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                 smp_mb();
 727                 if (atomic_read(&gl->gl_revokes) != 0)
 728                         goto out_sched;
 729                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 730                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 731                 gl->gl_target = gl->gl_demote_state;
 732         } else {
 733                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 734                         gfs2_demote_wake(gl);
 735                 ret = do_promote(gl);
 736                 if (ret == 0)
 737                         goto out_unlock;
 738                 if (ret == 2)
 739                         goto out;
 740                 gh = find_first_waiter(gl);
 741                 gl->gl_target = gh->gh_state;
 742                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 743                         do_error(gl, 0); /* Fail queued try locks */
 744         }
 745         do_xmote(gl, gh, gl->gl_target);
 746 out:
 747         return;
 748
 749 out_sched:
 750         clear_bit(GLF_LOCK, &gl->gl_flags);
 751         smp_mb__after_atomic();
 752         gl->gl_lockref.count++;
 753         __gfs2_glock_queue_work(gl, 0);
 754         return;
 755
 756 out_unlock:
 757         clear_bit(GLF_LOCK, &gl->gl_flags);
 758         smp_mb__after_atomic();
 759         return;
 760 }
 761
 762 static void delete_work_func(struct work_struct *work)
 763 {
 764         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 765         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 766         struct inode *inode;
 767         u64 no_addr = gl->gl_name.ln_number;
 768
 769         /* If someone's using this glock to create a new dinode, the block must
 770            have been freed by another node, then re-used, in which case our
 771            iopen callback is too late after the fact. Ignore it. */
 772         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 773                 goto out;
 774
 775         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 776         if (!IS_ERR_OR_NULL(inode)) {
 777                 d_prune_aliases(inode);
 778                 iput(inode);
 779         }
 780 out:
 781         gfs2_glock_put(gl);
 782 }
 783
 784 static void glock_work_func(struct work_struct *work)
 785 {
 786         unsigned long delay = 0;
 787         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 788         unsigned int drop_refs = 1;
 789
 790         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 791                 finish_xmote(gl, gl->gl_reply);
 792                 drop_refs++;
 793         }
 794         spin_lock(&gl->gl_lockref.lock);
 795         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 796             gl->gl_state != LM_ST_UNLOCKED &&
 797             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 798                 unsigned long holdtime, now = jiffies;
 799
 800                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 801                 if (time_before(now, holdtime))
 802                         delay = holdtime - now;
 803
 804                 if (!delay) {
 805                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 806                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 807                 }
 808         }
 809         run_queue(gl, 0);
 810         if (delay) {
 811                 /* Keep one glock reference for the work we requeue. */
 812                 drop_refs--;
 813                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 814                         delay = 0;
 815                 __gfs2_glock_queue_work(gl, delay);
 816         }
 817
 818         /*
 819          * Drop the remaining glock references manually here. (Mind that
 820          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 821          * here as well.)
 822          */
 823         gl->gl_lockref.count -= drop_refs;
 824         if (!gl->gl_lockref.count) {
 825                 __gfs2_glock_put(gl);
 826                 return;
 827         }
 828         spin_unlock(&gl->gl_lockref.lock);
 829 }
 830
 831 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 832                                             struct gfs2_glock *new)
 833 {
 834         struct wait_glock_queue wait;
 835         wait_queue_head_t *wq = glock_waitqueue(name);
 836         struct gfs2_glock *gl;
 837
 838         wait.name = name;
 839         init_wait(&wait.wait);
 840         wait.wait.func = glock_wake_function;
 841
 842 again:
 843         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 844         rcu_read_lock();
 845         if (new) {
 846                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 847                         &new->gl_node, ht_parms);
 848                 if (IS_ERR(gl))
 849                         goto out;
 850         } else {
 851                 gl = rhashtable_lookup_fast(&gl_hash_table,
 852                         name, ht_parms);
 853         }
 854         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 855                 rcu_read_unlock();
 856                 schedule();
 857                 goto again;
 858         }
 859 out:
 860         rcu_read_unlock();
 861         finish_wait(wq, &wait.wait);
 862         return gl;
 863 }
 864
 865 /**
 866  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 867  * @sdp: The GFS2 superblock
 868  * @number: the lock number
 869  * @glops: The glock_operations to use
 870  * @create: If 0, don't create the glock if it doesn't exist
 871  * @glp: the glock is returned here
 872  *
 873  * This does not lock a glock, just finds/creates structures for one.
 874  *
 875  * Returns: errno
 876  */
 877
 878 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 879                    const struct gfs2_glock_operations *glops, int create,
 880                    struct gfs2_glock **glp)
 881 {
 882         struct super_block *s = sdp->sd_vfs;
 883         struct lm_lockname name = { .ln_number = number,
 884                                     .ln_type = glops->go_type,
 885                                     .ln_sbd = sdp };
 886         struct gfs2_glock *gl, *tmp;
 887         struct address_space *mapping;
 888         struct kmem_cache *cachep;
 889         int ret = 0;
 890
 891         gl = find_insert_glock(&name, NULL);
 892         if (gl) {
 893                 *glp = gl;
 894                 return 0;
 895         }
 896         if (!create)
 897                 return -ENOENT;
 898
 899         if (glops->go_flags & GLOF_ASPACE)
 900                 cachep = gfs2_glock_aspace_cachep;
 901         else
 902                 cachep = gfs2_glock_cachep;
 903         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 904         if (!gl)
 905                 return -ENOMEM;
 906
 907         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 908
 909         if (glops->go_flags & GLOF_LVB) {
