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 /**
 137  * glock_blocked_by_withdraw - determine if we can still use a glock
 138  * @gl: the glock
 139  *
 140  * We need to allow some glocks to be enqueued, dequeued, promoted, and demoted
 141  * when we're withdrawn. For example, to maintain metadata integrity, we should
 142  * disallow the use of inode and rgrp glocks when withdrawn. Other glocks, like
 143  * iopen or the transaction glocks may be safely used because none of their
 144  * metadata goes through the journal. So in general, we should disallow all
 145  * glocks that are journaled, and allow all the others. One exception is:
 146  * we need to allow our active journal to be promoted and demoted so others
 147  * may recover it and we can reacquire it when they're done.
 148  */
 149 static bool glock_blocked_by_withdraw(struct gfs2_glock *gl)
 150 {
 151         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 152
 153         if (likely(!gfs2_withdrawn(sdp)))
 154                 return false;
 155         if (gl->gl_ops->go_flags & GLOF_NONDISK)
 156                 return false;
 157         if (!sdp->sd_jdesc ||
 158             gl->gl_name.ln_number == sdp->sd_jdesc->jd_no_addr)
 159                 return false;
 160         return true;
 161 }
 162
 163 void gfs2_glock_free(struct gfs2_glock *gl)
 164 {
 165         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 166
 167         BUG_ON(atomic_read(&gl->gl_revokes));
 168         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 169         smp_mb();
 170         wake_up_glock(gl);
 171         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 172         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 173                 wake_up(&sdp->sd_glock_wait);
 174 }
 175
 176 /**
 177  * gfs2_glock_hold() - increment reference count on glock
 178  * @gl: The glock to hold
 179  *
 180  */
 181
 182 void gfs2_glock_hold(struct gfs2_glock *gl)
 183 {
 184         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 185         lockref_get(&gl->gl_lockref);
 186 }
 187
 188 /**
 189  * demote_ok - Check to see if it's ok to unlock a glock
 190  * @gl: the glock
 191  *
 192  * Returns: 1 if it's ok
 193  */
 194
 195 static int demote_ok(const struct gfs2_glock *gl)
 196 {
 197         const struct gfs2_glock_operations *glops = gl->gl_ops;
 198
 199         if (gl->gl_state == LM_ST_UNLOCKED)
 200                 return 0;
 201         if (!list_empty(&gl->gl_holders))
 202                 return 0;
 203         if (glops->go_demote_ok)
 204                 return glops->go_demote_ok(gl);
 205         return 1;
 206 }
 207
 208
 209 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 210 {
 211         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 212                 return;
 213
 214         spin_lock(&lru_lock);
 215
 216         list_del(&gl->gl_lru);
 217         list_add_tail(&gl->gl_lru, &lru_list);
 218
 219         if (!test_bit(GLF_LRU, &gl->gl_flags)) {
 220                 set_bit(GLF_LRU, &gl->gl_flags);
 221                 atomic_inc(&lru_count);
 222         }
 223
 224         spin_unlock(&lru_lock);
 225 }
 226
 227 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 228 {
 229         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 230                 return;
 231
 232         spin_lock(&lru_lock);
 233         if (test_bit(GLF_LRU, &gl->gl_flags)) {
 234                 list_del_init(&gl->gl_lru);
 235                 atomic_dec(&lru_count);
 236                 clear_bit(GLF_LRU, &gl->gl_flags);
 237         }
 238         spin_unlock(&lru_lock);
 239 }
 240
 241 /*
 242  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 243  * work queue.
 244  */
 245 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 246         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 247                 /*
 248                  * We are holding the lockref spinlock, and the work was still
 249                  * queued above.  The queued work (glock_work_func) takes that
 250                  * spinlock before dropping its glock reference(s), so it
 251                  * cannot have dropped them in the meantime.
 252                  */
 253                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 254                 gl->gl_lockref.count--;
 255         }
 256 }
 257
 258 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 259         spin_lock(&gl->gl_lockref.lock);
 260         __gfs2_glock_queue_work(gl, delay);
 261         spin_unlock(&gl->gl_lockref.lock);
 262 }
 263
 264 static void __gfs2_glock_put(struct gfs2_glock *gl)
 265 {
 266         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 267         struct address_space *mapping = gfs2_glock2aspace(gl);
 268
 269         lockref_mark_dead(&gl->gl_lockref);
 270
 271         gfs2_glock_remove_from_lru(gl);
 272         spin_unlock(&gl->gl_lockref.lock);
 273         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 274         GLOCK_BUG_ON(gl, mapping && mapping->nrpages && !gfs2_withdrawn(sdp));
 275         trace_gfs2_glock_put(gl);
 276         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 277 }
 278
 279 /*
 280  * Cause the glock to be put in work queue context.
 281  */
 282 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 283 {
 284         gfs2_glock_queue_work(gl, 0);
 285 }
 286
 287 /**
 288  * gfs2_glock_put() - Decrement reference count on glock
 289  * @gl: The glock to put
 290  *
 291  */
 292
 293 void gfs2_glock_put(struct gfs2_glock *gl)
 294 {
 295         if (lockref_put_or_lock(&gl->gl_lockref))
 296                 return;
 297
 298         __gfs2_glock_put(gl);
 299 }
 300
 301 /**
 302  * may_grant - check if its ok to grant a new lock
 303  * @gl: The glock
 304  * @gh: The lock request which we wish to grant
 305  *
 306  * Returns: true if its ok to grant the lock
 307  */
 308
 309 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 310 {
 311         const struct gfs2_holder *gh_head = list_first_entry(&gl->gl_holders, const struct gfs2_holder, gh_list);
 312         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 313              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 314                 return 0;
 315         if (gl->gl_state == gh->gh_state)
 316                 return 1;
 317         if (gh->gh_flags & GL_EXACT)
 318                 return 0;
 319         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 320                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 321                         return 1;
 322                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 323                         return 1;
 324         }
 325         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 326                 return 1;
 327         return 0;
 328 }
 329
 330 static void gfs2_holder_wake(struct gfs2_holder *gh)
 331 {
 332         clear_bit(HIF_WAIT, &gh->gh_iflags);
 333         smp_mb__after_atomic();
 334         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 335         if (gh->gh_flags & GL_ASYNC) {
 336                 struct gfs2_sbd *sdp = gh->gh_gl->gl_name.ln_sbd;
 337
 338                 wake_up(&sdp->sd_async_glock_wait);
 339         }
 340 }
 341
 342 /**
 343  * do_error - Something unexpected has happened during a lock request
 344  *
 345  */
 346
 347 static void do_error(struct gfs2_glock *gl, const int ret)
 348 {
 349         struct gfs2_holder *gh, *tmp;
 350
 351         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 352                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 353                         continue;
 354                 if (ret & LM_OUT_ERROR)
 355                         gh->gh_error = -EIO;
 356                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 357                         gh->gh_error = GLR_TRYFAILED;
 358                 else
 359                         continue;
 360                 list_del_init(&gh->gh_list);
 361                 trace_gfs2_glock_queue(gh, 0);
 362                 gfs2_holder_wake(gh);
 363         }
 364 }
 365
 366 /**
 367  * do_promote - promote as many requests as possible on the current queue
 368  * @gl: The glock
 369  *
 370  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 371  *          if a type specific operation is underway.
