
1520
1521         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1522                 return 1;
1523         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1524                 return -1;
1525
1526         return 0;
1527 }
1528
1529 /**
1530  * gfs2_dispose_glock_lru - Demote a list of glocks
1531  * @list: The list to dispose of
1532  *
1533  * Disposing of glocks may involve disk accesses, so that here we sort
1534  * the glocks by number (i.e. disk location of the inodes) so that if
1535  * there are any such accesses, they'll be sent in order (mostly).
1536  *
1537  * Must be called under the lru_lock, but may drop and retake this
1538  * lock. While the lru_lock is dropped, entries may vanish from the
1539  * list, but no new entries will appear on the list (since it is
1540  * private)
1541  */
1542
1543 static void gfs2_dispose_glock_lru(struct list_head *list)
1544 __releases(&lru_lock)
1545 __acquires(&lru_lock)
1546 {
1547         struct gfs2_glock *gl;
1548
1549         list_sort(NULL, list, glock_cmp);
1550
1551         while(!list_empty(list)) {
1552                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1553                 list_del_init(&gl->gl_lru);
1554                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1555 add_back_to_lru:
1556                         list_add(&gl->gl_lru, &lru_list);
1557                         set_bit(GLF_LRU, &gl->gl_flags);
1558                         atomic_inc(&lru_count);
1559                         continue;
1560                 }
1561                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1562                         spin_unlock(&gl->gl_lockref.lock);
1563                         goto add_back_to_lru;
1564                 }
1565                 gl->gl_lockref.count++;
1566                 if (demote_ok(gl))
1567                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1568                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1569                 __gfs2_glock_queue_work(gl, 0);
1570                 spin_unlock(&gl->gl_lockref.lock);
1571                 cond_resched_lock(&lru_lock);
1572         }
1573 }
1574
1575 /**
1576  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1577  * @nr: The number of entries to scan
1578  *
1579  * This function selects the entries on the LRU which are able to
1580  * be demoted, and then kicks off the process by calling
1581  * gfs2_dispose_glock_lru() above.
1582  */
1583
1584 static long gfs2_scan_glock_lru(int nr)
1585 {
1586         struct gfs2_glock *gl;
1587         LIST_HEAD(skipped);
1588         LIST_HEAD(dispose);
1589         long freed = 0;
1590
1591         spin_lock(&lru_lock);
1592         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1593                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1594
1595                 /* Test for being demotable */
1596                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1597                         list_move(&gl->gl_lru, &dispose);
1598                         atomic_dec(&lru_count);
1599                         clear_bit(GLF_LRU, &gl->gl_flags);
1600                         freed++;
1601                         continue;
1602                 }
1603
1604                 list_move(&gl->gl_lru, &skipped);
1605         }
1606         list_splice(&skipped, &lru_list);
1607         if (!list_empty(&dispose))
1608                 gfs2_dispose_glock_lru(&dispose);
1609         spin_unlock(&lru_lock);
1610
1611         return freed;
1612 }
1613
1614 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1615                                             struct shrink_control *sc)
1616 {
1617         if (!(sc->gfp_mask & __GFP_FS))
1618                 return SHRINK_STOP;
1619         return gfs2_scan_glock_lru(sc->nr_to_scan);
1620 }
1621
1622 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1623                                              struct shrink_control *sc)
1624 {
1625         return vfs_pressure_ratio(atomic_read(&lru_count));
1626 }
1627
1628 static struct shrinker glock_shrinker = {
1629         .seeks = DEFAULT_SEEKS,
1630         .count_objects = gfs2_glock_shrink_count,
1631         .scan_objects = gfs2_glock_shrink_scan,
1632 };
1633
1634 /**
1635  * examine_bucket - Call a function for glock in a hash bucket
1636  * @examiner: the function
1637  * @sdp: the filesystem
1638  * @bucket: the bucket
1639  *
1640  * Note that the function can be called multiple times on the same
1641  * object.  So the user must ensure that the function can cope with
1642  * that.