 910                 gl->gl_lksb.sb_lvbptr = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
 911                 if (!gl->gl_lksb.sb_lvbptr) {
 912                         kmem_cache_free(cachep, gl);
 913                         return -ENOMEM;
 914                 }
 915         }
 916
 917         atomic_inc(&sdp->sd_glock_disposal);
 918         gl->gl_node.next = NULL;
 919         gl->gl_flags = 0;
 920         gl->gl_name = name;
 921         gl->gl_lockref.count = 1;
 922         gl->gl_state = LM_ST_UNLOCKED;
 923         gl->gl_target = LM_ST_UNLOCKED;
 924         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 925         gl->gl_ops = glops;
 926         gl->gl_dstamp = 0;
 927         preempt_disable();
 928         /* We use the global stats to estimate the initial per-glock stats */
 929         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 930         preempt_enable();
 931         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 932         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 933         gl->gl_tchange = jiffies;
 934         gl->gl_object = NULL;
 935         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 936         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 937         INIT_WORK(&gl->gl_delete, delete_work_func);
 938
 939         mapping = gfs2_glock2aspace(gl);
 940         if (mapping) {
 941                 mapping->a_ops = &gfs2_meta_aops;
 942                 mapping->host = s->s_bdev->bd_inode;
 943                 mapping->flags = 0;
 944                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 945                 mapping->private_data = NULL;
 946                 mapping->writeback_index = 0;
 947         }
 948
 949         tmp = find_insert_glock(&name, gl);
 950         if (!tmp) {
 951                 *glp = gl;
 952                 goto out;
 953         }
 954         if (IS_ERR(tmp)) {
 955                 ret = PTR_ERR(tmp);
 956                 goto out_free;
 957         }
 958         *glp = tmp;
 959
 960 out_free:
 961         kfree(gl->gl_lksb.sb_lvbptr);
 962         kmem_cache_free(cachep, gl);
 963         atomic_dec(&sdp->sd_glock_disposal);
 964
 965 out:
 966         return ret;
 967 }
 968
 969 /**
 970  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 971  * @gl: the glock
 972  * @state: the state we're requesting
 973  * @flags: the modifier flags
 974  * @gh: the holder structure
 975  *
 976  */
 977
 978 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 979                       struct gfs2_holder *gh)
 980 {
 981         INIT_LIST_HEAD(&gh->gh_list);
 982         gh->gh_gl = gl;
 983         gh->gh_ip = _RET_IP_;
 984         gh->gh_owner_pid = get_pid(task_pid(current));
 985         gh->gh_state = state;
 986         gh->gh_flags = flags;
 987         gh->gh_error = 0;
 988         gh->gh_iflags = 0;
 989         gfs2_glock_hold(gl);
 990 }
 991
 992 /**
 993  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 994  * @state: the state we're requesting
 995  * @flags: the modifier flags
 996  * @gh: the holder structure
 997  *
 998  * Don't mess with the glock.
 999  *
1000  */
1001
1002 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
1003 {
1004         gh->gh_state = state;
1005         gh->gh_flags = flags;
1006         gh->gh_iflags = 0;
1007         gh->gh_ip = _RET_IP_;
1008         put_pid(gh->gh_owner_pid);
1009         gh->gh_owner_pid = get_pid(task_pid(current));
1010 }
1011
1012 /**
1013  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
1014  * @gh: the holder structure
1015  *
1016  */
1017
1018 void gfs2_holder_uninit(struct gfs2_holder *gh)
1019 {
1020         put_pid(gh->gh_owner_pid);
1021         gfs2_glock_put(gh->gh_gl);
1022         gfs2_holder_mark_uninitialized(gh);
1023         gh->gh_ip = 0;
1024 }
1025
1026 static void gfs2_glock_update_hold_time(struct gfs2_glock *gl,
1027                                         unsigned long start_time)
1028 {
1029         /* Have we waited longer that a second? */
1030         if (time_after(jiffies, start_time + HZ)) {
1031                 /* Lengthen the minimum hold time. */
1032                 gl->gl_hold_time = min(gl->gl_hold_time + GL_GLOCK_HOLD_INCR,
1033                                        GL_GLOCK_MAX_HOLD);
1034         }
1035 }
1036
1037 /**
1038  * gfs2_glock_wait - wait on a glock acquisition
1039  * @gh: the glock holder
1040  *
1041  * Returns: 0 on success
1042  */
1043
1044 int gfs2_glock_wait(struct gfs2_holder *gh)
1045 {
1046         unsigned long start_time = jiffies;
1047
1048         might_sleep();
1049         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
1050         gfs2_glock_update_hold_time(gh->gh_gl, start_time);
1051         return gh->gh_error;
1052 }
1053
1054 static int glocks_pending(unsigned int num_gh, struct gfs2_holder *ghs)
1055 {
1056         int i;
1057
1058         for (i = 0; i < num_gh; i++)
1059                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags))
1060                         return 1;
1061         return 0;
1062 }
1063
1064 /**
1065  * gfs2_glock_async_wait - wait on multiple asynchronous glock acquisitions
1066  * @num_gh: the number of holders in the array
1067  * @ghs: the glock holder array
1068  *
1069  * Returns: 0 on success, meaning all glocks have been granted and are held.
1070  *          -ESTALE if the request timed out, meaning all glocks were released,
1071  *          and the caller should retry the operation.
1072  */
1073
1074 int gfs2_glock_async_wait(unsigned int num_gh, struct gfs2_holder *ghs)
1075 {
1076         struct gfs2_sbd *sdp = ghs[0].gh_gl->gl_name.ln_sbd;
1077         int i, ret = 0, timeout = 0;
1078         unsigned long start_time = jiffies;
1079         bool keep_waiting;
1080
1081         might_sleep();
1082         /*
1083          * Total up the (minimum hold time * 2) of all glocks and use that to
1084          * determine the max amount of time we should wait.
1085          */
1086         for (i = 0; i < num_gh; i++)
1087                 timeout += ghs[i].gh_gl->gl_hold_time << 1;
1088
1089 wait_for_dlm:
1090         if (!wait_event_timeout(sdp->sd_async_glock_wait,
1091                                 !glocks_pending(num_gh, ghs), timeout))
1092                 ret = -ESTALE; /* request timed out. */
1093
1094         /*
1095          * If dlm granted all our requests, we need to adjust the glock
1096          * minimum hold time values according to how long we waited.
1097          *
1098          * If our request timed out, we need to repeatedly release any held
1099          * glocks we acquired thus far to allow dlm to acquire the remaining
1100          * glocks without deadlocking.  We cannot currently cancel outstanding
1101          * glock acquisitions.
1102          *
1103          * The HIF_WAIT bit tells us which requests still need a response from
1104          * dlm.
1105          *
1106          * If dlm sent us any errors, we return the first error we find.