 372  */
 373
 374 static int do_promote(struct gfs2_glock *gl)
 375 __releases(&gl->gl_lockref.lock)
 376 __acquires(&gl->gl_lockref.lock)
 377 {
 378         const struct gfs2_glock_operations *glops = gl->gl_ops;
 379         struct gfs2_holder *gh, *tmp;
 380         int ret;
 381
 382 restart:
 383         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 384                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 385                         continue;
 386                 if (may_grant(gl, gh)) {
 387                         if (gh->gh_list.prev == &gl->gl_holders &&
 388                             glops->go_lock) {
 389                                 spin_unlock(&gl->gl_lockref.lock);
 390                                 /* FIXME: eliminate this eventually */
 391                                 ret = glops->go_lock(gh);
 392                                 spin_lock(&gl->gl_lockref.lock);
 393                                 if (ret) {
 394                                         if (ret == 1)
 395                                                 return 2;
 396                                         gh->gh_error = ret;
 397                                         list_del_init(&gh->gh_list);
 398                                         trace_gfs2_glock_queue(gh, 0);
 399                                         gfs2_holder_wake(gh);
 400                                         goto restart;
 401                                 }
 402                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 403                                 trace_gfs2_promote(gh, 1);
 404                                 gfs2_holder_wake(gh);
 405                                 goto restart;
 406                         }
 407                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 408                         trace_gfs2_promote(gh, 0);
 409                         gfs2_holder_wake(gh);
 410                         continue;
 411                 }
 412                 if (gh->gh_list.prev == &gl->gl_holders)
 413                         return 1;
 414                 do_error(gl, 0);
 415                 break;
 416         }
 417         return 0;
 418 }
 419
 420 /**
 421  * find_first_waiter - find the first gh that's waiting for the glock
 422  * @gl: the glock
 423  */
 424
 425 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 426 {
 427         struct gfs2_holder *gh;
 428
 429         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 430                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 431                         return gh;
 432         }
 433         return NULL;
 434 }
 435
 436 /**
 437  * state_change - record that the glock is now in a different state
 438  * @gl: the glock
 439  * @new_state the new state
 440  *
 441  */
 442
 443 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 444 {
 445         int held1, held2;
 446
 447         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 448         held2 = (new_state != LM_ST_UNLOCKED);
 449
 450         if (held1 != held2) {
 451                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 452                 if (held2)
 453                         gl->gl_lockref.count++;
 454                 else
 455                         gl->gl_lockref.count--;
 456         }
 457         if (held1 && held2 && list_empty(&gl->gl_holders))
 458                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 459
 460         if (new_state != gl->gl_target)
 461                 /* shorten our minimum hold time */
 462                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 463                                        GL_GLOCK_MIN_HOLD);
 464         gl->gl_state = new_state;
 465         gl->gl_tchange = jiffies;
 466 }
 467
 468 static void gfs2_demote_wake(struct gfs2_glock *gl)
 469 {
 470         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 471         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 472         smp_mb__after_atomic();
 473         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 474 }
 475
 476 /**
 477  * finish_xmote - The DLM has replied to one of our lock requests
 478  * @gl: The glock
 479  * @ret: The status from the DLM
 480  *
 481  */
 482
 483 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 484 {
 485         const struct gfs2_glock_operations *glops = gl->gl_ops;
 486         struct gfs2_holder *gh;
 487         unsigned state = ret & LM_OUT_ST_MASK;
 488         int rv;
 489
 490         spin_lock(&gl->gl_lockref.lock);
 491         trace_gfs2_glock_state_change(gl, state);
 492         state_change(gl, state);
 493         gh = find_first_waiter(gl);
 494
 495         /* Demote to UN request arrived during demote to SH or DF */
 496         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 497             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 498                 gl->gl_target = LM_ST_UNLOCKED;
 499
 500         /* Check for state != intended state */
 501         if (unlikely(state != gl->gl_target)) {
 502                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 503                         /* move to back of queue and try next entry */
 504                         if (ret & LM_OUT_CANCELED) {
 505                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 506                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 507                                 gh = find_first_waiter(gl);
 508                                 gl->gl_target = gh->gh_state;
 509                                 goto retry;
 510                         }
 511                         /* Some error or failed "try lock" - report it */
 512                         if ((ret & LM_OUT_ERROR) ||
 513                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 514                                 gl->gl_target = gl->gl_state;
 515                                 do_error(gl, ret);
 516                                 goto out;
 517                         }
 518                 }
 519                 switch(state) {
 520                 /* Unlocked due to conversion deadlock, try again */
 521                 case LM_ST_UNLOCKED:
 522 retry:
 523                         do_xmote(gl, gh, gl->gl_target);
 524                         break;
 525                 /* Conversion fails, unlock and try again */
 526                 case LM_ST_SHARED:
 527                 case LM_ST_DEFERRED:
 528                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 529                         break;
 530                 default: /* Everything else */
 531                         fs_err(gl->gl_name.ln_sbd, "wanted %u got %u\n",
 532                                gl->gl_target, state);
 533                         GLOCK_BUG_ON(gl, 1);
 534                 }
 535                 spin_unlock(&gl->gl_lockref.lock);
 536                 return;
 537         }
 538
 539         /* Fast path - we got what we asked for */
 540         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 541                 gfs2_demote_wake(gl);
 542         if (state != LM_ST_UNLOCKED) {
 543                 if (glops->go_xmote_bh) {
 544                         spin_unlock(&gl->gl_lockref.lock);
 545                         rv = glops->go_xmote_bh(gl, gh);
 546                         spin_lock(&gl->gl_lockref.lock);
 547                         if (rv) {
 548                                 do_error(gl, rv);
 549                                 goto out;
 550                         }
 551                 }
 552                 rv = do_promote(gl);
 553                 if (rv == 2)
 554                         goto out_locked;
 555         }
 556 out:
 557         clear_bit(GLF_LOCK, &gl->gl_flags);
 558 out_locked:
 559         spin_unlock(&gl->gl_lockref.lock);
 560 }
 561
 562 /**
 563  * do_xmote - Calls the DLM to change the state of a lock
 564  * @gl: The lock state
 565  * @gh: The holder (only for promotes)
 566  * @target: The target lock state
 567  *
 568  */
 569
 570 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 571 __releases(&gl->gl_lockref.lock)
 572 __acquires(&gl->gl_lockref.lock)
 573 {
 574         const struct gfs2_glock_operations *glops = gl->gl_ops;
 575         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 576         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 577         int ret;
 578
 579         if (target != LM_ST_UNLOCKED && glock_blocked_by_withdraw(gl) &&
 580             gh && !(gh->gh_flags & LM_FLAG_NOEXP))
 581                 return;
 582         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 583                       LM_FLAG_PRIORITY);
 584         GLOCK_BUG_ON(gl, gl->gl_state == target);
 585         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 586         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 587             glops->go_inval) {
 588                 /*
 589                  * If another process is already doing the invalidate, let that
 590                  * finish first.  The glock state machine will get back to this
 591                  * holder again later.
 592                  */
 593                 if (test_and_set_bit(GLF_INVALIDATE_IN_PROGRESS,
 594                                      &gl->gl_flags))
 595                         return;
 596                 do_error(gl, 0); /* Fail queued try locks */
 597         }
 598         gl->gl_req = target;
 599         set_bit(GLF_BLOCKING, &gl->gl_flags);
 600         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 601             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 602             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 603                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 604         spin_unlock(&gl->gl_lockref.lock);
 605         if (glops->go_sync) {
 606                 ret = glops->go_sync(gl);
 607                 /* If we had a problem syncing (due to io errors or whatever,
 608                  * we should not invalidate the metadata or tell dlm to
 609                  * release the glock to other nodes.
 610                  */
 611                 if (ret) {
 612                         if (cmpxchg(&sdp->sd_log_error, 0, ret)) {
 613                                 fs_err(sdp, "Error %d syncing glock \n", ret);
 614                                 gfs2_dump_glock(NULL, gl, true);
 615                         }
 616                         goto skip_inval;
 617                 }
 618         }
 619         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags)) {
 620                 /*
 621                  * The call to go_sync should have cleared out the ail list.
 622                  * If there are still items, we have a problem. We ought to
 623                  * withdraw, but we can't because the withdraw code also uses
 624                  * glocks. Warn about the error, dump the glock, then fall
 625                  * through and wait for logd to do the withdraw for us.
 626                  */
 627                 if ((atomic_read(&gl->gl_ail_count) != 0) &&
 628                     (!cmpxchg(&sdp->sd_log_error, 0, -EIO))) {
 629                         gfs2_assert_warn(sdp, !atomic_read(&gl->gl_ail_count));
 630                         gfs2_dump_glock(NULL, gl, true);
 631                 }
 632                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 633                 clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 634         }
 635
 636 skip_inval:
 637         gfs2_glock_hold(gl);
 638         /*
 639          * Check for an error encountered since we called go_sync and go_inval.
 640          * If so, we can't withdraw from the glock code because the withdraw
 641          * code itself uses glocks (see function signal_our_withdraw) to
 642          * change the mount to read-only. Most importantly, we must not call
 643          * dlm to unlock the glock until the journal is in a known good state
 644          * (after journal replay) otherwise other nodes may use the object
 645          * (rgrp or dinode) and then later, journal replay will corrupt the
 646          * file system. The best we can do here is wait for the logd daemon
 647          * to see sd_log_error and withdraw, and in the meantime, requeue the
 648          * work for later.
 649          *
 650          * However, if we're just unlocking the lock (say, for unmount, when
 651          * gfs2_gl_hash_clear calls clear_glock) and recovery is complete
 652          * then it's okay to tell dlm to unlock it.