1643  */
1644
1645 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1646 {
1647         struct gfs2_glock *gl;
1648         struct rhashtable_iter iter;
1649
1650         rhashtable_walk_enter(&gl_hash_table, &iter);
1651
1652         do {
1653                 rhashtable_walk_start(&iter);
1654
1655                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1656                         if (gl->gl_name.ln_sbd == sdp &&
1657                             lockref_get_not_dead(&gl->gl_lockref))
1658                                 examiner(gl);
1659
1660                 rhashtable_walk_stop(&iter);
1661         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1662
1663         rhashtable_walk_exit(&iter);
1664 }
1665
1666 /**
1667  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1668  * @gl: The glock to thaw
1669  *
1670  */
1671
1672 static void thaw_glock(struct gfs2_glock *gl)
1673 {
1674         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1675                 gfs2_glock_put(gl);
1676                 return;
1677         }
1678         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1679         gfs2_glock_queue_work(gl, 0);
1680 }
1681
1682 /**
1683  * clear_glock - look at a glock and see if we can free it from glock cache
1684  * @gl: the glock to look at
1685  *
1686  */
1687
1688 static void clear_glock(struct gfs2_glock *gl)
1689 {
1690         gfs2_glock_remove_from_lru(gl);
1691
1692         spin_lock(&gl->gl_lockref.lock);
1693         if (gl->gl_state != LM_ST_UNLOCKED)
1694                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1695         __gfs2_glock_queue_work(gl, 0);
1696         spin_unlock(&gl->gl_lockref.lock);
1697 }
1698
1699 /**
1700  * gfs2_glock_thaw - Thaw any frozen glocks
1701  * @sdp: The super block
1702  *
1703  */
1704
1705 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1706 {
1707         glock_hash_walk(thaw_glock, sdp);
1708 }
1709
1710 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1711 {
1712         spin_lock(&gl->gl_lockref.lock);
1713         gfs2_dump_glock(seq, gl, fsid);
1714         spin_unlock(&gl->gl_lockref.lock);
1715 }
1716
1717 static void dump_glock_func(struct gfs2_glock *gl)
1718 {
1719         dump_glock(NULL, gl, true);
1720 }
1721
1722 /**
1723  * gfs2_gl_hash_clear - Empty out the glock hash table
1724  * @sdp: the filesystem
1725  * @wait: wait until it's all gone
1726  *
1727  * Called when unmounting the filesystem.
1728  */
1729
1730 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1731 {
1732         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1733         flush_workqueue(glock_workqueue);
1734         glock_hash_walk(clear_glock, sdp);
1735         flush_workqueue(glock_workqueue);
1736         wait_event_timeout(sdp->sd_glock_wait,
1737                            atomic_read(&sdp->sd_glock_disposal) == 0,
1738                            HZ * 600);
1739         glock_hash_walk(dump_glock_func, sdp);
1740 }
1741
1742 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1743 {
1744         struct gfs2_glock *gl = ip->i_gl;
1745         int ret;
1746
1747         ret = gfs2_truncatei_resume(ip);
1748         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1749
1750         spin_lock(&gl->gl_lockref.lock);
1751         clear_bit(GLF_LOCK, &gl->gl_flags);
1752         run_queue(gl, 1);
1753         spin_unlock(&gl->gl_lockref.lock);
1754 }
1755
1756 static const char *state2str(unsigned state)
1757 {
1758         switch(state) {
1759         case LM_ST_UNLOCKED:
1760                 return "UN";
1761         case LM_ST_SHARED:
1762                 return "SH";
1763         case LM_ST_DEFERRED:
1764                 return "DF";
1765         case LM_ST_EXCLUSIVE:
1766                 return "EX";
1767         }
1768         return "??";
1769 }
1770
1771 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1772 {
1773         char *p = buf;
1774         if (flags & LM_FLAG_TRY)
1775                 *p++ = 't';
