fs/gfs2/glock.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/buffer_head.h>
  14 #include <linux/delay.h>
  15 #include <linux/sort.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 <asm/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
  33 #include "gfs2.h"
  34 #include "incore.h"
  35 #include "glock.h"
  36 #include "glops.h"
  37 #include "inode.h"
  38 #include "lops.h"
  39 #include "meta_io.h"
  40 #include "quota.h"
  41 #include "super.h"
  42 #include "util.h"
  43 #include "bmap.h"
  44 #define CREATE_TRACE_POINTS
  45 #include "trace_gfs2.h"
  46
  47 struct gfs2_glock_iter {
  48         int hash;                       /* hash bucket index         */
  49         struct gfs2_sbd *sdp;           /* incore superblock         */
  50         struct gfs2_glock *gl;          /* current glock struct      */
  51         char string[512];               /* scratch space             */
  52 };
  53
  54 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  55
  56 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl);
  57 #define GLOCK_BUG_ON(gl,x) do { if (unlikely(x)) { __dump_glock(NULL, gl); BUG(); } } while(0)
  58 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  59
  60 static struct dentry *gfs2_root;
  61 static struct workqueue_struct *glock_workqueue;
  62 struct workqueue_struct *gfs2_delete_workqueue;
  63 static LIST_HEAD(lru_list);
  64 static atomic_t lru_count = ATOMIC_INIT(0);
  65 static DEFINE_SPINLOCK(lru_lock);
  66
  67 #define GFS2_GL_HASH_SHIFT      15
  68 #define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
  69 #define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
  70
  71 static struct hlist_bl_head gl_hash_table[GFS2_GL_HASH_SIZE];
  72 static struct dentry *gfs2_root;
  73
  74 /**
  75  * gl_hash() - Turn glock number into hash bucket number
  76  * @lock: The glock number
  77  *
  78  * Returns: The number of the corresponding hash bucket
  79  */
  80
  81 static unsigned int gl_hash(const struct gfs2_sbd *sdp,
  82                             const struct lm_lockname *name)
  83 {
  84         unsigned int h;
  85
  86         h = jhash(&name->ln_number, sizeof(u64), 0);
  87         h = jhash(&name->ln_type, sizeof(unsigned int), h);
  88         h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
  89         h &= GFS2_GL_HASH_MASK;
  90
  91         return h;
  92 }
  93
  94 static inline void spin_lock_bucket(unsigned int hash)
  95 {
  96         struct hlist_bl_head *bl = &gl_hash_table[hash];
  97         bit_spin_lock(0, (unsigned long *)bl);
  98 }
  99
 100 static inline void spin_unlock_bucket(unsigned int hash)
 101 {
 102         struct hlist_bl_head *bl = &gl_hash_table[hash];
 103         __bit_spin_unlock(0, (unsigned long *)bl);
 104 }
 105
 106 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 107 {
 108         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 109
 110         if (gl->gl_ops->go_flags & GLOF_ASPACE)
 111                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 112         else
 113                 kmem_cache_free(gfs2_glock_cachep, gl);
 114 }
 115
 116 void gfs2_glock_free(struct gfs2_glock *gl)
 117 {
 118         struct gfs2_sbd *sdp = gl->gl_sbd;
 119
 120         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 121         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 122                 wake_up(&sdp->sd_glock_wait);
 123 }
 124
 125 /**
 126  * gfs2_glock_hold() - increment reference count on glock
 127  * @gl: The glock to hold
 128  *
 129  */
 130
 131 void gfs2_glock_hold(struct gfs2_glock *gl)
 132 {
 133         GLOCK_BUG_ON(gl, atomic_read(&gl->gl_ref) == 0);
 134         atomic_inc(&gl->gl_ref);
 135 }
 136
 137 /**
 138  * demote_ok - Check to see if it's ok to unlock a glock
 139  * @gl: the glock
 140  *
 141  * Returns: 1 if it's ok
 142  */
 143
 144 static int demote_ok(const struct gfs2_glock *gl)
 145 {
 146         const struct gfs2_glock_operations *glops = gl->gl_ops;
 147
 148         /* assert_spin_locked(&gl->gl_spin); */
 149
 150         if (gl->gl_state == LM_ST_UNLOCKED)
 151                 return 0;
 152         if (test_bit(GLF_LFLUSH, &gl->gl_flags))
 153                 return 0;
 154         if ((gl->gl_name.ln_type != LM_TYPE_INODE) &&
 155             !list_empty(&gl->gl_holders))
 156                 return 0;
 157         if (glops->go_demote_ok)
 158                 return glops->go_demote_ok(gl);
 159         return 1;
 160 }
 161
 162
 163 /**
 164  * __gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
 165  * @gl: the glock
 166  *
 167  * If the glock is demotable, then we add it (or move it) to the end
 168  * of the glock LRU list.