1107          */
1108         keep_waiting = false;
1109         for (i = 0; i < num_gh; i++) {
1110                 /* Skip holders we have already dequeued below. */
1111                 if (!gfs2_holder_queued(&ghs[i]))
1112                         continue;
1113                 /* Skip holders with a pending DLM response. */
1114                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags)) {
1115                         keep_waiting = true;
1116                         continue;
1117                 }
1118
1119                 if (test_bit(HIF_HOLDER, &ghs[i].gh_iflags)) {
1120                         if (ret == -ESTALE)
1121                                 gfs2_glock_dq(&ghs[i]);
1122                         else
1123                                 gfs2_glock_update_hold_time(ghs[i].gh_gl,
1124                                                             start_time);
1125                 }
1126                 if (!ret)
1127                         ret = ghs[i].gh_error;
1128         }
1129
1130         if (keep_waiting)
1131                 goto wait_for_dlm;
1132
1133         /*
1134          * At this point, we've either acquired all locks or released them all.
1135          */
1136         return ret;
1137 }
1138
1139 /**
1140  * handle_callback - process a demote request
1141  * @gl: the glock
1142  * @state: the state the caller wants us to change to
1143  *
1144  * There are only two requests that we are going to see in actual
1145  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
1146  */
1147
1148 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
1149                             unsigned long delay, bool remote)
1150 {
1151         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
1152
1153         set_bit(bit, &gl->gl_flags);
1154         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
1155                 gl->gl_demote_state = state;
1156                 gl->gl_demote_time = jiffies;
1157         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
1158                         gl->gl_demote_state != state) {
1159                 gl->gl_demote_state = LM_ST_UNLOCKED;
1160         }
1161         if (gl->gl_ops->go_callback)
1162                 gl->gl_ops->go_callback(gl, remote);
1163         trace_gfs2_demote_rq(gl, remote);
1164 }
1165
1166 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
1167 {
1168         struct va_format vaf;
1169         va_list args;
1170
1171         va_start(args, fmt);
1172
1173         if (seq) {
1174                 seq_vprintf(seq, fmt, args);
1175         } else {
1176                 vaf.fmt = fmt;
1177                 vaf.va = &args;
1178
1179                 pr_err("%pV", &vaf);
1180         }
1181
1182         va_end(args);
1183 }
1184
1185 /**
1186  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1187  * @gh: the holder structure to add
1188  *
1189  * Eventually we should move the recursive locking trap to a
1190  * debugging option or something like that. This is the fast
1191  * path and needs to have the minimum number of distractions.
1192  *
1193  */
1194
1195 static inline void add_to_queue(struct gfs2_holder *gh)
1196 __releases(&gl->gl_lockref.lock)
1197 __acquires(&gl->gl_lockref.lock)
1198 {
1199         struct gfs2_glock *gl = gh->gh_gl;
1200         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1201         struct list_head *insert_pt = NULL;
1202         struct gfs2_holder *gh2;
1203         int try_futile = 0;
1204
1205         GLOCK_BUG_ON(gl, gh->gh_owner_pid == NULL);
1206         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1207                 GLOCK_BUG_ON(gl, true);
1208
1209         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1210                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1211                         try_futile = !may_grant(gl, gh);
1212                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1213                         goto fail;
1214         }
1215
1216         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1217                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1218                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1219                         goto trap_recursive;
1220                 if (try_futile &&
1221                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1222 fail:
1223                         gh->gh_error = GLR_TRYFAILED;
1224                         gfs2_holder_wake(gh);
1225                         return;
1226                 }
1227                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1228                         continue;
1229                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1230                         insert_pt = &gh2->gh_list;
1231         }
1232         set_bit(GLF_QUEUED, &gl->gl_flags);
1233         trace_gfs2_glock_queue(gh, 1);
1234         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1235         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1236         if (likely(insert_pt == NULL)) {
1237                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1238                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1239                         goto do_cancel;
1240                 return;
1241         }
1242         list_add_tail(&gh->gh_list, insert_pt);
1243 do_cancel:
1244         gh = list_first_entry(&gl->gl_holders, struct gfs2_holder, gh_list);
1245         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1246                 spin_unlock(&gl->gl_lockref.lock);
1247                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1248                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1249                 spin_lock(&gl->gl_lockref.lock);
1250         }
1251         return;
1252
1253 trap_recursive:
1254         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1255         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1256         fs_err(sdp, "lock type: %d req lock state : %d\n",
1257                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1258         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1259         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1260         fs_err(sdp, "lock type: %d req lock state : %d\n",
1261                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1262         gfs2_dump_glock(NULL, gl, true);
1263         BUG();
1264 }
1265
1266 /**
1267  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1268  * @gh: the holder structure
1269  *
1270  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1271  *
1272  * Returns: 0, GLR_TRYFAILED, or errno on failure
1273  */
1274
1275 int gfs2_glock_nq(struct gfs2_holder *gh)
1276 {
1277         struct gfs2_glock *gl = gh->gh_gl;
1278         int error = 0;
1279
1280         if (glock_blocked_by_withdraw(gl) && !(gh->gh_flags & LM_FLAG_NOEXP))
1281                 return -EIO;
1282
1283         if (test_bit(GLF_LRU, &gl->gl_flags))
1284                 gfs2_glock_remove_from_lru(gl);
1285
1286         spin_lock(&gl->gl_lockref.lock);
1287         add_to_queue(gh);
1288         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1289                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1290                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1291                 gl->gl_lockref.count++;
1292                 __gfs2_glock_queue_work(gl, 0);
1293         }
1294         run_queue(gl, 1);
1295         spin_unlock(&gl->gl_lockref.lock);
1296
1297         if (!(gh->gh_flags & GL_ASYNC))
1298                 error = gfs2_glock_wait(gh);
1299
1300         return error;
1301 }
1302
1303 /**
1304  * gfs2_glock_poll - poll to see if an async request has been completed
1305  * @gh: the holder
1306  *
1307  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1308  */
1309
1310 int gfs2_glock_poll(struct gfs2_holder *gh)
1311 {
1312         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1313 }
1314
1315 /**
1316  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1317  * @gh: the glock holder
1318  *
1319  */
1320
1321 void gfs2_glock_dq(struct gfs2_holder *gh)
1322 {
1323         struct gfs2_glock *gl = gh->gh_gl;
1324         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1325         unsigned delay = 0;
1326         int fast_path = 0;
1327
1328         spin_lock(&gl->gl_lockref.lock);
1329         /*
1330          * If we're in the process of file system withdraw, we cannot just
1331          * dequeue any glocks until our journal is recovered, lest we
1332          * introduce file system corruption. We need two exceptions to this
1333          * rule: We need to allow unlocking of nondisk glocks and the glock
1334          * for our own journal that needs recovery.