1776         if (flags & LM_FLAG_TRY_1CB)
1777                 *p++ = 'T';
1778         if (flags & LM_FLAG_NOEXP)
1779                 *p++ = 'e';
1780         if (flags & LM_FLAG_ANY)
1781                 *p++ = 'A';
1782         if (flags & LM_FLAG_PRIORITY)
1783                 *p++ = 'p';
1784         if (flags & GL_ASYNC)
1785                 *p++ = 'a';
1786         if (flags & GL_EXACT)
1787                 *p++ = 'E';
1788         if (flags & GL_NOCACHE)
1789                 *p++ = 'c';
1790         if (test_bit(HIF_HOLDER, &iflags))
1791                 *p++ = 'H';
1792         if (test_bit(HIF_WAIT, &iflags))
1793                 *p++ = 'W';
1794         if (test_bit(HIF_FIRST, &iflags))
1795                 *p++ = 'F';
1796         *p = 0;
1797         return buf;
1798 }
1799
1800 /**
1801  * dump_holder - print information about a glock holder
1802  * @seq: the seq_file struct
1803  * @gh: the glock holder
1804  * @fs_id_buf: pointer to file system id (if requested)
1805  *
1806  */
1807
1808 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh,
1809                         const char *fs_id_buf)
1810 {
1811         struct task_struct *gh_owner = NULL;
1812         char flags_buf[32];
1813
1814         rcu_read_lock();
1815         if (gh->gh_owner_pid)
1816                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1817         gfs2_print_dbg(seq, "%s H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1818                        fs_id_buf, state2str(gh->gh_state),
1819                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1820                        gh->gh_error,
1821                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1822                        gh_owner ? gh_owner->comm : "(ended)",
1823                        (void *)gh->gh_ip);
1824         rcu_read_unlock();
1825 }
1826
1827 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1828 {
1829         const unsigned long *gflags = &gl->gl_flags;
1830         char *p = buf;
1831
1832         if (test_bit(GLF_LOCK, gflags))
1833                 *p++ = 'l';
1834         if (test_bit(GLF_DEMOTE, gflags))
1835                 *p++ = 'D';
1836         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1837                 *p++ = 'd';
1838         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1839                 *p++ = 'p';
1840         if (test_bit(GLF_DIRTY, gflags))
1841                 *p++ = 'y';
1842         if (test_bit(GLF_LFLUSH, gflags))
1843                 *p++ = 'f';
1844         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1845                 *p++ = 'i';
1846         if (test_bit(GLF_REPLY_PENDING, gflags))
1847                 *p++ = 'r';
1848         if (test_bit(GLF_INITIAL, gflags))
1849                 *p++ = 'I';
1850         if (test_bit(GLF_FROZEN, gflags))
1851                 *p++ = 'F';
1852         if (test_bit(GLF_QUEUED, gflags))
1853                 *p++ = 'q';
1854         if (test_bit(GLF_LRU, gflags))
1855                 *p++ = 'L';
1856         if (gl->gl_object)
1857                 *p++ = 'o';
1858         if (test_bit(GLF_BLOCKING, gflags))
1859                 *p++ = 'b';
1860         *p = 0;
1861         return buf;
1862 }
1863
1864 /**
1865  * gfs2_dump_glock - print information about a glock
1866  * @seq: The seq_file struct
1867  * @gl: the glock
1868  * @fsid: If true, also dump the file system id
1869  *
1870  * The file format is as follows:
1871  * One line per object, capital letters are used to indicate objects
1872  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1873  * other objects are indented by a single space and follow the glock to
1874  * which they are related. Fields are indicated by lower case letters
1875  * followed by a colon and the field value, except for strings which are in
1876  * [] so that its possible to see if they are composed of spaces for
1877  * example. The field's are n = number (id of the object), f = flags,
1878  * t = type, s = state, r = refcount, e = error, p = pid.