 169  */
 170
 171 static void __gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
 172 {
 173         if (demote_ok(gl)) {
 174                 spin_lock(&lru_lock);
 175
 176                 if (!list_empty(&gl->gl_lru))
 177                         list_del_init(&gl->gl_lru);
 178                 else
 179                         atomic_inc(&lru_count);
 180
 181                 list_add_tail(&gl->gl_lru, &lru_list);
 182                 spin_unlock(&lru_lock);
 183         }
 184 }
 185
 186 void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
 187 {
 188         spin_lock(&gl->gl_spin);
 189         __gfs2_glock_schedule_for_reclaim(gl);
 190         spin_unlock(&gl->gl_spin);
 191 }
 192
 193 /**
 194  * gfs2_glock_put_nolock() - Decrement reference count on glock
 195  * @gl: The glock to put
 196  *
 197  * This function should only be used if the caller has its own reference
 198  * to the glock, in addition to the one it is dropping.
 199  */
 200
 201 void gfs2_glock_put_nolock(struct gfs2_glock *gl)
 202 {
 203         if (atomic_dec_and_test(&gl->gl_ref))
 204                 GLOCK_BUG_ON(gl, 1);
 205 }
 206
 207 /**
 208  * gfs2_glock_put() - Decrement reference count on glock
 209  * @gl: The glock to put
 210  *
 211  */
 212
 213 void gfs2_glock_put(struct gfs2_glock *gl)
 214 {
 215         struct gfs2_sbd *sdp = gl->gl_sbd;
 216         struct address_space *mapping = gfs2_glock2aspace(gl);
 217
 218         if (atomic_dec_and_test(&gl->gl_ref)) {
 219                 spin_lock_bucket(gl->gl_hash);
 220                 hlist_bl_del_rcu(&gl->gl_list);
 221                 spin_unlock_bucket(gl->gl_hash);
 222                 spin_lock(&lru_lock);
 223                 if (!list_empty(&gl->gl_lru)) {
 224                         list_del_init(&gl->gl_lru);
 225                         atomic_dec(&lru_count);
 226                 }
 227                 spin_unlock(&lru_lock);
 228                 GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 229                 GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
 230                 trace_gfs2_glock_put(gl);
 231                 sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 232         }
 233 }
 234
 235 /**
 236  * search_bucket() - Find struct gfs2_glock by lock number
 237  * @bucket: the bucket to search
 238  * @name: The lock name
 239  *
 240  * Returns: NULL, or the struct gfs2_glock with the requested number
 241  */
 242
 243 static struct gfs2_glock *search_bucket(unsigned int hash,
 244                                         const struct gfs2_sbd *sdp,
 245                                         const struct lm_lockname *name)
 246 {
 247         struct gfs2_glock *gl;
 248         struct hlist_bl_node *h;
 249
 250         hlist_bl_for_each_entry_rcu(gl, h, &gl_hash_table[hash], gl_list) {
 251                 if (!lm_name_equal(&gl->gl_name, name))
 252                         continue;
 253                 if (gl->gl_sbd != sdp)
 254                         continue;
 255                 if (atomic_inc_not_zero(&gl->gl_ref))
 256                         return gl;
 257         }
 258
 259         return NULL;
 260 }
 261
 262 /**
 263  * may_grant - check if its ok to grant a new lock
 264  * @gl: The glock
 265  * @gh: The lock request which we wish to grant
 266  *
 267  * Returns: true if its ok to grant the lock
 268  */
 269
 270 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 271 {
 272         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 273         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 274              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 275                 return 0;
 276         if (gl->gl_state == gh->gh_state)
 277                 return 1;
 278         if (gh->gh_flags & GL_EXACT)
 279                 return 0;
 280         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 281                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 282                         return 1;
 283                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 284                         return 1;
 285         }
 286         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 287                 return 1;
 288         return 0;
 289 }
 290
 291 static void gfs2_holder_wake(struct gfs2_holder *gh)
 292 {
 293         clear_bit(HIF_WAIT, &gh->gh_iflags);