 653          */
 654         if (unlikely(sdp->sd_log_error && !gfs2_withdrawn(sdp)))
 655                 gfs2_withdraw_delayed(sdp);
 656         if (glock_blocked_by_withdraw(gl)) {
 657                 if (target != LM_ST_UNLOCKED ||
 658                     test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags)) {
 659                         gfs2_glock_queue_work(gl, GL_GLOCK_DFT_HOLD);
 660                         goto out;
 661                 }
 662         }
 663
 664         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 665                 /* lock_dlm */
 666                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 667                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 668                     target == LM_ST_UNLOCKED &&
 669                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 670                         finish_xmote(gl, target);
 671                         gfs2_glock_queue_work(gl, 0);
 672                 } else if (ret) {
 673                         fs_err(sdp, "lm_lock ret %d\n", ret);
 674                         GLOCK_BUG_ON(gl, !gfs2_withdrawn(sdp));
 675                 }
 676         } else { /* lock_nolock */
 677                 finish_xmote(gl, target);
 678                 gfs2_glock_queue_work(gl, 0);
 679         }
 680 out:
 681         spin_lock(&gl->gl_lockref.lock);
 682 }
 683
 684 /**
 685  * find_first_holder - find the first "holder" gh
 686  * @gl: the glock
 687  */
 688
 689 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 690 {
 691         struct gfs2_holder *gh;
 692
 693         if (!list_empty(&gl->gl_holders)) {
 694                 gh = list_first_entry(&gl->gl_holders, struct gfs2_holder, gh_list);
 695                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 696                         return gh;
 697         }
 698         return NULL;
 699 }
 700
 701 /**
 702  * run_queue - do all outstanding tasks related to a glock
 703  * @gl: The glock in question
 704  * @nonblock: True if we must not block in run_queue
 705  *
 706  */
 707
 708 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 709 __releases(&gl->gl_lockref.lock)
 710 __acquires(&gl->gl_lockref.lock)
 711 {
 712         struct gfs2_holder *gh = NULL;
 713         int ret;
 714
 715         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 716                 return;
 717
 718         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 719
 720         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 721             gl->gl_demote_state != gl->gl_state) {
 722                 if (find_first_holder(gl))
 723                         goto out_unlock;
 724                 if (nonblock)
 725                         goto out_sched;
 726                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 727                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 728                 gl->gl_target = gl->gl_demote_state;
 729         } else {
 730                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 731                         gfs2_demote_wake(gl);
 732                 ret = do_promote(gl);
 733                 if (ret == 0)
 734                         goto out_unlock;
 735                 if (ret == 2)
 736                         goto out;
 737                 gh = find_first_waiter(gl);
 738                 gl->gl_target = gh->gh_state;
 739                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 740                         do_error(gl, 0); /* Fail queued try locks */
 741         }
 742         do_xmote(gl, gh, gl->gl_target);
 743 out:
 744         return;
 745
 746 out_sched:
 747         clear_bit(GLF_LOCK, &gl->gl_flags);
 748         smp_mb__after_atomic();
 749         gl->gl_lockref.count++;
 750         __gfs2_glock_queue_work(gl, 0);
 751         return;
 752
 753 out_unlock:
 754         clear_bit(GLF_LOCK, &gl->gl_flags);
 755         smp_mb__after_atomic();
 756         return;
 757 }
 758
 759 static void delete_work_func(struct work_struct *work)
 760 {
 761         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 762         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 763         struct inode *inode;
 764         u64 no_addr = gl->gl_name.ln_number;
 765
 766         /* If someone's using this glock to create a new dinode, the block must
 767            have been freed by another node, then re-used, in which case our
 768            iopen callback is too late after the fact. Ignore it. */
 769         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 770                 goto out;
 771
 772         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 773         if (!IS_ERR_OR_NULL(inode)) {
 774                 d_prune_aliases(inode);
 775                 iput(inode);
 776         }
 777 out:
 778         gfs2_glock_put(gl);
 779 }
 780
 781 static void glock_work_func(struct work_struct *work)
 782 {
 783         unsigned long delay = 0;
 784         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 785         unsigned int drop_refs = 1;
 786
 787         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 788                 finish_xmote(gl, gl->gl_reply);
 789                 drop_refs++;
 790         }
 791         spin_lock(&gl->gl_lockref.lock);
 792         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 793             gl->gl_state != LM_ST_UNLOCKED &&
 794             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 795                 unsigned long holdtime, now = jiffies;
 796
 797                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 798                 if (time_before(now, holdtime))
 799                         delay = holdtime - now;
 800
 801                 if (!delay) {
 802                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 803                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 804                 }
 805         }
 806         run_queue(gl, 0);
 807         if (delay) {
 808                 /* Keep one glock reference for the work we requeue. */
 809                 drop_refs--;
 810                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 811                         delay = 0;
 812                 __gfs2_glock_queue_work(gl, delay);
 813         }
 814
 815         /*
 816          * Drop the remaining glock references manually here. (Mind that
 817          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 818          * here as well.)
 819          */
 820         gl->gl_lockref.count -= drop_refs;
 821         if (!gl->gl_lockref.count) {
 822                 __gfs2_glock_put(gl);
 823                 return;
 824         }
 825         spin_unlock(&gl->gl_lockref.lock);
 826 }
 827
 828 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 829                                             struct gfs2_glock *new)
 830 {
 831         struct wait_glock_queue wait;
 832         wait_queue_head_t *wq = glock_waitqueue(name);
 833         struct gfs2_glock *gl;
 834
 835         wait.name = name;
 836         init_wait(&wait.wait);
 837         wait.wait.func = glock_wake_function;
 838
 839 again:
 840         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 841         rcu_read_lock();
 842         if (new) {
 843                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 844                         &new->gl_node, ht_parms);
 845                 if (IS_ERR(gl))
 846                         goto out;
 847         } else {
 848                 gl = rhashtable_lookup_fast(&gl_hash_table,
 849                         name, ht_parms);
 850         }
 851         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 852                 rcu_read_unlock();
 853                 schedule();
 854                 goto again;
 855         }
 856 out:
 857         rcu_read_unlock();
 858         finish_wait(wq, &wait.wait);
 859         return gl;
 860 }
 861
 862 /**
 863  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 864  * @sdp: The GFS2 superblock
 865  * @number: the lock number
 866  * @glops: The glock_operations to use
 867  * @create: If 0, don't create the glock if it doesn't exist
 868  * @glp: the glock is returned here
 869  *
 870  * This does not lock a glock, just finds/creates structures for one.
 871  *
 872  * Returns: errno
 873  */
 874
 875 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 876                    const struct gfs2_glock_operations *glops, int create,
 877                    struct gfs2_glock **glp)
 878 {
 879         struct super_block *s = sdp->sd_vfs;
 880         struct lm_lockname name = { .ln_number = number,
 881                                     .ln_type = glops->go_type,
 882                                     .ln_sbd = sdp };
 883         struct gfs2_glock *gl, *tmp;
 884         struct address_space *mapping;
 885         struct kmem_cache *cachep;
 886         int ret = 0;
 887
 888         gl = find_insert_glock(&name, NULL);
 889         if (gl) {
 890                 *glp = gl;
 891                 return 0;
 892         }
 893         if (!create)
 894                 return -ENOENT;
 895
 896         if (glops->go_flags & GLOF_ASPACE)
 897                 cachep = gfs2_glock_aspace_cachep;
 898         else
 899                 cachep = gfs2_glock_cachep;
 900         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 901         if (!gl)
 902                 return -ENOMEM;
 903
 904         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 905
 906         if (glops->go_flags & GLOF_LVB) {
 907                 gl->gl_lksb.sb_lvbptr = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
 908                 if (!gl->gl_lksb.sb_lvbptr) {
 909                         kmem_cache_free(cachep, gl);
 910                         return -ENOMEM;
 911                 }
 912         }
 913
 914         atomic_inc(&sdp->sd_glock_disposal);
 915         gl->gl_node.next = NULL;
 916         gl->gl_flags = 0;
 917         gl->gl_name = name;
 918         gl->gl_lockref.count = 1;
 919         gl->gl_state = LM_ST_UNLOCKED;
 920         gl->gl_target = LM_ST_UNLOCKED;
 921         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 922         gl->gl_ops = glops;
 923         gl->gl_dstamp = 0;
 924         preempt_disable();
 925         /* We use the global stats to estimate the initial per-glock stats */
 926         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 927         preempt_enable();
 928         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 929         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 930         gl->gl_tchange = jiffies;
 931         gl->gl_object = NULL;
 932         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 933         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 934         INIT_WORK(&gl->gl_delete, delete_work_func);
 935
 936         mapping = gfs2_glock2aspace(gl);
 937         if (mapping) {
 938                 mapping->a_ops = &gfs2_meta_aops;
 939                 mapping->host = s->s_bdev->bd_inode;
 940                 mapping->flags = 0;
 941                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 942                 mapping->private_data = NULL;
 943                 mapping->writeback_index = 0;
 944         }
 945
 946         tmp = find_insert_glock(&name, gl);
 947         if (!tmp) {
 948                 *glp = gl;
 949                 goto out;
 950         }
 951         if (IS_ERR(tmp)) {
 952                 ret = PTR_ERR(tmp);
 953                 goto out_free;
 954         }
 955         *glp = tmp;
 956
 957 out_free:
 958         kfree(gl->gl_lksb.sb_lvbptr);
 959         kmem_cache_free(cachep, gl);
 960         atomic_dec(&sdp->sd_glock_disposal);
 961
 962 out:
 963         return ret;
 964 }
 965
 966 /**
 967  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 968  * @gl: the glock
 969  * @state: the state we're requesting
 970  * @flags: the modifier flags
 971  * @gh: the holder structure
 972  *
 973  */
 974
 975 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 976                       struct gfs2_holder *gh)
 977 {
 978         INIT_LIST_HEAD(&gh->gh_list);
 979         gh->gh_gl = gl;
 980         gh->gh_ip = _RET_IP_;
 981         gh->gh_owner_pid = get_pid(task_pid(current));
 982         gh->gh_state = state;
 983         gh->gh_flags = flags;
 984         gh->gh_error = 0;
 985         gh->gh_iflags = 0;
 986         gfs2_glock_hold(gl);
 987 }
 988
 989 /**
 990  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 991  * @state: the state we're requesting
 992  * @flags: the modifier flags
 993  * @gh: the holder structure
 994  *
 995  * Don't mess with the glock.