1879  *
1880  */
1881
1882 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1883 {
1884         const struct gfs2_glock_operations *glops = gl->gl_ops;
1885         unsigned long long dtime;
1886         const struct gfs2_holder *gh;
1887         char gflags_buf[32];
1888         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1889         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
1890
1891         memset(fs_id_buf, 0, sizeof(fs_id_buf));
1892         if (fsid && sdp) /* safety precaution */
1893                 sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
1894         dtime = jiffies - gl->gl_demote_time;
1895         dtime *= 1000000/HZ; /* demote time in uSec */
1896         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1897                 dtime = 0;
1898         gfs2_print_dbg(seq, "%sG:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d "
1899                        "v:%d r:%d m:%ld\n", fs_id_buf, state2str(gl->gl_state),
1900                   gl->gl_name.ln_type,
1901                   (unsigned long long)gl->gl_name.ln_number,
1902                   gflags2str(gflags_buf, gl),
1903                   state2str(gl->gl_target),
1904                   state2str(gl->gl_demote_state), dtime,
1905                   atomic_read(&gl->gl_ail_count),
1906                   atomic_read(&gl->gl_revokes),
1907                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1908
1909         list_for_each_entry(gh, &gl->gl_holders, gh_list)
1910                 dump_holder(seq, gh, fs_id_buf);
1911
1912         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1913                 glops->go_dump(seq, gl, fs_id_buf);
1914 }
1915
1916 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
1917 {
1918         struct gfs2_glock *gl = iter_ptr;
1919
1920         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
1921                    gl->gl_name.ln_type,
1922                    (unsigned long long)gl->gl_name.ln_number,
1923                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
1924                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
1925                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
1926                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
1927                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
1928                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
1929                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
1930                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
1931         return 0;
1932 }
1933
1934 static const char *gfs2_gltype[] = {
1935         "type",
1936         "reserved",
1937         "nondisk",
1938         "inode",
1939         "rgrp",
1940         "meta",
1941         "iopen",
1942         "flock",
1943         "plock",
1944         "quota",
1945         "journal",
1946 };
1947
1948 static const char *gfs2_stype[] = {
1949         [GFS2_LKS_SRTT]         = "srtt",
1950         [GFS2_LKS_SRTTVAR]      = "srttvar",
1951         [GFS2_LKS_SRTTB]        = "srttb",
1952         [GFS2_LKS_SRTTVARB]     = "srttvarb",
1953         [GFS2_LKS_SIRT]         = "sirt",
1954         [GFS2_LKS_SIRTVAR]      = "sirtvar",
1955         [GFS2_LKS_DCOUNT]       = "dlm",
1956         [GFS2_LKS_QCOUNT]       = "queue",
1957 };
1958
1959 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
1960
1961 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
1962 {
1963         struct gfs2_sbd *sdp = seq->private;
1964         loff_t pos = *(loff_t *)iter_ptr;
1965         unsigned index = pos >> 3;
1966         unsigned subindex = pos & 0x07;
1967         int i;
1968
1969         if (index == 0 && subindex != 0)
1970                 return 0;
1971
1972         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
1973                    (index == 0) ? "cpu": gfs2_stype[subindex]);
1974
1975         for_each_possible_cpu(i) {
1976                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
1977
1978                 if (index == 0)
1979                         seq_printf(seq, " %15u", i);
1980                 else
1981                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
1982                                    lkstats[index - 1].stats[subindex]);
1983         }
1984         seq_putc(seq, '\n');
1985         return 0;
1986 }
1987
1988 int __init gfs2_glock_init(void)
1989 {
1990         int i, ret;
1991
1992         ret = rhashtable_init(&gl_hash_table, &ht_parms);
1993         if (ret < 0)
1994                 return ret;
1995