 294         smp_mb__after_clear_bit();
 295         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 296 }
 297
 298 /**
 299  * do_error - Something unexpected has happened during a lock request
 300  *
 301  */
 302
 303 static inline void do_error(struct gfs2_glock *gl, const int ret)
 304 {
 305         struct gfs2_holder *gh, *tmp;
 306
 307         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 308                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 309                         continue;
 310                 if (ret & LM_OUT_ERROR)
 311                         gh->gh_error = -EIO;
 312                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 313                         gh->gh_error = GLR_TRYFAILED;
 314                 else
 315                         continue;
 316                 list_del_init(&gh->gh_list);
 317                 trace_gfs2_glock_queue(gh, 0);
 318                 gfs2_holder_wake(gh);
 319         }
 320 }
 321
 322 /**
 323  * do_promote - promote as many requests as possible on the current queue
 324  * @gl: The glock
 325  *
 326  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 327  *          if a type specific operation is underway.
 328  */
 329
 330 static int do_promote(struct gfs2_glock *gl)
 331 __releases(&gl->gl_spin)
 332 __acquires(&gl->gl_spin)
 333 {
 334         const struct gfs2_glock_operations *glops = gl->gl_ops;
 335         struct gfs2_holder *gh, *tmp;
 336         int ret;
 337
 338 restart:
 339         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 340                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 341                         continue;
 342                 if (may_grant(gl, gh)) {
 343                         if (gh->gh_list.prev == &gl->gl_holders &&
 344                             glops->go_lock) {
 345                                 spin_unlock(&gl->gl_spin);
 346                                 /* FIXME: eliminate this eventually */
 347                                 ret = glops->go_lock(gh);
 348                                 spin_lock(&gl->gl_spin);
 349                                 if (ret) {
 350                                         if (ret == 1)
 351                                                 return 2;
 352                                         gh->gh_error = ret;
 353                                         list_del_init(&gh->gh_list);
 354                                         trace_gfs2_glock_queue(gh, 0);
 355                                         gfs2_holder_wake(gh);
 356                                         goto restart;
 357                                 }
 358                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 359                                 trace_gfs2_promote(gh, 1);
 360                                 gfs2_holder_wake(gh);
 361                                 goto restart;
 362                         }
 363                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 364                         trace_gfs2_promote(gh, 0);
 365                         gfs2_holder_wake(gh);
 366                         continue;
 367                 }
 368                 if (gh->gh_list.prev == &gl->gl_holders)
 369                         return 1;
 370                 do_error(gl, 0);
 371                 break;
 372         }
 373         return 0;
 374 }
 375
 376 /**
 377  * find_first_waiter - find the first gh that's waiting for the glock
 378  * @gl: the glock
 379  */
 380
 381 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 382 {
 383         struct gfs2_holder *gh;
 384
 385         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 386                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 387                         return gh;
 388         }
 389         return NULL;
 390 }
 391
 392 /**
 393  * state_change - record that the glock is now in a different state
 394  * @gl: the glock
 395  * @new_state the new state
 396  *
 397  */
 398
 399 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 400 {
 401         int held1, held2;
 402
 403         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 404         held2 = (new_state != LM_ST_UNLOCKED);
 405
 406         if (held1 != held2) {
 407                 if (held2)
 408                         gfs2_glock_hold(gl);