1335          */
1336         if (test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags) &&
1337             glock_blocked_by_withdraw(gl) &&
1338             gh->gh_gl != sdp->sd_jinode_gl) {
1339                 sdp->sd_glock_dqs_held++;
1340                 might_sleep();
1341                 wait_on_bit(&sdp->sd_flags, SDF_WITHDRAW_RECOVERY,
1342                             TASK_UNINTERRUPTIBLE);
1343         }
1344         if (gh->gh_flags & GL_NOCACHE)
1345                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1346
1347         list_del_init(&gh->gh_list);
1348         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1349         if (find_first_holder(gl) == NULL) {
1350                 if (list_empty(&gl->gl_holders) &&
1351                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1352                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1353                         fast_path = 1;
1354         }
1355         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1356                 gfs2_glock_add_to_lru(gl);
1357
1358         trace_gfs2_glock_queue(gh, 0);
1359         if (unlikely(!fast_path)) {
1360                 gl->gl_lockref.count++;
1361                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1362                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1363                     gl->gl_name.ln_type == LM_TYPE_INODE)
1364                         delay = gl->gl_hold_time;
1365                 __gfs2_glock_queue_work(gl, delay);
1366         }
1367         spin_unlock(&gl->gl_lockref.lock);
1368 }
1369
1370 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1371 {
1372         struct gfs2_glock *gl = gh->gh_gl;
1373         gfs2_glock_dq(gh);
1374         might_sleep();
1375         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1376 }
1377
1378 /**
1379  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1380  * @gh: the holder structure
1381  *
1382  */
1383
1384 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1385 {
1386         gfs2_glock_dq(gh);
1387         gfs2_holder_uninit(gh);
1388 }
1389
1390 /**
1391  * gfs2_glock_nq_num - acquire a glock based on lock number
1392  * @sdp: the filesystem
1393  * @number: the lock number
1394  * @glops: the glock operations for the type of glock
1395  * @state: the state to acquire the glock in
1396  * @flags: modifier flags for the acquisition
1397  * @gh: the struct gfs2_holder
1398  *
1399  * Returns: errno
1400  */
1401
1402 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1403                       const struct gfs2_glock_operations *glops,
1404                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1405 {
1406         struct gfs2_glock *gl;
1407         int error;
1408
1409         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1410         if (!error) {
1411                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1412                 gfs2_glock_put(gl);
1413         }
1414
1415         return error;
1416 }
1417
1418 /**
1419  * glock_compare - Compare two struct gfs2_glock structures for sorting
1420  * @arg_a: the first structure
1421  * @arg_b: the second structure
1422  *
1423  */
1424
1425 static int glock_compare(const void *arg_a, const void *arg_b)
1426 {
1427         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1428         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1429         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1430         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1431
1432         if (a->ln_number > b->ln_number)
1433                 return 1;
1434         if (a->ln_number < b->ln_number)
1435                 return -1;
1436         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1437         return 0;
1438 }
1439
1440 /**
1441  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1442  * @num_gh: the number of structures
1443  * @ghs: an array of struct gfs2_holder structures
1444  *
1445  * Returns: 0 on success (all glocks acquired),
1446  *          errno on failure (no glocks acquired)
1447  */
1448
1449 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1450                      struct gfs2_holder **p)
1451 {
1452         unsigned int x;
1453         int error = 0;
1454
1455         for (x = 0; x < num_gh; x++)
1456                 p[x] = &ghs[x];
1457
1458         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1459
1460         for (x = 0; x < num_gh; x++) {
1461                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1462
1463                 error = gfs2_glock_nq(p[x]);
1464                 if (error) {
1465                         while (x--)
1466                                 gfs2_glock_dq(p[x]);
1467                         break;
1468                 }
1469         }
1470
1471         return error;
1472 }
1473
1474 /**
1475  * gfs2_glock_nq_m - acquire multiple glocks
1476  * @num_gh: the number of structures
1477  * @ghs: an array of struct gfs2_holder structures
1478  *
1479  *
1480  * Returns: 0 on success (all glocks acquired),
1481  *          errno on failure (no glocks acquired)
1482  */
1483
1484 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1485 {
1486         struct gfs2_holder *tmp[4];
1487         struct gfs2_holder **pph = tmp;
1488         int error = 0;
1489
1490         switch(num_gh) {
1491         case 0:
1492                 return 0;
1493         case 1:
1494                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1495                 return gfs2_glock_nq(ghs);
1496         default:
1497                 if (num_gh <= 4)
1498                         break;
1499                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1500                                     GFP_NOFS);
1501                 if (!pph)
1502                         return -ENOMEM;
1503         }
1504
1505         error = nq_m_sync(num_gh, ghs, pph);
1506
1507         if (pph != tmp)
1508                 kfree(pph);
1509
1510         return error;
1511 }
1512
1513 /**
1514  * gfs2_glock_dq_m - release multiple glocks
1515  * @num_gh: the number of structures
1516  * @ghs: an array of struct gfs2_holder structures
1517  *
1518  */
1519
1520 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1521 {
1522         while (num_gh--)
1523                 gfs2_glock_dq(&ghs[num_gh]);
1524 }
1525
1526 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1527 {
1528         unsigned long delay = 0;
1529         unsigned long holdtime;
1530         unsigned long now = jiffies;
1531
1532         gfs2_glock_hold(gl);
1533         holdtime = gl->gl_tchange + gl->gl_hold_time;
1534         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1535             gl->gl_name.ln_type == LM_TYPE_INODE) {
1536                 if (time_before(now, holdtime))
1537                         delay = holdtime - now;
1538                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1539                         delay = gl->gl_hold_time;
1540         }
1541
1542         spin_lock(&gl->gl_lockref.lock);
1543         handle_callback(gl, state, delay, true);
1544         __gfs2_glock_queue_work(gl, delay);
1545         spin_unlock(&gl->gl_lockref.lock);
1546 }
1547
1548 /**
1549  * gfs2_should_freeze - Figure out if glock should be frozen
1550  * @gl: The glock in question
1551  *
1552  * Glocks are not frozen if (a) the result of the dlm operation is
1553  * an error, (b) the locking operation was an unlock operation or
1554  * (c) if there is a "noexp" flagged request anywhere in the queue
1555  *
1556  * Returns: 1 if freezing should occur, 0 otherwise
1557  */
1558
1559 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1560 {
1561         const struct gfs2_holder *gh;
1562
1563         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1564                 return 0;
1565         if (gl->gl_target == LM_ST_UNLOCKED)
1566                 return 0;
1567
1568         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1569                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1570                         continue;
1571                 if (LM_FLAG_NOEXP & gh->gh_flags)
1572                         return 0;
1573         }
1574
1575         return 1;
1576 }
1577
1578 /**
1579  * gfs2_glock_complete - Callback used by locking
1580  * @gl: Pointer to the glock
1581  * @ret: The return value from the dlm
1582  *
1583  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1584  * to use a bitfield shared with other glock state fields.