1996         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1997                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1998         if (!glock_workqueue) {
1999                 rhashtable_destroy(&gl_hash_table);
2000                 return -ENOMEM;
2001         }
2002         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
2003                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
2004                                                 0);
2005         if (!gfs2_delete_workqueue) {
2006                 destroy_workqueue(glock_workqueue);
2007                 rhashtable_destroy(&gl_hash_table);
2008                 return -ENOMEM;
2009         }
2010
2011         ret = register_shrinker(&glock_shrinker);
2012         if (ret) {
2013                 destroy_workqueue(gfs2_delete_workqueue);
2014                 destroy_workqueue(glock_workqueue);
2015                 rhashtable_destroy(&gl_hash_table);
2016                 return ret;
2017         }
2018
2019         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
2020                 init_waitqueue_head(glock_wait_table + i);
2021
2022         return 0;
2023 }
2024
2025 void gfs2_glock_exit(void)
2026 {
2027         unregister_shrinker(&glock_shrinker);
2028         rhashtable_destroy(&gl_hash_table);
2029         destroy_workqueue(glock_workqueue);
2030         destroy_workqueue(gfs2_delete_workqueue);
2031 }
2032
2033 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
2034 {
2035         struct gfs2_glock *gl = gi->gl;
2036
2037         if (gl) {
2038                 if (n == 0)
2039                         return;
2040                 if (!lockref_put_not_zero(&gl->gl_lockref))
2041                         gfs2_glock_queue_put(gl);
2042         }
2043         for (;;) {
2044                 gl = rhashtable_walk_next(&gi->hti);
2045                 if (IS_ERR_OR_NULL(gl)) {
2046                         if (gl == ERR_PTR(-EAGAIN)) {
2047                                 n = 1;
2048                                 continue;
2049                         }
2050                         gl = NULL;
2051                         break;
2052                 }
2053                 if (gl->gl_name.ln_sbd != gi->sdp)
2054                         continue;
2055                 if (n <= 1) {
2056                         if (!lockref_get_not_dead(&gl->gl_lockref))
2057                                 continue;
2058                         break;
2059                 } else {
2060                         if (__lockref_is_dead(&gl->gl_lockref))
2061                                 continue;
2062                         n--;
2063                 }
2064         }
2065         gi->gl = gl;
2066 }
2067
2068 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
2069         __acquires(RCU)
2070 {
2071         struct gfs2_glock_iter *gi = seq->private;
2072         loff_t n;
2073
2074         /*
2075          * We can either stay where we are, skip to the next hash table
2076          * entry, or start from the beginning.
2077          */
2078         if (*pos < gi->last_pos) {
2079                 rhashtable_walk_exit(&gi->hti);
2080                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2081                 n = *pos + 1;
2082         } else {
2083                 n = *pos - gi->last_pos;
2084         }
2085
2086         rhashtable_walk_start(&gi->hti);
2087
2088         gfs2_glock_iter_next(gi, n);
2089         gi->last_pos = *pos;
2090         return gi->gl;
2091 }
2092
2093 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
2094                                  loff_t *pos)
2095 {
2096         struct gfs2_glock_iter *gi = seq->private;
2097
2098         (*pos)++;
2099         gi->last_pos = *pos;
2100         gfs2_glock_iter_next(gi, 1);
2101         return gi->gl;
2102 }
2103
2104 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2105         __releases(RCU)
2106 {
2107         struct gfs2_glock_iter *gi = seq->private;
2108
2109         rhashtable_walk_stop(&gi->hti);
2110 }
2111
2112 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2113 {
2114         dump_glock(seq, iter_ptr, false);
2115         return 0;
2116 }
2117
2118 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2119 {
2120         preempt_disable();
2121         if (*pos >= GFS2_NR_SBSTATS)
2122                 return NULL;
2123         return pos;
2124 }
2125
2126 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2127                                    loff_t *pos)
2128 {
2129         (*pos)++;
2130         if (*pos >= GFS2_NR_SBSTATS)
2131                 return NULL;
2132         return pos;
2133 }
2134
2135 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2136 {