 996  *
 997  */
 998
 999 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
1000 {
1001         gh->gh_state = state;
1002         gh->gh_flags = flags;
1003         gh->gh_iflags = 0;
1004         gh->gh_ip = _RET_IP_;
1005         put_pid(gh->gh_owner_pid);
1006         gh->gh_owner_pid = get_pid(task_pid(current));
1007 }
1008
1009 /**
1010  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
1011  * @gh: the holder structure
1012  *
1013  */
1014
1015 void gfs2_holder_uninit(struct gfs2_holder *gh)
1016 {
1017         put_pid(gh->gh_owner_pid);
1018         gfs2_glock_put(gh->gh_gl);
1019         gfs2_holder_mark_uninitialized(gh);
1020         gh->gh_ip = 0;
1021 }
1022
1023 static void gfs2_glock_update_hold_time(struct gfs2_glock *gl,
1024                                         unsigned long start_time)
1025 {
1026         /* Have we waited longer that a second? */
1027         if (time_after(jiffies, start_time + HZ)) {
1028                 /* Lengthen the minimum hold time. */
1029                 gl->gl_hold_time = min(gl->gl_hold_time + GL_GLOCK_HOLD_INCR,
1030                                        GL_GLOCK_MAX_HOLD);
1031         }
1032 }
1033
1034 /**
1035  * gfs2_glock_wait - wait on a glock acquisition
1036  * @gh: the glock holder
1037  *
1038  * Returns: 0 on success
1039  */
1040
1041 int gfs2_glock_wait(struct gfs2_holder *gh)
1042 {
1043         unsigned long start_time = jiffies;
1044
1045         might_sleep();
1046         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
1047         gfs2_glock_update_hold_time(gh->gh_gl, start_time);
1048         return gh->gh_error;
1049 }
1050
1051 static int glocks_pending(unsigned int num_gh, struct gfs2_holder *ghs)
1052 {
1053         int i;
1054
1055         for (i = 0; i < num_gh; i++)
1056                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags))
1057                         return 1;
1058         return 0;
1059 }
1060
1061 /**
1062  * gfs2_glock_async_wait - wait on multiple asynchronous glock acquisitions
1063  * @num_gh: the number of holders in the array
1064  * @ghs: the glock holder array
1065  *
1066  * Returns: 0 on success, meaning all glocks have been granted and are held.
1067  *          -ESTALE if the request timed out, meaning all glocks were released,
1068  *          and the caller should retry the operation.
1069  */
1070
1071 int gfs2_glock_async_wait(unsigned int num_gh, struct gfs2_holder *ghs)
1072 {
1073         struct gfs2_sbd *sdp = ghs[0].gh_gl->gl_name.ln_sbd;
1074         int i, ret = 0, timeout = 0;
1075         unsigned long start_time = jiffies;
1076         bool keep_waiting;
1077
1078         might_sleep();
1079         /*
1080          * Total up the (minimum hold time * 2) of all glocks and use that to
1081          * determine the max amount of time we should wait.
1082          */
1083         for (i = 0; i < num_gh; i++)
1084                 timeout += ghs[i].gh_gl->gl_hold_time << 1;
1085
1086 wait_for_dlm:
1087         if (!wait_event_timeout(sdp->sd_async_glock_wait,
1088                                 !glocks_pending(num_gh, ghs), timeout))
1089                 ret = -ESTALE; /* request timed out. */
1090
1091         /*
1092          * If dlm granted all our requests, we need to adjust the glock
1093          * minimum hold time values according to how long we waited.
1094          *
1095          * If our request timed out, we need to repeatedly release any held
1096          * glocks we acquired thus far to allow dlm to acquire the remaining
1097          * glocks without deadlocking.  We cannot currently cancel outstanding
1098          * glock acquisitions.
1099          *
1100          * The HIF_WAIT bit tells us which requests still need a response from
1101          * dlm.
1102          *
1103          * If dlm sent us any errors, we return the first error we find.
1104          */
1105         keep_waiting = false;
1106         for (i = 0; i < num_gh; i++) {
1107                 /* Skip holders we have already dequeued below. */
1108                 if (!gfs2_holder_queued(&ghs[i]))
1109                         continue;
1110                 /* Skip holders with a pending DLM response. */
1111                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags)) {
1112                         keep_waiting = true;
1113                         continue;
1114                 }
1115
1116                 if (test_bit(HIF_HOLDER, &ghs[i].gh_iflags)) {
1117                         if (ret == -ESTALE)
1118                                 gfs2_glock_dq(&ghs[i]);
1119                         else
1120                                 gfs2_glock_update_hold_time(ghs[i].gh_gl,
1121                                                             start_time);
1122                 }
1123                 if (!ret)
1124                         ret = ghs[i].gh_error;
1125         }
1126
1127         if (keep_waiting)
1128                 goto wait_for_dlm;
1129
1130         /*
1131          * At this point, we've either acquired all locks or released them all.
1132          */
1133         return ret;
1134 }
1135
1136 /**
1137  * handle_callback - process a demote request
1138  * @gl: the glock
1139  * @state: the state the caller wants us to change to
1140  *
1141  * There are only two requests that we are going to see in actual
1142  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
1143  */
1144
1145 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
1146                             unsigned long delay, bool remote)
1147 {
1148         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
1149
1150         set_bit(bit, &gl->gl_flags);
1151         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
1152                 gl->gl_demote_state = state;
1153                 gl->gl_demote_time = jiffies;
1154         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
1155                         gl->gl_demote_state != state) {
1156                 gl->gl_demote_state = LM_ST_UNLOCKED;
1157         }
1158         if (gl->gl_ops->go_callback)
1159                 gl->gl_ops->go_callback(gl, remote);
1160         trace_gfs2_demote_rq(gl, remote);
1161 }
1162
1163 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
1164 {
1165         struct va_format vaf;
1166         va_list args;
1167
1168         va_start(args, fmt);
1169
1170         if (seq) {
1171                 seq_vprintf(seq, fmt, args);
1172         } else {
1173                 vaf.fmt = fmt;
1174                 vaf.va = &args;
1175
1176                 pr_err("%pV", &vaf);
1177         }
1178
1179         va_end(args);
1180 }
1181
1182 /**
1183  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1184  * @gh: the holder structure to add
1185  *
1186  * Eventually we should move the recursive locking trap to a
1187  * debugging option or something like that. This is the fast
1188  * path and needs to have the minimum number of distractions.