 409                 else
 410                         gfs2_glock_put_nolock(gl);
 411         }
 412         if (held1 && held2 && list_empty(&gl->gl_holders))
 413                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 414
 415         gl->gl_state = new_state;
 416         gl->gl_tchange = jiffies;
 417 }
 418
 419 static void gfs2_demote_wake(struct gfs2_glock *gl)
 420 {
 421         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 422         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 423         smp_mb__after_clear_bit();
 424         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 425 }
 426
 427 /**
 428  * finish_xmote - The DLM has replied to one of our lock requests
 429  * @gl: The glock
 430  * @ret: The status from the DLM
 431  *
 432  */
 433
 434 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 435 {
 436         const struct gfs2_glock_operations *glops = gl->gl_ops;
 437         struct gfs2_holder *gh;
 438         unsigned state = ret & LM_OUT_ST_MASK;
 439         int rv;
 440
 441         spin_lock(&gl->gl_spin);
 442         trace_gfs2_glock_state_change(gl, state);
 443         state_change(gl, state);
 444         gh = find_first_waiter(gl);
 445
 446         /* Demote to UN request arrived during demote to SH or DF */
 447         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 448             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 449                 gl->gl_target = LM_ST_UNLOCKED;
 450
 451         /* Check for state != intended state */
 452         if (unlikely(state != gl->gl_target)) {
 453                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 454                         /* move to back of queue and try next entry */
 455                         if (ret & LM_OUT_CANCELED) {
 456                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 457                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 458                                 gh = find_first_waiter(gl);
 459                                 gl->gl_target = gh->gh_state;
 460                                 goto retry;
 461                         }
 462                         /* Some error or failed "try lock" - report it */
 463                         if ((ret & LM_OUT_ERROR) ||
 464                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 465                                 gl->gl_target = gl->gl_state;
 466                                 do_error(gl, ret);
 467                                 goto out;
 468                         }
 469                 }
 470                 switch(state) {
 471                 /* Unlocked due to conversion deadlock, try again */
 472                 case LM_ST_UNLOCKED:
 473 retry:
 474                         do_xmote(gl, gh, gl->gl_target);
 475                         break;
 476                 /* Conversion fails, unlock and try again */
 477                 case LM_ST_SHARED:
 478                 case LM_ST_DEFERRED:
 479                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 480                         break;
 481                 default: /* Everything else */
 482                         printk(KERN_ERR "GFS2: wanted %u got %u\n", gl->gl_target, state);
 483                         GLOCK_BUG_ON(gl, 1);
 484                 }
 485                 spin_unlock(&gl->gl_spin);
 486                 return;
 487         }
 488
 489         /* Fast path - we got what we asked for */
 490         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 491                 gfs2_demote_wake(gl);
 492         if (state != LM_ST_UNLOCKED) {
 493                 if (glops->go_xmote_bh) {
 494                         spin_unlock(&gl->gl_spin);
 495                         rv = glops->go_xmote_bh(gl, gh);
 496                         spin_lock(&gl->gl_spin);
 497                         if (rv) {
 498                                 do_error(gl, rv);
 499                                 goto out;
 500                         }
 501                 }
 502                 rv = do_promote(gl);
 503                 if (rv == 2)
 504                         goto out_locked;
 505         }
 506 out:
 507         clear_bit(GLF_LOCK, &gl->gl_flags);
 508 out_locked:
 509         spin_unlock(&gl->gl_spin);
 510 }
 511
 512 /**
 513  * do_xmote - Calls the DLM to change the state of a lock