1585  */
1586
1587 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1588 {
1589         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1590
1591         spin_lock(&gl->gl_lockref.lock);
1592         gl->gl_reply = ret;
1593
1594         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1595                 if (gfs2_should_freeze(gl)) {
1596                         set_bit(GLF_FROZEN, &gl->gl_flags);
1597                         spin_unlock(&gl->gl_lockref.lock);
1598                         return;
1599                 }
1600         }
1601
1602         gl->gl_lockref.count++;
1603         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1604         __gfs2_glock_queue_work(gl, 0);
1605         spin_unlock(&gl->gl_lockref.lock);
1606 }
1607
1608 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1609 {
1610         struct gfs2_glock *gla, *glb;
1611
1612         gla = list_entry(a, struct gfs2_glock, gl_lru);
1613         glb = list_entry(b, struct gfs2_glock, gl_lru);
1614
1615         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1616                 return 1;
1617         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1618                 return -1;
1619
1620         return 0;
1621 }
1622
1623 /**
1624  * gfs2_dispose_glock_lru - Demote a list of glocks
1625  * @list: The list to dispose of
1626  *
1627  * Disposing of glocks may involve disk accesses, so that here we sort
1628  * the glocks by number (i.e. disk location of the inodes) so that if
1629  * there are any such accesses, they'll be sent in order (mostly).
1630  *
1631  * Must be called under the lru_lock, but may drop and retake this
1632  * lock. While the lru_lock is dropped, entries may vanish from the
1633  * list, but no new entries will appear on the list (since it is
1634  * private)
1635  */
1636
1637 static void gfs2_dispose_glock_lru(struct list_head *list)
1638 __releases(&lru_lock)
1639 __acquires(&lru_lock)
1640 {
1641         struct gfs2_glock *gl;
1642
1643         list_sort(NULL, list, glock_cmp);
1644
1645         while(!list_empty(list)) {
1646                 gl = list_first_entry(list, struct gfs2_glock, gl_lru);
1647                 list_del_init(&gl->gl_lru);
1648                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1649 add_back_to_lru:
1650                         list_add(&gl->gl_lru, &lru_list);
1651                         set_bit(GLF_LRU, &gl->gl_flags);
1652                         atomic_inc(&lru_count);
1653                         continue;
1654                 }
1655                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1656                         spin_unlock(&gl->gl_lockref.lock);
1657                         goto add_back_to_lru;
1658                 }
1659                 gl->gl_lockref.count++;
1660                 if (demote_ok(gl))
1661                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1662                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1663                 __gfs2_glock_queue_work(gl, 0);
1664                 spin_unlock(&gl->gl_lockref.lock);
1665                 cond_resched_lock(&lru_lock);
1666         }
1667 }
1668
1669 /**
1670  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1671  * @nr: The number of entries to scan
1672  *
1673  * This function selects the entries on the LRU which are able to
1674  * be demoted, and then kicks off the process by calling
1675  * gfs2_dispose_glock_lru() above.
1676  */
1677
1678 static long gfs2_scan_glock_lru(int nr)
1679 {
1680         struct gfs2_glock *gl;
1681         LIST_HEAD(skipped);
1682         LIST_HEAD(dispose);
1683         long freed = 0;
1684
1685         spin_lock(&lru_lock);
1686         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1687                 gl = list_first_entry(&lru_list, struct gfs2_glock, gl_lru);
1688
1689                 /* Test for being demotable */
1690                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1691                         list_move(&gl->gl_lru, &dispose);
1692                         atomic_dec(&lru_count);
1693                         clear_bit(GLF_LRU, &gl->gl_flags);
1694                         freed++;
1695                         continue;
1696                 }
1697
1698                 list_move(&gl->gl_lru, &skipped);
1699         }
1700         list_splice(&skipped, &lru_list);
1701         if (!list_empty(&dispose))
1702                 gfs2_dispose_glock_lru(&dispose);
1703         spin_unlock(&lru_lock);
1704
1705         return freed;
1706 }
1707
1708 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1709                                             struct shrink_control *sc)
1710 {
1711         if (!(sc->gfp_mask & __GFP_FS))
1712                 return SHRINK_STOP;
1713         return gfs2_scan_glock_lru(sc->nr_to_scan);
1714 }
1715
1716 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1717                                              struct shrink_control *sc)
1718 {
1719         return vfs_pressure_ratio(atomic_read(&lru_count));
1720 }
1721
1722 static struct shrinker glock_shrinker = {
1723         .seeks = DEFAULT_SEEKS,
1724         .count_objects = gfs2_glock_shrink_count,
1725         .scan_objects = gfs2_glock_shrink_scan,
1726 };
1727
1728 /**
1729  * examine_bucket - Call a function for glock in a hash bucket
1730  * @examiner: the function
1731  * @sdp: the filesystem
1732  * @bucket: the bucket
1733  *
1734  * Note that the function can be called multiple times on the same
1735  * object.  So the user must ensure that the function can cope with
1736  * that.