1189  *
1190  */
1191
1192 static inline void add_to_queue(struct gfs2_holder *gh)
1193 __releases(&gl->gl_lockref.lock)
1194 __acquires(&gl->gl_lockref.lock)
1195 {
1196         struct gfs2_glock *gl = gh->gh_gl;
1197         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1198         struct list_head *insert_pt = NULL;
1199         struct gfs2_holder *gh2;
1200         int try_futile = 0;
1201
1202         GLOCK_BUG_ON(gl, gh->gh_owner_pid == NULL);
1203         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1204                 GLOCK_BUG_ON(gl, true);
1205
1206         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1207                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1208                         try_futile = !may_grant(gl, gh);
1209                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1210                         goto fail;
1211         }
1212
1213         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1214                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1215                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1216                         goto trap_recursive;
1217                 if (try_futile &&
1218                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1219 fail:
1220                         gh->gh_error = GLR_TRYFAILED;
1221                         gfs2_holder_wake(gh);
1222                         return;
1223                 }
1224                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1225                         continue;
1226                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1227                         insert_pt = &gh2->gh_list;
1228         }
1229         set_bit(GLF_QUEUED, &gl->gl_flags);
1230         trace_gfs2_glock_queue(gh, 1);
1231         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1232         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1233         if (likely(insert_pt == NULL)) {
1234                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1235                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1236                         goto do_cancel;
1237                 return;
1238         }
1239         list_add_tail(&gh->gh_list, insert_pt);
1240 do_cancel:
1241         gh = list_first_entry(&gl->gl_holders, struct gfs2_holder, gh_list);
1242         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1243                 spin_unlock(&gl->gl_lockref.lock);
1244                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1245                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1246                 spin_lock(&gl->gl_lockref.lock);
1247         }
1248         return;
1249
1250 trap_recursive:
1251         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1252         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1253         fs_err(sdp, "lock type: %d req lock state : %d\n",
1254                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1255         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1256         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1257         fs_err(sdp, "lock type: %d req lock state : %d\n",
1258                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1259         gfs2_dump_glock(NULL, gl, true);
1260         BUG();
1261 }
1262
1263 /**
1264  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1265  * @gh: the holder structure
1266  *
1267  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1268  *
1269  * Returns: 0, GLR_TRYFAILED, or errno on failure
1270  */
1271
1272 int gfs2_glock_nq(struct gfs2_holder *gh)
1273 {
1274         struct gfs2_glock *gl = gh->gh_gl;
1275         int error = 0;
1276
1277         if (glock_blocked_by_withdraw(gl) && !(gh->gh_flags & LM_FLAG_NOEXP))
1278                 return -EIO;
1279
1280         if (test_bit(GLF_LRU, &gl->gl_flags))
1281                 gfs2_glock_remove_from_lru(gl);
1282
1283         spin_lock(&gl->gl_lockref.lock);
1284         add_to_queue(gh);
1285         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1286                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1287                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1288                 gl->gl_lockref.count++;
1289                 __gfs2_glock_queue_work(gl, 0);
1290         }
1291         run_queue(gl, 1);
1292         spin_unlock(&gl->gl_lockref.lock);
1293
1294         if (!(gh->gh_flags & GL_ASYNC))
1295                 error = gfs2_glock_wait(gh);
1296
1297         return error;
1298 }
1299
1300 /**
1301  * gfs2_glock_poll - poll to see if an async request has been completed
1302  * @gh: the holder
1303  *
1304  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1305  */
1306
1307 int gfs2_glock_poll(struct gfs2_holder *gh)
1308 {
1309         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1310 }
1311
1312 /**
1313  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1314  * @gh: the glock holder
1315  *
1316  */
1317
1318 void gfs2_glock_dq(struct gfs2_holder *gh)
1319 {
1320         struct gfs2_glock *gl = gh->gh_gl;
1321         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1322         unsigned delay = 0;
1323         int fast_path = 0;
1324
1325         spin_lock(&gl->gl_lockref.lock);
1326         /*
1327          * If we're in the process of file system withdraw, we cannot just
1328          * dequeue any glocks until our journal is recovered, lest we
1329          * introduce file system corruption. We need two exceptions to this
1330          * rule: We need to allow unlocking of nondisk glocks and the glock
1331          * for our own journal that needs recovery.
1332          */
1333         if (test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags) &&
1334             glock_blocked_by_withdraw(gl) &&
1335             gh->gh_gl != sdp->sd_jinode_gl) {
1336                 sdp->sd_glock_dqs_held++;
1337                 might_sleep();
1338                 wait_on_bit(&sdp->sd_flags, SDF_WITHDRAW_RECOVERY,
1339                             TASK_UNINTERRUPTIBLE);
1340         }
1341         if (gh->gh_flags & GL_NOCACHE)
1342                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1343
1344         list_del_init(&gh->gh_list);
1345         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1346         if (find_first_holder(gl) == NULL) {
1347                 if (list_empty(&gl->gl_holders) &&
1348                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1349                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1350                         fast_path = 1;
1351         }
1352         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1353                 gfs2_glock_add_to_lru(gl);
1354
1355         trace_gfs2_glock_queue(gh, 0);
1356         if (unlikely(!fast_path)) {
1357                 gl->gl_lockref.count++;
1358                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1359                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1360                     gl->gl_name.ln_type == LM_TYPE_INODE)
1361                         delay = gl->gl_hold_time;
1362                 __gfs2_glock_queue_work(gl, delay);
1363         }
1364         spin_unlock(&gl->gl_lockref.lock);
1365 }
1366
1367 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1368 {
1369         struct gfs2_glock *gl = gh->gh_gl;
1370         gfs2_glock_dq(gh);
1371         might_sleep();
1372         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1373 }
1374
1375 /**
1376  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1377  * @gh: the holder structure
1378  *
1379  */
1380
1381 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1382 {
1383         gfs2_glock_dq(gh);
1384         gfs2_holder_uninit(gh);
1385 }
1386
1387 /**
1388  * gfs2_glock_nq_num - acquire a glock based on lock number
1389  * @sdp: the filesystem
1390  * @number: the lock number
1391  * @glops: the glock operations for the type of glock
1392  * @state: the state to acquire the glock in
1393  * @flags: modifier flags for the acquisition
1394  * @gh: the struct gfs2_holder
1395  *
1396  * Returns: errno
1397  */
1398
1399 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1400                       const struct gfs2_glock_operations *glops,
1401                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1402 {
1403         struct gfs2_glock *gl;
1404         int error;
1405
1406         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1407         if (!error) {
1408                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1409                 gfs2_glock_put(gl);
1410         }
1411
1412         return error;
1413 }
1414
1415 /**
1416  * glock_compare - Compare two struct gfs2_glock structures for sorting
1417  * @arg_a: the first structure
1418  * @arg_b: the second structure
1419  *
1420  */
1421
1422 static int glock_compare(const void *arg_a, const void *arg_b)
1423 {
1424         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1425         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1426         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1427         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1428
1429         if (a->ln_number > b->ln_number)
1430                 return 1;
1431         if (a->ln_number < b->ln_number)
1432                 return -1;
1433         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1434         return 0;
1435 }
1436
1437 /**
1438  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1439  * @num_gh: the number of structures
1440  * @ghs: an array of struct gfs2_holder structures
1441  *
1442  * Returns: 0 on success (all glocks acquired),
1443  *          errno on failure (no glocks acquired)
1444  */
1445
1446 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1447                      struct gfs2_holder **p)
1448 {
1449         unsigned int x;
1450         int error = 0;
1451
1452         for (x = 0; x < num_gh; x++)
1453                 p[x] = &ghs[x];
1454
1455         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1456
1457         for (x = 0; x < num_gh; x++) {
1458                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1459
1460                 error = gfs2_glock_nq(p[x]);
1461                 if (error) {
1462                         while (x--)
1463                                 gfs2_glock_dq(p[x]);
1464                         break;
1465                 }
1466         }
1467
1468         return error;
1469 }
1470
1471 /**
1472  * gfs2_glock_nq_m - acquire multiple glocks
1473  * @num_gh: the number of structures
1474  * @ghs: an array of struct gfs2_holder structures
1475  *
1476  *
1477  * Returns: 0 on success (all glocks acquired),
1478  *          errno on failure (no glocks acquired)
1479  */
1480
1481 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1482 {
1483         struct gfs2_holder *tmp[4];
1484         struct gfs2_holder **pph = tmp;
1485         int error = 0;
1486
1487         switch(num_gh) {
1488         case 0:
1489                 return 0;
1490         case 1:
1491                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1492                 return gfs2_glock_nq(ghs);
1493         default:
1494                 if (num_gh <= 4)
1495                         break;
1496                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1497                                     GFP_NOFS);
1498                 if (!pph)
1499                         return -ENOMEM;
1500         }
1501
1502         error = nq_m_sync(num_gh, ghs, pph);
1503
1504         if (pph != tmp)
1505                 kfree(pph);
1506
1507         return error;
1508 }
1509
1510 /**
1511  * gfs2_glock_dq_m - release multiple glocks
1512  * @num_gh: the number of structures
1513  * @ghs: an array of struct gfs2_holder structures
1514  *
1515  */
1516
1517 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1518 {
1519         while (num_gh--)
1520                 gfs2_glock_dq(&ghs[num_gh]);
1521 }
1522
1523 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1524 {
1525         unsigned long delay = 0;
1526         unsigned long holdtime;
1527         unsigned long now = jiffies;
1528
1529         gfs2_glock_hold(gl);
1530         holdtime = gl->gl_tchange + gl->gl_hold_time;
1531         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1532             gl->gl_name.ln_type == LM_TYPE_INODE) {
1533                 if (time_before(now, holdtime))
1534                         delay = holdtime - now;
1535                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1536                         delay = gl->gl_hold_time;
1537         }
1538
1539         spin_lock(&gl->gl_lockref.lock);
1540         handle_callback(gl, state, delay, true);
1541         __gfs2_glock_queue_work(gl, delay);
1542         spin_unlock(&gl->gl_lockref.lock);
1543 }
1544
1545 /**
1546  * gfs2_should_freeze - Figure out if glock should be frozen
1547  * @gl: The glock in question
1548  *
1549  * Glocks are not frozen if (a) the result of the dlm operation is
1550  * an error, (b) the locking operation was an unlock operation or
1551  * (c) if there is a "noexp" flagged request anywhere in the queue
1552  *
1553  * Returns: 1 if freezing should occur, 0 otherwise
1554  */
1555
1556 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1557 {
1558         const struct gfs2_holder *gh;
1559
1560         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1561                 return 0;
1562         if (gl->gl_target == LM_ST_UNLOCKED)
1563                 return 0;
1564
1565         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1566                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1567                         continue;
1568                 if (LM_FLAG_NOEXP & gh->gh_flags)
1569                         return 0;
1570         }
1571
1572         return 1;
1573 }
1574
1575 /**
1576  * gfs2_glock_complete - Callback used by locking
1577  * @gl: Pointer to the glock
1578  * @ret: The return value from the dlm
1579  *
1580  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1581  * to use a bitfield shared with other glock state fields.