 514  * @gl: The lock state
 515  * @gh: The holder (only for promotes)
 516  * @target: The target lock state
 517  *
 518  */
 519
 520 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 521 __releases(&gl->gl_spin)
 522 __acquires(&gl->gl_spin)
 523 {
 524         const struct gfs2_glock_operations *glops = gl->gl_ops;
 525         struct gfs2_sbd *sdp = gl->gl_sbd;
 526         unsigned int lck_flags = gh ? gh->gh_flags : 0;
 527         int ret;
 528
 529         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 530                       LM_FLAG_PRIORITY);
 531         GLOCK_BUG_ON(gl, gl->gl_state == target);
 532         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 533         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 534             glops->go_inval) {
 535                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 536                 do_error(gl, 0); /* Fail queued try locks */
 537         }
 538         gl->gl_req = target;
 539         spin_unlock(&gl->gl_spin);
 540         if (glops->go_xmote_th)
 541                 glops->go_xmote_th(gl);
 542         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 543                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 544         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 545
 546         gfs2_glock_hold(gl);
 547         if (target != LM_ST_UNLOCKED && (gl->gl_state == LM_ST_SHARED ||
 548             gl->gl_state == LM_ST_DEFERRED) &&
 549             !(lck_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 550                 lck_flags |= LM_FLAG_TRY_1CB;
 551
 552         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 553                 /* lock_dlm */
 554                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 555                 GLOCK_BUG_ON(gl, ret);
 556         } else { /* lock_nolock */
 557                 finish_xmote(gl, target);
 558                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
 559                         gfs2_glock_put(gl);
 560         }
 561
 562         spin_lock(&gl->gl_spin);
 563 }
 564
 565 /**
 566  * find_first_holder - find the first "holder" gh
 567  * @gl: the glock
 568  */
 569
 570 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 571 {
 572         struct gfs2_holder *gh;
 573
 574         if (!list_empty(&gl->gl_holders)) {
 575                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 576                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 577                         return gh;
 578         }
 579         return NULL;
 580 }
 581
 582 /**
 583  * run_queue - do all outstanding tasks related to a glock
 584  * @gl: The glock in question
 585  * @nonblock: True if we must not block in run_queue
 586  *
 587  */
 588
 589 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 590 __releases(&gl->gl_spin)
 591 __acquires(&gl->gl_spin)
 592 {
 593         struct gfs2_holder *gh = NULL;
 594         int ret;
 595
 596         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 597                 return;
 598
 599         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 600
 601         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 602             gl->gl_demote_state != gl->gl_state) {
 603                 if (find_first_holder(gl))
 604                         goto out_unlock;
 605                 if (nonblock)
 606                         goto out_sched;
 607                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 608                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 609                 gl->gl_target = gl->gl_demote_state;
 610         } else {
 611                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 612                         gfs2_demote_wake(gl);
 613                 ret = do_promote(gl);
 614                 if (ret == 0)
 615                         goto out_unlock;
 616                 if (ret == 2)
 617                         goto out;
 618                 gh = find_first_waiter(gl);
 619                 gl->gl_target = gh->gh_state;
 620                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 621                         do_error(gl, 0); /* Fail queued try locks */
 622         }
 623         do_xmote(gl, gh, gl->gl_target);
 624 out:
 625         return;
 626
 627 out_sched:
 628         clear_bit(GLF_LOCK, &gl->gl_flags);