1737  */
1738
1739 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1740 {
1741         struct gfs2_glock *gl;
1742         struct rhashtable_iter iter;
1743
1744         rhashtable_walk_enter(&gl_hash_table, &iter);
1745
1746         do {
1747                 rhashtable_walk_start(&iter);
1748
1749                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1750                         if (gl->gl_name.ln_sbd == sdp &&
1751                             lockref_get_not_dead(&gl->gl_lockref))
1752                                 examiner(gl);
1753
1754                 rhashtable_walk_stop(&iter);
1755         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1756
1757         rhashtable_walk_exit(&iter);
1758 }
1759
1760 /**
1761  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1762  * @gl: The glock to thaw
1763  *
1764  */
1765
1766 static void thaw_glock(struct gfs2_glock *gl)
1767 {
1768         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1769                 gfs2_glock_put(gl);
1770                 return;
1771         }
1772         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1773         gfs2_glock_queue_work(gl, 0);
1774 }
1775
1776 /**
1777  * clear_glock - look at a glock and see if we can free it from glock cache
1778  * @gl: the glock to look at
1779  *
1780  */
1781
1782 static void clear_glock(struct gfs2_glock *gl)
1783 {
1784         gfs2_glock_remove_from_lru(gl);
1785
1786         spin_lock(&gl->gl_lockref.lock);
1787         if (gl->gl_state != LM_ST_UNLOCKED)
1788                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1789         __gfs2_glock_queue_work(gl, 0);
1790         spin_unlock(&gl->gl_lockref.lock);
1791 }
1792
1793 /**
1794  * gfs2_glock_thaw - Thaw any frozen glocks
1795  * @sdp: The super block
1796  *
1797  */
1798
1799 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1800 {
1801         glock_hash_walk(thaw_glock, sdp);
1802 }
1803
1804 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1805 {
1806         spin_lock(&gl->gl_lockref.lock);
1807         gfs2_dump_glock(seq, gl, fsid);
1808         spin_unlock(&gl->gl_lockref.lock);
1809 }
1810
1811 static void dump_glock_func(struct gfs2_glock *gl)
1812 {
1813         dump_glock(NULL, gl, true);
1814 }
1815
1816 /**
1817  * gfs2_gl_hash_clear - Empty out the glock hash table
1818  * @sdp: the filesystem
1819  * @wait: wait until it's all gone
1820  *
1821  * Called when unmounting the filesystem.
1822  */
1823
1824 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1825 {
1826         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1827         flush_workqueue(glock_workqueue);
1828         glock_hash_walk(clear_glock, sdp);
1829         flush_workqueue(glock_workqueue);
1830         wait_event_timeout(sdp->sd_glock_wait,
1831                            atomic_read(&sdp->sd_glock_disposal) == 0,
1832                            HZ * 600);
1833         glock_hash_walk(dump_glock_func, sdp);
1834 }
1835
1836 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1837 {
1838         struct gfs2_glock *gl = ip->i_gl;
1839         int ret;
1840
1841         ret = gfs2_truncatei_resume(ip);
1842         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1843
1844         spin_lock(&gl->gl_lockref.lock);
1845         clear_bit(GLF_LOCK, &gl->gl_flags);
1846         run_queue(gl, 1);
1847         spin_unlock(&gl->gl_lockref.lock);
1848 }
1849
1850 static const char *state2str(unsigned state)
1851 {
1852         switch(state) {
1853         case LM_ST_UNLOCKED:
1854                 return "UN";
1855         case LM_ST_SHARED:
1856                 return "SH";
1857         case LM_ST_DEFERRED:
1858                 return "DF";
1859         case LM_ST_EXCLUSIVE:
1860                 return "EX";
1861         }
1862         return "??";
1863 }
1864
1865 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1866 {
1867         char *p = buf;
1868         if (flags & LM_FLAG_TRY)
1869                 *p++ = 't';
1870         if (flags & LM_FLAG_TRY_1CB)
1871                 *p++ = 'T';
1872         if (flags & LM_FLAG_NOEXP)
1873                 *p++ = 'e';
1874         if (flags & LM_FLAG_ANY)
1875                 *p++ = 'A';
1876         if (flags & LM_FLAG_PRIORITY)
1877                 *p++ = 'p';
1878         if (flags & GL_ASYNC)
1879                 *p++ = 'a';
1880         if (flags & GL_EXACT)
1881                 *p++ = 'E';
1882         if (flags & GL_NOCACHE)
1883                 *p++ = 'c';
1884         if (test_bit(HIF_HOLDER, &iflags))
1885                 *p++ = 'H';
1886         if (test_bit(HIF_WAIT, &iflags))
1887                 *p++ = 'W';
1888         if (test_bit(HIF_FIRST, &iflags))
1889                 *p++ = 'F';
1890         *p = 0;
1891         return buf;
1892 }
1893
1894 /**
1895  * dump_holder - print information about a glock holder
1896  * @seq: the seq_file struct
1897  * @gh: the glock holder
1898  * @fs_id_buf: pointer to file system id (if requested)
1899  *
1900  */
1901
1902 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh,
1903                         const char *fs_id_buf)
1904 {
1905         struct task_struct *gh_owner = NULL;
1906         char flags_buf[32];
1907
1908         rcu_read_lock();
1909         if (gh->gh_owner_pid)
1910                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1911         gfs2_print_dbg(seq, "%s H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1912                        fs_id_buf, state2str(gh->gh_state),
1913                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1914                        gh->gh_error,
1915                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1916                        gh_owner ? gh_owner->comm : "(ended)",
1917                        (void *)gh->gh_ip);
1918         rcu_read_unlock();
1919 }
1920
1921 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1922 {
1923         const unsigned long *gflags = &gl->gl_flags;
1924         char *p = buf;
1925
1926         if (test_bit(GLF_LOCK, gflags))
1927                 *p++ = 'l';
1928         if (test_bit(GLF_DEMOTE, gflags))
1929                 *p++ = 'D';
1930         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1931                 *p++ = 'd';
1932         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1933                 *p++ = 'p';
1934         if (test_bit(GLF_DIRTY, gflags))
1935                 *p++ = 'y';
1936         if (test_bit(GLF_LFLUSH, gflags))
1937                 *p++ = 'f';
1938         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1939                 *p++ = 'i';
1940         if (test_bit(GLF_REPLY_PENDING, gflags))
1941                 *p++ = 'r';
1942         if (test_bit(GLF_INITIAL, gflags))
1943                 *p++ = 'I';
1944         if (test_bit(GLF_FROZEN, gflags))
1945                 *p++ = 'F';
1946         if (test_bit(GLF_QUEUED, gflags))
1947                 *p++ = 'q';
1948         if (test_bit(GLF_LRU, gflags))
1949                 *p++ = 'L';
1950         if (gl->gl_object)
1951                 *p++ = 'o';
1952         if (test_bit(GLF_BLOCKING, gflags))
1953                 *p++ = 'b';
1954         *p = 0;
1955         return buf;
1956 }
1957
1958 /**
1959  * gfs2_dump_glock - print information about a glock
1960  * @seq: The seq_file struct
1961  * @gl: the glock
1962  * @fsid: If true, also dump the file system id
1963  *
1964  * The file format is as follows:
1965  * One line per object, capital letters are used to indicate objects
1966  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1967  * other objects are indented by a single space and follow the glock to
1968  * which they are related. Fields are indicated by lower case letters
1969  * followed by a colon and the field value, except for strings which are in
1970  * [] so that its possible to see if they are composed of spaces for
1971  * example. The field's are n = number (id of the object), f = flags,
1972  * t = type, s = state, r = refcount, e = error, p = pid.