1582  */
1583
1584 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1585 {
1586         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1587
1588         spin_lock(&gl->gl_lockref.lock);
1589         gl->gl_reply = ret;
1590
1591         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1592                 if (gfs2_should_freeze(gl)) {
1593                         set_bit(GLF_FROZEN, &gl->gl_flags);
1594                         spin_unlock(&gl->gl_lockref.lock);
1595                         return;
1596                 }
1597         }
1598
1599         gl->gl_lockref.count++;
1600         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1601         __gfs2_glock_queue_work(gl, 0);
1602         spin_unlock(&gl->gl_lockref.lock);
1603 }
1604
1605 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1606 {
1607         struct gfs2_glock *gla, *glb;
1608
1609         gla = list_entry(a, struct gfs2_glock, gl_lru);
1610         glb = list_entry(b, struct gfs2_glock, gl_lru);
1611
1612         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1613                 return 1;
1614         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1615                 return -1;
1616
1617         return 0;
1618 }
1619
1620 /**
1621  * gfs2_dispose_glock_lru - Demote a list of glocks
1622  * @list: The list to dispose of
1623  *
1624  * Disposing of glocks may involve disk accesses, so that here we sort
1625  * the glocks by number (i.e. disk location of the inodes) so that if
1626  * there are any such accesses, they'll be sent in order (mostly).
1627  *
1628  * Must be called under the lru_lock, but may drop and retake this
1629  * lock. While the lru_lock is dropped, entries may vanish from the
1630  * list, but no new entries will appear on the list (since it is
1631  * private)
1632  */
1633
1634 static void gfs2_dispose_glock_lru(struct list_head *list)
1635 __releases(&lru_lock)
1636 __acquires(&lru_lock)
1637 {
1638         struct gfs2_glock *gl;
1639
1640         list_sort(NULL, list, glock_cmp);
1641
1642         while(!list_empty(list)) {
1643                 gl = list_first_entry(list, struct gfs2_glock, gl_lru);
1644                 list_del_init(&gl->gl_lru);
1645                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1646 add_back_to_lru:
1647                         list_add(&gl->gl_lru, &lru_list);
1648                         set_bit(GLF_LRU, &gl->gl_flags);
1649                         atomic_inc(&lru_count);
1650                         continue;
1651                 }
1652                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1653                         spin_unlock(&gl->gl_lockref.lock);
1654                         goto add_back_to_lru;
1655                 }
1656                 gl->gl_lockref.count++;
1657                 if (demote_ok(gl))
1658                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1659                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1660                 __gfs2_glock_queue_work(gl, 0);
1661                 spin_unlock(&gl->gl_lockref.lock);
1662                 cond_resched_lock(&lru_lock);
1663         }
1664 }
1665
1666 /**
1667  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1668  * @nr: The number of entries to scan
1669  *
1670  * This function selects the entries on the LRU which are able to
1671  * be demoted, and then kicks off the process by calling
1672  * gfs2_dispose_glock_lru() above.
1673  */
1674
1675 static long gfs2_scan_glock_lru(int nr)
1676 {
1677         struct gfs2_glock *gl;
1678         LIST_HEAD(skipped);
1679         LIST_HEAD(dispose);
1680         long freed = 0;
1681
1682         spin_lock(&lru_lock);
1683         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1684                 gl = list_first_entry(&lru_list, struct gfs2_glock, gl_lru);
1685
1686                 /* Test for being demotable */
1687                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1688                         list_move(&gl->gl_lru, &dispose);
1689                         atomic_dec(&lru_count);
1690                         clear_bit(GLF_LRU, &gl->gl_flags);
1691                         freed++;
1692                         continue;
1693                 }
1694
1695                 list_move(&gl->gl_lru, &skipped);
1696         }
1697         list_splice(&skipped, &lru_list);
1698         if (!list_empty(&dispose))
1699                 gfs2_dispose_glock_lru(&dispose);
1700         spin_unlock(&lru_lock);
1701
1702         return freed;
1703 }
1704
1705 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1706                                             struct shrink_control *sc)
1707 {
1708         if (!(sc->gfp_mask & __GFP_FS))
1709                 return SHRINK_STOP;
1710         return gfs2_scan_glock_lru(sc->nr_to_scan);
1711 }
1712
1713 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1714                                              struct shrink_control *sc)
1715 {
1716         return vfs_pressure_ratio(atomic_read(&lru_count));
1717 }
1718
1719 static struct shrinker glock_shrinker = {
1720         .seeks = DEFAULT_SEEKS,
1721         .count_objects = gfs2_glock_shrink_count,
1722         .scan_objects = gfs2_glock_shrink_scan,
1723 };
1724
1725 /**
1726  * examine_bucket - Call a function for glock in a hash bucket
1727  * @examiner: the function
1728  * @sdp: the filesystem
1729  * @bucket: the bucket
1730  *
1731  * Note that the function can be called multiple times on the same
1732  * object.  So the user must ensure that the function can cope with
1733  * that.
1734  */
1735
1736 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1737 {
1738         struct gfs2_glock *gl;
1739         struct rhashtable_iter iter;
1740
1741         rhashtable_walk_enter(&gl_hash_table, &iter);
1742
1743         do {
1744                 rhashtable_walk_start(&iter);
1745
1746                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1747                         if (gl->gl_name.ln_sbd == sdp &&
1748                             lockref_get_not_dead(&gl->gl_lockref))
1749                                 examiner(gl);
1750
1751                 rhashtable_walk_stop(&iter);
1752         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1753
1754         rhashtable_walk_exit(&iter);
1755 }
1756
1757 /**
1758  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1759  * @gl: The glock to thaw
1760  *
1761  */
1762
1763 static void thaw_glock(struct gfs2_glock *gl)
1764 {
1765         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1766                 gfs2_glock_put(gl);
1767                 return;
1768         }
1769         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1770         gfs2_glock_queue_work(gl, 0);
1771 }
1772
1773 /**
1774  * clear_glock - look at a glock and see if we can free it from glock cache
1775  * @gl: the glock to look at
1776  *
1777  */
1778
1779 static void clear_glock(struct gfs2_glock *gl)
1780 {
1781         gfs2_glock_remove_from_lru(gl);
1782
1783         spin_lock(&gl->gl_lockref.lock);
1784         if (gl->gl_state != LM_ST_UNLOCKED)
1785                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1786         __gfs2_glock_queue_work(gl, 0);
1787         spin_unlock(&gl->gl_lockref.lock);
1788 }
1789
1790 /**
1791  * gfs2_glock_thaw - Thaw any frozen glocks
1792  * @sdp: The super block
1793  *
1794  */
1795
1796 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1797 {
1798         glock_hash_walk(thaw_glock, sdp);
1799 }
1800
1801 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1802 {
1803         spin_lock(&gl->gl_lockref.lock);
1804         gfs2_dump_glock(seq, gl, fsid);
1805         spin_unlock(&gl->gl_lockref.lock);
1806 }
1807
1808 static void dump_glock_func(struct gfs2_glock *gl)
1809 {
1810         dump_glock(NULL, gl, true);
1811 }
1812
1813 /**
1814  * gfs2_gl_hash_clear - Empty out the glock hash table
1815  * @sdp: the filesystem
1816  * @wait: wait until it's all gone
1817  *
1818  * Called when unmounting the filesystem.