 629         smp_mb__after_clear_bit();
 630         gfs2_glock_hold(gl);
 631         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
 632                 gfs2_glock_put_nolock(gl);
 633         return;
 634
 635 out_unlock:
 636         clear_bit(GLF_LOCK, &gl->gl_flags);
 637         smp_mb__after_clear_bit();
 638         return;
 639 }
 640
 641 static void delete_work_func(struct work_struct *work)
 642 {
 643         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 644         struct gfs2_sbd *sdp = gl->gl_sbd;
 645         struct gfs2_inode *ip;
 646         struct inode *inode;
 647         u64 no_addr = gl->gl_name.ln_number;
 648
 649         ip = gl->gl_object;
 650         /* Note: Unsafe to dereference ip as we don't hold right refs/locks */
 651
 652         if (ip)
 653                 inode = gfs2_ilookup(sdp->sd_vfs, no_addr);
 654         else
 655                 inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 656         if (inode && !IS_ERR(inode)) {
 657                 d_prune_aliases(inode);
 658                 iput(inode);
 659         }
 660         gfs2_glock_put(gl);
 661 }
 662
 663 static void glock_work_func(struct work_struct *work)
 664 {
 665         unsigned long delay = 0;
 666         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 667         int drop_ref = 0;
 668
 669         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 670                 finish_xmote(gl, gl->gl_reply);
 671                 drop_ref = 1;
 672         }
 673         spin_lock(&gl->gl_spin);
 674         if (test_and_clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 675             gl->gl_state != LM_ST_UNLOCKED &&
 676             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 677                 unsigned long holdtime, now = jiffies;
 678                 holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
 679                 if (time_before(now, holdtime))
 680                         delay = holdtime - now;
 681                 set_bit(delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE, &gl->gl_flags);
 682         }
 683         run_queue(gl, 0);
 684         spin_unlock(&gl->gl_spin);
 685         if (!delay ||
 686             queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
 687                 gfs2_glock_put(gl);
 688         if (drop_ref)
 689                 gfs2_glock_put(gl);
 690 }
 691
 692 /**
 693  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 694  * @sdp: The GFS2 superblock
 695  * @number: the lock number
 696  * @glops: The glock_operations to use
 697  * @create: If 0, don't create the glock if it doesn't exist
 698  * @glp: the glock is returned here
 699  *
 700  * This does not lock a glock, just finds/creates structures for one.
 701  *
 702  * Returns: errno
 703  */
 704
 705 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 706                    const struct gfs2_glock_operations *glops, int create,
 707                    struct gfs2_glock **glp)
 708 {
 709         struct super_block *s = sdp->sd_vfs;
 710         struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
 711         struct gfs2_glock *gl, *tmp;
 712         unsigned int hash = gl_hash(sdp, &name);
 713         struct address_space *mapping;
 714         struct kmem_cache *cachep;
 715
 716         rcu_read_lock();
 717         gl = search_bucket(hash, sdp, &name);
 718         rcu_read_unlock();
 719
 720         *glp = gl;
 721         if (gl)
 722                 return 0;
 723         if (!create)
 724                 return -ENOENT;
 725
 726         if (glops->go_flags & GLOF_ASPACE)
 727                 cachep = gfs2_glock_aspace_cachep;
 728         else
 729                 cachep = gfs2_glock_cachep;
 730         gl = kmem_cache_alloc(cachep, GFP_KERNEL);
 731         if (!gl)
 732                 return -ENOMEM;
 733
 734         atomic_inc(&sdp->sd_glock_disposal);
 735         gl->gl_flags = 0;
 736         gl->gl_name = name;
 737         atomic_set(&gl->gl_ref, 1);
 738         gl->gl_state = LM_ST_UNLOCKED;
 739         gl->gl_target = LM_ST_UNLOCKED;
 740         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 741         gl->gl_hash = hash;
 742         gl->gl_ops = glops;
 743         snprintf(gl->gl_strname, GDLM_STRNAME_BYTES, "%8x%16llx", name.ln_type, (unsigned long long)number);