1973  *
1974  */
1975
1976 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1977 {
1978         const struct gfs2_glock_operations *glops = gl->gl_ops;
1979         unsigned long long dtime;
1980         const struct gfs2_holder *gh;
1981         char gflags_buf[32];
1982         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1983         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
1984
1985         memset(fs_id_buf, 0, sizeof(fs_id_buf));
1986         if (fsid && sdp) /* safety precaution */
1987                 sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
1988         dtime = jiffies - gl->gl_demote_time;
1989         dtime *= 1000000/HZ; /* demote time in uSec */
1990         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1991                 dtime = 0;
1992         gfs2_print_dbg(seq, "%sG:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d "
1993                        "v:%d r:%d m:%ld\n", fs_id_buf, state2str(gl->gl_state),
1994                   gl->gl_name.ln_type,
1995                   (unsigned long long)gl->gl_name.ln_number,
1996                   gflags2str(gflags_buf, gl),
1997                   state2str(gl->gl_target),
1998                   state2str(gl->gl_demote_state), dtime,
1999                   atomic_read(&gl->gl_ail_count),
2000                   atomic_read(&gl->gl_revokes),
2001                   (int)gl->gl_lockref.count, gl->gl_hold_time);
2002
2003         list_for_each_entry(gh, &gl->gl_holders, gh_list)
2004                 dump_holder(seq, gh, fs_id_buf);
2005
2006         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
2007                 glops->go_dump(seq, gl, fs_id_buf);
2008 }
2009
2010 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
2011 {
2012         struct gfs2_glock *gl = iter_ptr;
2013
2014         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
2015                    gl->gl_name.ln_type,
2016                    (unsigned long long)gl->gl_name.ln_number,
2017                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
2018                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
2019                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
2020                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
2021                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
2022                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
2023                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
2024                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
2025         return 0;
2026 }
2027
2028 static const char *gfs2_gltype[] = {
2029         "type",
2030         "reserved",
2031         "nondisk",
2032         "inode",
2033         "rgrp",
2034         "meta",
2035         "iopen",
2036         "flock",
2037         "plock",
2038         "quota",
2039         "journal",
2040 };
2041
2042 static const char *gfs2_stype[] = {
2043         [GFS2_LKS_SRTT]         = "srtt",
2044         [GFS2_LKS_SRTTVAR]      = "srttvar",
2045         [GFS2_LKS_SRTTB]        = "srttb",
2046         [GFS2_LKS_SRTTVARB]     = "srttvarb",
2047         [GFS2_LKS_SIRT]         = "sirt",
2048         [GFS2_LKS_SIRTVAR]      = "sirtvar",
2049         [GFS2_LKS_DCOUNT]       = "dlm",
2050         [GFS2_LKS_QCOUNT]       = "queue",
2051 };
2052
2053 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
2054
2055 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
2056 {
2057         struct gfs2_sbd *sdp = seq->private;
2058         loff_t pos = *(loff_t *)iter_ptr;
2059         unsigned index = pos >> 3;
2060         unsigned subindex = pos & 0x07;
2061         int i;
2062
2063         if (index == 0 && subindex != 0)
2064                 return 0;
2065
2066         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
2067                    (index == 0) ? "cpu": gfs2_stype[subindex]);
2068
2069         for_each_possible_cpu(i) {
2070                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
2071
2072                 if (index == 0)
2073                         seq_printf(seq, " %15u", i);
2074                 else
2075                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
2076                                    lkstats[index - 1].stats[subindex]);
2077         }
2078         seq_putc(seq, '\n');
2079         return 0;
2080 }
2081
2082 int __init gfs2_glock_init(void)
2083 {
2084         int i, ret;
2085
2086         ret = rhashtable_init(&gl_hash_table, &ht_parms);
2087         if (ret < 0)
2088                 return ret;
2089
2090         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
2091                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
2092         if (!glock_workqueue) {
2093                 rhashtable_destroy(&gl_hash_table);
2094                 return -ENOMEM;
2095         }
2096         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
2097                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
2098                                                 0);
2099         if (!gfs2_delete_workqueue) {
2100                 destroy_workqueue(glock_workqueue);
2101                 rhashtable_destroy(&gl_hash_table);
2102                 return -ENOMEM;
2103         }
2104
2105         ret = register_shrinker(&glock_shrinker);
2106         if (ret) {
2107                 destroy_workqueue(gfs2_delete_workqueue);
2108                 destroy_workqueue(glock_workqueue);
2109                 rhashtable_destroy(&gl_hash_table);
2110                 return ret;
2111         }
2112
2113         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
2114                 init_waitqueue_head(glock_wait_table + i);
2115
2116         return 0;
2117 }
2118
2119 void gfs2_glock_exit(void)
2120 {
2121         unregister_shrinker(&glock_shrinker);
2122         rhashtable_destroy(&gl_hash_table);
2123         destroy_workqueue(glock_workqueue);
2124         destroy_workqueue(gfs2_delete_workqueue);
2125 }
2126
2127 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
2128 {
2129         struct gfs2_glock *gl = gi->gl;
2130
2131         if (gl) {
2132                 if (n == 0)
2133                         return;
2134                 if (!lockref_put_not_zero(&gl->gl_lockref))
2135                         gfs2_glock_queue_put(gl);
2136         }
2137         for (;;) {
2138                 gl = rhashtable_walk_next(&gi->hti);
2139                 if (IS_ERR_OR_NULL(gl)) {
2140                         if (gl == ERR_PTR(-EAGAIN)) {
2141                                 n = 1;
2142                                 continue;
2143                         }
2144                         gl = NULL;
2145                         break;
2146                 }
2147                 if (gl->gl_name.ln_sbd != gi->sdp)
2148                         continue;
2149                 if (n <= 1) {
2150                         if (!lockref_get_not_dead(&gl->gl_lockref))
2151                                 continue;
2152                         break;
2153                 } else {
2154                         if (__lockref_is_dead(&gl->gl_lockref))
2155                                 continue;
2156                         n--;
2157                 }
2158         }
2159         gi->gl = gl;
2160 }
2161
2162 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
2163         __acquires(RCU)
2164 {
2165         struct gfs2_glock_iter *gi = seq->private;
2166         loff_t n;
2167
2168         /*
2169          * We can either stay where we are, skip to the next hash table
2170          * entry, or start from the beginning.