2137         preempt_enable();
2138 }
2139
2140 static const struct seq_operations gfs2_glock_seq_ops = {
2141         .start = gfs2_glock_seq_start,
2142         .next  = gfs2_glock_seq_next,
2143         .stop  = gfs2_glock_seq_stop,
2144         .show  = gfs2_glock_seq_show,
2145 };
2146
2147 static const struct seq_operations gfs2_glstats_seq_ops = {
2148         .start = gfs2_glock_seq_start,
2149         .next  = gfs2_glock_seq_next,
2150         .stop  = gfs2_glock_seq_stop,
2151         .show  = gfs2_glstats_seq_show,
2152 };
2153
2154 static const struct seq_operations gfs2_sbstats_seq_ops = {
2155         .start = gfs2_sbstats_seq_start,
2156         .next  = gfs2_sbstats_seq_next,
2157         .stop  = gfs2_sbstats_seq_stop,
2158         .show  = gfs2_sbstats_seq_show,
2159 };
2160
2161 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2162
2163 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2164                               const struct seq_operations *ops)
2165 {
2166         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2167         if (ret == 0) {
2168                 struct seq_file *seq = file->private_data;
2169                 struct gfs2_glock_iter *gi = seq->private;
2170
2171                 gi->sdp = inode->i_private;
2172                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2173                 if (seq->buf)
2174                         seq->size = GFS2_SEQ_GOODSIZE;
2175                 /*
2176                  * Initially, we are "before" the first hash table entry; the
2177                  * first call to rhashtable_walk_next gets us the first entry.
2178                  */
2179                 gi->last_pos = -1;
2180                 gi->gl = NULL;
2181                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2182         }
2183         return ret;
2184 }
2185
2186 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2187 {
2188         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2189 }
2190
2191 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2192 {
2193         struct seq_file *seq = file->private_data;
2194         struct gfs2_glock_iter *gi = seq->private;
2195
2196         if (gi->gl)
2197                 gfs2_glock_put(gi->gl);
2198         rhashtable_walk_exit(&gi->hti);
2199         return seq_release_private(inode, file);
2200 }
2201
2202 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2203 {
2204         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2205 }
2206
2207 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2208 {
2209         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2210         if (ret == 0) {
2211                 struct seq_file *seq = file->private_data;
2212                 seq->private = inode->i_private;  /* sdp */
2213         }
2214         return ret;
2215 }
2216
2217 static const struct file_operations gfs2_glocks_fops = {
2218         .owner   = THIS_MODULE,
2219         .open    = gfs2_glocks_open,
2220         .read    = seq_read,
2221         .llseek  = seq_lseek,
2222         .release = gfs2_glocks_release,
2223 };
2224
2225 static const struct file_operations gfs2_glstats_fops = {
2226         .owner   = THIS_MODULE,
2227         .open    = gfs2_glstats_open,
2228         .read    = seq_read,
2229         .llseek  = seq_lseek,
2230         .release = gfs2_glocks_release,
2231 };
2232
2233 static const struct file_operations gfs2_sbstats_fops = {
2234         .owner   = THIS_MODULE,
2235         .open    = gfs2_sbstats_open,
2236         .read    = seq_read,
2237         .llseek  = seq_lseek,
2238         .release = seq_release,
2239 };
2240
2241 void gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2242 {
2243         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2244
2245         debugfs_create_file("glocks", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2246                             &gfs2_glocks_fops);
2247
2248         debugfs_create_file("glstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2249                             &gfs2_glstats_fops);
2250
2251         debugfs_create_file("sbstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2252                             &gfs2_sbstats_fops);
2253 }
2254
2255 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2256 {
2257         debugfs_remove_recursive(sdp->debugfs_dir);
2258         sdp->debugfs_dir = NULL;
2259 }
2260
2261 void gfs2_register_debugfs(void)
2262 {
2263         gfs2_root = debugfs_create_dir("gfs2", NULL);
2264 }
2265
2266 void gfs2_unregister_debugfs(void)
2267 {
2268         debugfs_remove(gfs2_root);
2269         gfs2_root = NULL;
2270 }