1819  */
1820
1821 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1822 {
1823         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1824         flush_workqueue(glock_workqueue);
1825         glock_hash_walk(clear_glock, sdp);
1826         flush_workqueue(glock_workqueue);
1827         wait_event_timeout(sdp->sd_glock_wait,
1828                            atomic_read(&sdp->sd_glock_disposal) == 0,
1829                            HZ * 600);
1830         glock_hash_walk(dump_glock_func, sdp);
1831 }
1832
1833 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1834 {
1835         struct gfs2_glock *gl = ip->i_gl;
1836         int ret;
1837
1838         ret = gfs2_truncatei_resume(ip);
1839         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1840
1841         spin_lock(&gl->gl_lockref.lock);
1842         clear_bit(GLF_LOCK, &gl->gl_flags);
1843         run_queue(gl, 1);
1844         spin_unlock(&gl->gl_lockref.lock);
1845 }
1846
1847 static const char *state2str(unsigned state)
1848 {
1849         switch(state) {
1850         case LM_ST_UNLOCKED:
1851                 return "UN";
1852         case LM_ST_SHARED:
1853                 return "SH";
1854         case LM_ST_DEFERRED:
1855                 return "DF";
1856         case LM_ST_EXCLUSIVE:
1857                 return "EX";
1858         }
1859         return "??";
1860 }
1861
1862 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1863 {
1864         char *p = buf;
1865         if (flags & LM_FLAG_TRY)
1866                 *p++ = 't';
1867         if (flags & LM_FLAG_TRY_1CB)
1868                 *p++ = 'T';
1869         if (flags & LM_FLAG_NOEXP)
1870                 *p++ = 'e';
1871         if (flags & LM_FLAG_ANY)
1872                 *p++ = 'A';
1873         if (flags & LM_FLAG_PRIORITY)
1874                 *p++ = 'p';
1875         if (flags & GL_ASYNC)
1876                 *p++ = 'a';
1877         if (flags & GL_EXACT)
1878                 *p++ = 'E';
1879         if (flags & GL_NOCACHE)
1880                 *p++ = 'c';
1881         if (test_bit(HIF_HOLDER, &iflags))
1882                 *p++ = 'H';
1883         if (test_bit(HIF_WAIT, &iflags))
1884                 *p++ = 'W';
1885         if (test_bit(HIF_FIRST, &iflags))
1886                 *p++ = 'F';
1887         *p = 0;
1888         return buf;
1889 }
1890
1891 /**
1892  * dump_holder - print information about a glock holder
1893  * @seq: the seq_file struct
1894  * @gh: the glock holder
1895  * @fs_id_buf: pointer to file system id (if requested)
1896  *
1897  */
1898
1899 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh,
1900                         const char *fs_id_buf)
1901 {
1902         struct task_struct *gh_owner = NULL;
1903         char flags_buf[32];
1904
1905         rcu_read_lock();
1906         if (gh->gh_owner_pid)
1907                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1908         gfs2_print_dbg(seq, "%s H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1909                        fs_id_buf, state2str(gh->gh_state),
1910                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1911                        gh->gh_error,
1912                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1913                        gh_owner ? gh_owner->comm : "(ended)",
1914                        (void *)gh->gh_ip);
1915         rcu_read_unlock();
1916 }
1917
1918 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1919 {
1920         const unsigned long *gflags = &gl->gl_flags;
1921         char *p = buf;
1922
1923         if (test_bit(GLF_LOCK, gflags))
1924                 *p++ = 'l';
1925         if (test_bit(GLF_DEMOTE, gflags))
1926                 *p++ = 'D';
1927         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1928                 *p++ = 'd';
1929         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1930                 *p++ = 'p';
1931         if (test_bit(GLF_DIRTY, gflags))
1932                 *p++ = 'y';
1933         if (test_bit(GLF_LFLUSH, gflags))
1934                 *p++ = 'f';
1935         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1936                 *p++ = 'i';
1937         if (test_bit(GLF_REPLY_PENDING, gflags))
1938                 *p++ = 'r';
1939         if (test_bit(GLF_INITIAL, gflags))
1940                 *p++ = 'I';
1941         if (test_bit(GLF_FROZEN, gflags))
1942                 *p++ = 'F';
1943         if (test_bit(GLF_QUEUED, gflags))
1944                 *p++ = 'q';
1945         if (test_bit(GLF_LRU, gflags))
1946                 *p++ = 'L';
1947         if (gl->gl_object)
1948                 *p++ = 'o';
1949         if (test_bit(GLF_BLOCKING, gflags))
1950                 *p++ = 'b';
1951         *p = 0;
1952         return buf;
1953 }
1954
1955 /**
1956  * gfs2_dump_glock - print information about a glock
1957  * @seq: The seq_file struct
1958  * @gl: the glock
1959  * @fsid: If true, also dump the file system id
1960  *
1961  * The file format is as follows:
1962  * One line per object, capital letters are used to indicate objects
1963  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1964  * other objects are indented by a single space and follow the glock to
1965  * which they are related. Fields are indicated by lower case letters
1966  * followed by a colon and the field value, except for strings which are in
1967  * [] so that its possible to see if they are composed of spaces for
1968  * example. The field's are n = number (id of the object), f = flags,
1969  * t = type, s = state, r = refcount, e = error, p = pid.
1970  *
1971  */
1972
1973 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1974 {
1975         const struct gfs2_glock_operations *glops = gl->gl_ops;
1976         unsigned long long dtime;
1977         const struct gfs2_holder *gh;
1978         char gflags_buf[32];
1979         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1980         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
1981
1982         memset(fs_id_buf, 0, sizeof(fs_id_buf));
1983         if (fsid && sdp) /* safety precaution */
1984                 sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
1985         dtime = jiffies - gl->gl_demote_time;
1986         dtime *= 1000000/HZ; /* demote time in uSec */
1987         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1988                 dtime = 0;
1989         gfs2_print_dbg(seq, "%sG:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d "
1990                        "v:%d r:%d m:%ld\n", fs_id_buf, state2str(gl->gl_state),
1991                   gl->gl_name.ln_type,
1992                   (unsigned long long)gl->gl_name.ln_number,
1993                   gflags2str(gflags_buf, gl),
1994                   state2str(gl->gl_target),
1995                   state2str(gl->gl_demote_state), dtime,
1996                   atomic_read(&gl->gl_ail_count),
1997                   atomic_read(&gl->gl_revokes),
1998                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1999
2000         list_for_each_entry(gh, &gl->gl_holders, gh_list)
2001                 dump_holder(seq, gh, fs_id_buf);
2002
2003         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
2004                 glops->go_dump(seq, gl, fs_id_buf);
2005 }
2006
2007 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
2008 {
2009         struct gfs2_glock *gl = iter_ptr;
2010
2011         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
2012                    gl->gl_name.ln_type,
2013                    (unsigned long long)gl->gl_name.ln_number,
2014                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
2015                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
2016                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
2017                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
2018                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
2019                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
2020                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
2021                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
2022         return 0;
2023 }
2024
2025 static const char *gfs2_gltype[] = {
2026         "type",
2027         "reserved",
2028         "nondisk",
2029         "inode",
2030         "rgrp",
2031         "meta",
2032         "iopen",
2033         "flock",
2034         "plock",
2035         "quota",
2036         "journal",
2037 };
2038
2039 static const char *gfs2_stype[] = {
2040         [GFS2_LKS_SRTT]         = "srtt",
2041         [GFS2_LKS_SRTTVAR]      = "srttvar",
2042         [GFS2_LKS_SRTTB]        = "srttb",
2043         [GFS2_LKS_SRTTVARB]     = "srttvarb",
2044         [GFS2_LKS_SIRT]         = "sirt",
2045         [GFS2_LKS_SIRTVAR]      = "sirtvar",
2046         [GFS2_LKS_DCOUNT]       = "dlm",
2047         [GFS2_LKS_QCOUNT]       = "queue",
2048 };
2049
2050 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
2051
2052 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
2053 {
2054         struct gfs2_sbd *sdp = seq->private;
2055         loff_t pos = *(loff_t *)iter_ptr;
2056         unsigned index = pos >> 3;
2057         unsigned subindex = pos & 0x07;
2058         int i;
2059
2060         if (index == 0 && subindex != 0)
2061                 return 0;
2062
2063         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
2064                    (index == 0) ? "cpu": gfs2_stype[subindex]);
2065
2066         for_each_possible_cpu(i) {
2067                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
2068
2069                 if (index == 0)
2070                         seq_printf(seq, " %15u", i);
2071                 else
2072                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