2171          */
2172         if (*pos < gi->last_pos) {
2173                 rhashtable_walk_exit(&gi->hti);
2174                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2175                 n = *pos + 1;
2176         } else {
2177                 n = *pos - gi->last_pos;
2178         }
2179
2180         rhashtable_walk_start(&gi->hti);
2181
2182         gfs2_glock_iter_next(gi, n);
2183         gi->last_pos = *pos;
2184         return gi->gl;
2185 }
2186
2187 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
2188                                  loff_t *pos)
2189 {
2190         struct gfs2_glock_iter *gi = seq->private;
2191
2192         (*pos)++;
2193         gi->last_pos = *pos;
2194         gfs2_glock_iter_next(gi, 1);
2195         return gi->gl;
2196 }
2197
2198 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2199         __releases(RCU)
2200 {
2201         struct gfs2_glock_iter *gi = seq->private;
2202
2203         rhashtable_walk_stop(&gi->hti);
2204 }
2205
2206 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2207 {
2208         dump_glock(seq, iter_ptr, false);
2209         return 0;
2210 }
2211
2212 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2213 {
2214         preempt_disable();
2215         if (*pos >= GFS2_NR_SBSTATS)
2216                 return NULL;
2217         return pos;
2218 }
2219
2220 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2221                                    loff_t *pos)
2222 {
2223         (*pos)++;
2224         if (*pos >= GFS2_NR_SBSTATS)
2225                 return NULL;
2226         return pos;
2227 }
2228
2229 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2230 {
2231         preempt_enable();
2232 }
2233
2234 static const struct seq_operations gfs2_glock_seq_ops = {
2235         .start = gfs2_glock_seq_start,
2236         .next  = gfs2_glock_seq_next,
2237         .stop  = gfs2_glock_seq_stop,
2238         .show  = gfs2_glock_seq_show,
2239 };
2240
2241 static const struct seq_operations gfs2_glstats_seq_ops = {
2242         .start = gfs2_glock_seq_start,
2243         .next  = gfs2_glock_seq_next,
2244         .stop  = gfs2_glock_seq_stop,
2245         .show  = gfs2_glstats_seq_show,
2246 };
2247
2248 static const struct seq_operations gfs2_sbstats_seq_ops = {
2249         .start = gfs2_sbstats_seq_start,
2250         .next  = gfs2_sbstats_seq_next,
2251         .stop  = gfs2_sbstats_seq_stop,
2252         .show  = gfs2_sbstats_seq_show,
2253 };
2254
2255 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2256
2257 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2258                               const struct seq_operations *ops)
2259 {
2260         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2261         if (ret == 0) {
2262                 struct seq_file *seq = file->private_data;
2263                 struct gfs2_glock_iter *gi = seq->private;
2264
2265                 gi->sdp = inode->i_private;
2266                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2267                 if (seq->buf)
2268                         seq->size = GFS2_SEQ_GOODSIZE;
2269                 /*
2270                  * Initially, we are "before" the first hash table entry; the
2271                  * first call to rhashtable_walk_next gets us the first entry.
2272                  */
2273                 gi->last_pos = -1;
2274                 gi->gl = NULL;
2275                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2276         }
2277         return ret;
2278 }
2279
2280 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2281 {
2282         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2283 }
2284
2285 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2286 {
2287         struct seq_file *seq = file->private_data;
2288         struct gfs2_glock_iter *gi = seq->private;
2289
2290         if (gi->gl)
2291                 gfs2_glock_put(gi->gl);
2292         rhashtable_walk_exit(&gi->hti);
2293         return seq_release_private(inode, file);
2294 }
2295
2296 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2297 {
2298         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2299 }
2300
2301 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2302 {
2303         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2304         if (ret == 0) {
2305                 struct seq_file *seq = file->private_data;
2306                 seq->private = inode->i_private;  /* sdp */
2307         }
2308         return ret;
2309 }
2310
2311 static const struct file_operations gfs2_glocks_fops = {
2312         .owner   = THIS_MODULE,
2313         .open    = gfs2_glocks_open,
2314         .read    = seq_read,
2315         .llseek  = seq_lseek,
2316         .release = gfs2_glocks_release,
2317 };
2318
2319 static const struct file_operations gfs2_glstats_fops = {
2320         .owner   = THIS_MODULE,
2321         .open    = gfs2_glstats_open,
2322         .read    = seq_read,
2323         .llseek  = seq_lseek,
2324         .release = gfs2_glocks_release,
2325 };
2326
2327 static const struct file_operations gfs2_sbstats_fops = {
2328         .owner   = THIS_MODULE,
2329         .open    = gfs2_sbstats_open,
2330         .read    = seq_read,
2331         .llseek  = seq_lseek,
2332         .release = seq_release,
2333 };
2334
2335 void gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2336 {
2337         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2338
2339         debugfs_create_file("glocks", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2340                             &gfs2_glocks_fops);
2341
2342         debugfs_create_file("glstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2343                             &gfs2_glstats_fops);
2344
2345         debugfs_create_file("sbstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2346                             &gfs2_sbstats_fops);
2347 }
2348
2349 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2350 {
2351         debugfs_remove_recursive(sdp->debugfs_dir);
2352         sdp->debugfs_dir = NULL;
2353 }
2354
2355 void gfs2_register_debugfs(void)
2356 {
2357         gfs2_root = debugfs_create_dir("gfs2", NULL);
2358 }
2359
2360 void gfs2_unregister_debugfs(void)
2361 {
2362         debugfs_remove(gfs2_root);
2363         gfs2_root = NULL;
2364 }