2073                                    lkstats[index - 1].stats[subindex]);
2074         }
2075         seq_putc(seq, '\n');
2076         return 0;
2077 }
2078
2079 int __init gfs2_glock_init(void)
2080 {
2081         int i, ret;
2082
2083         ret = rhashtable_init(&gl_hash_table, &ht_parms);
2084         if (ret < 0)
2085                 return ret;
2086
2087         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
2088                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
2089         if (!glock_workqueue) {
2090                 rhashtable_destroy(&gl_hash_table);
2091                 return -ENOMEM;
2092         }
2093         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
2094                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
2095                                                 0);
2096         if (!gfs2_delete_workqueue) {
2097                 destroy_workqueue(glock_workqueue);
2098                 rhashtable_destroy(&gl_hash_table);
2099                 return -ENOMEM;
2100         }
2101
2102         ret = register_shrinker(&glock_shrinker);
2103         if (ret) {
2104                 destroy_workqueue(gfs2_delete_workqueue);
2105                 destroy_workqueue(glock_workqueue);
2106                 rhashtable_destroy(&gl_hash_table);
2107                 return ret;
2108         }
2109
2110         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
2111                 init_waitqueue_head(glock_wait_table + i);
2112
2113         return 0;
2114 }
2115
2116 void gfs2_glock_exit(void)
2117 {
2118         unregister_shrinker(&glock_shrinker);
2119         rhashtable_destroy(&gl_hash_table);
2120         destroy_workqueue(glock_workqueue);
2121         destroy_workqueue(gfs2_delete_workqueue);
2122 }
2123
2124 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
2125 {
2126         struct gfs2_glock *gl = gi->gl;
2127
2128         if (gl) {
2129                 if (n == 0)
2130                         return;
2131                 if (!lockref_put_not_zero(&gl->gl_lockref))
2132                         gfs2_glock_queue_put(gl);
2133         }
2134         for (;;) {
2135                 gl = rhashtable_walk_next(&gi->hti);
2136                 if (IS_ERR_OR_NULL(gl)) {
2137                         if (gl == ERR_PTR(-EAGAIN)) {
2138                                 n = 1;
2139                                 continue;
2140                         }
2141                         gl = NULL;
2142                         break;
2143                 }
2144                 if (gl->gl_name.ln_sbd != gi->sdp)
2145                         continue;
2146                 if (n <= 1) {
2147                         if (!lockref_get_not_dead(&gl->gl_lockref))
2148                                 continue;
2149                         break;
2150                 } else {
2151                         if (__lockref_is_dead(&gl->gl_lockref))
2152                                 continue;
2153                         n--;
2154                 }
2155         }
2156         gi->gl = gl;
2157 }
2158
2159 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
2160         __acquires(RCU)
2161 {
2162         struct gfs2_glock_iter *gi = seq->private;
2163         loff_t n;
2164
2165         /*
2166          * We can either stay where we are, skip to the next hash table
2167          * entry, or start from the beginning.
2168          */
2169         if (*pos < gi->last_pos) {
2170                 rhashtable_walk_exit(&gi->hti);
2171                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2172                 n = *pos + 1;
2173         } else {
2174                 n = *pos - gi->last_pos;
2175         }
2176
2177         rhashtable_walk_start(&gi->hti);
2178
2179         gfs2_glock_iter_next(gi, n);
2180         gi->last_pos = *pos;
2181         return gi->gl;
2182 }
2183
2184 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
2185                                  loff_t *pos)
2186 {
2187         struct gfs2_glock_iter *gi = seq->private;
2188
2189         (*pos)++;
2190         gi->last_pos = *pos;
2191         gfs2_glock_iter_next(gi, 1);
2192         return gi->gl;
2193 }
2194
2195 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2196         __releases(RCU)
2197 {
2198         struct gfs2_glock_iter *gi = seq->private;
2199
2200         rhashtable_walk_stop(&gi->hti);
2201 }
2202
2203 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2204 {
2205         dump_glock(seq, iter_ptr, false);
2206         return 0;
2207 }
2208
2209 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2210 {
2211         preempt_disable();
2212         if (*pos >= GFS2_NR_SBSTATS)
2213                 return NULL;
2214         return pos;
2215 }
2216
2217 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2218                                    loff_t *pos)
2219 {
2220         (*pos)++;
2221         if (*pos >= GFS2_NR_SBSTATS)
2222                 return NULL;
2223         return pos;
2224 }
2225
2226 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2227 {
2228         preempt_enable();
2229 }
2230
2231 static const struct seq_operations gfs2_glock_seq_ops = {
2232         .start = gfs2_glock_seq_start,
2233         .next  = gfs2_glock_seq_next,
2234         .stop  = gfs2_glock_seq_stop,
2235         .show  = gfs2_glock_seq_show,
2236 };
2237
2238 static const struct seq_operations gfs2_glstats_seq_ops = {
2239         .start = gfs2_glock_seq_start,
2240         .next  = gfs2_glock_seq_next,
2241         .stop  = gfs2_glock_seq_stop,
2242         .show  = gfs2_glstats_seq_show,
2243 };
2244
2245 static const struct seq_operations gfs2_sbstats_seq_ops = {
2246         .start = gfs2_sbstats_seq_start,
2247         .next  = gfs2_sbstats_seq_next,
2248         .stop  = gfs2_sbstats_seq_stop,
2249         .show  = gfs2_sbstats_seq_show,
2250 };
2251
2252 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2253
2254 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2255                               const struct seq_operations *ops)
2256 {
2257         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2258         if (ret == 0) {
2259                 struct seq_file *seq = file->private_data;
2260                 struct gfs2_glock_iter *gi = seq->private;
2261
2262                 gi->sdp = inode->i_private;
2263                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2264                 if (seq->buf)
2265                         seq->size = GFS2_SEQ_GOODSIZE;
2266                 /*
2267                  * Initially, we are "before" the first hash table entry; the
2268                  * first call to rhashtable_walk_next gets us the first entry.
2269                  */
2270                 gi->last_pos = -1;
2271                 gi->gl = NULL;
2272                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2273         }
2274         return ret;
2275 }
2276
2277 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2278 {
2279         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2280 }
2281
2282 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2283 {
2284         struct seq_file *seq = file->private_data;
2285         struct gfs2_glock_iter *gi = seq->private;
2286
2287         if (gi->gl)
2288                 gfs2_glock_put(gi->gl);
2289         rhashtable_walk_exit(&gi->hti);
2290         return seq_release_private(inode, file);
2291 }
2292
2293 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2294 {
2295         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2296 }
2297
2298 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2299 {
2300         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2301         if (ret == 0) {
2302                 struct seq_file *seq = file->private_data;
2303                 seq->private = inode->i_private;  /* sdp */
2304         }
2305         return ret;
2306 }
2307
2308 static const struct file_operations gfs2_glocks_fops = {
2309         .owner   = THIS_MODULE,
2310         .open    = gfs2_glocks_open,
2311         .read    = seq_read,
2312         .llseek  = seq_lseek,
2313         .release = gfs2_glocks_release,
2314 };
2315
2316 static const struct file_operations gfs2_glstats_fops = {
2317         .owner   = THIS_MODULE,
2318         .open    = gfs2_glstats_open,
2319         .read    = seq_read,
2320         .llseek  = seq_lseek,
2321         .release = gfs2_glocks_release,
2322 };
2323
2324 static const struct file_operations gfs2_sbstats_fops = {
2325         .owner   = THIS_MODULE,
2326         .open    = gfs2_sbstats_open,
2327         .read    = seq_read,
2328         .llseek  = seq_lseek,
2329         .release = seq_release,
2330 };
2331
2332 void gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2333 {
2334         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2335
2336         debugfs_create_file("glocks", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2337                             &gfs2_glocks_fops);
2338
2339         debugfs_create_file("glstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2340                             &gfs2_glstats_fops);
2341
2342         debugfs_create_file("sbstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2343                             &gfs2_sbstats_fops);
2344 }
2345
2346 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2347 {
2348         debugfs_remove_recursive(sdp->debugfs_dir);
2349         sdp->debugfs_dir = NULL;
2350 }
2351
2352 void gfs2_register_debugfs(void)
2353 {
2354         gfs2_root = debugfs_create_dir("gfs2", NULL);
2355 }
2356
2357 void gfs2_unregister_debugfs(void)
2358 {
2359         debugfs_remove(gfs2_root);
2360         gfs2_root = NULL;
2361 }