 744         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 745         gl->gl_lksb.sb_lvbptr = gl->gl_lvb;
 746         gl->gl_tchange = jiffies;
 747         gl->gl_object = NULL;
 748         gl->gl_sbd = sdp;
 749         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 750         INIT_WORK(&gl->gl_delete, delete_work_func);
 751
 752         mapping = gfs2_glock2aspace(gl);
 753         if (mapping) {
 754                 mapping->a_ops = &gfs2_meta_aops;
 755                 mapping->host = s->s_bdev->bd_inode;
 756                 mapping->flags = 0;
 757                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 758                 mapping->assoc_mapping = NULL;
 759                 mapping->backing_dev_info = s->s_bdi;
 760                 mapping->writeback_index = 0;
 761         }
 762
 763         spin_lock_bucket(hash);
 764         tmp = search_bucket(hash, sdp, &name);
 765         if (tmp) {
 766                 spin_unlock_bucket(hash);
 767                 kmem_cache_free(cachep, gl);
 768                 atomic_dec(&sdp->sd_glock_disposal);
 769                 gl = tmp;
 770         } else {
 771                 hlist_bl_add_head_rcu(&gl->gl_list, &gl_hash_table[hash]);
 772                 spin_unlock_bucket(hash);
 773         }
 774
 775         *glp = gl;
 776
 777         return 0;
 778 }
 779
 780 /**
 781  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 782  * @gl: the glock
 783  * @state: the state we're requesting
 784  * @flags: the modifier flags
 785  * @gh: the holder structure
 786  *
 787  */
 788
 789 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
 790                       struct gfs2_holder *gh)
 791 {
 792         INIT_LIST_HEAD(&gh->gh_list);
 793         gh->gh_gl = gl;
 794         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 795         gh->gh_owner_pid = get_pid(task_pid(current));
 796         gh->gh_state = state;
 797         gh->gh_flags = flags;
 798         gh->gh_error = 0;
 799         gh->gh_iflags = 0;
 800         gfs2_glock_hold(gl);
 801 }
 802
 803 /**
 804  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 805  * @state: the state we're requesting
 806  * @flags: the modifier flags
 807  * @gh: the holder structure
 808  *
 809  * Don't mess with the glock.
 810  *
 811  */
 812
 813 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
 814 {
 815         gh->gh_state = state;
 816         gh->gh_flags = flags;
 817         gh->gh_iflags = 0;
 818         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 819         if (gh->gh_owner_pid)
 820                 put_pid(gh->gh_owner_pid);
 821         gh->gh_owner_pid = get_pid(task_pid(current));
 822 }
 823
 824 /**
 825  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 826  * @gh: the holder structure
 827  *
 828  */
 829
 830 void gfs2_holder_uninit(struct gfs2_holder *gh)
 831 {
 832         put_pid(gh->gh_owner_pid);
 833         gfs2_glock_put(gh->gh_gl);
 834         gh->gh_gl = NULL;
 835         gh->gh_ip = 0;
 836 }
 837
 838 /**
 839  * gfs2_glock_holder_wait
 840  * @word: unused
 841  *
 842  * This function and gfs2_glock_demote_wait both show up in the WCHAN
 843  * field. Thus I've separated these otherwise identical functions in
 844  * order to be more informative to the user.
 845  */
 846
 847 static int gfs2_glock_holder_wait(void *word)
 848 {
 849         schedule();
 850         return 0;
 851 }
 852
 853 static int gfs2_glock_demote_wait(void *word)
 854 {
 855         schedule();
 856         return 0;
 857 }
 858
 859 static void wait_on_holder(struct gfs2_holder *gh)
 860 {
 861         might_sleep();
 862         wait_on_bit(&gh->gh_iflags, HIF_WAIT, gfs2_glock_holder_wait, TASK_UNINTERRUPTIBLE);
 863 }
 864
 865 static void wait_on_demote(struct gfs2_glock *gl)
 866 {
 867         might_sleep();
 868         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, gfs2_glock_demote_wait, TASK_UNINTERRUPTIBLE);
 869 }
 870
 871 /**
 872  * handle_callback - process a demote request
 873  * @gl: the glock
 874  * @state: the state the caller wants us to change to
 875  *
 876  * There are only two requests that we are going to see in actual
 877  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 878  */
 879
 880 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 881                             unsigned long delay)
 882 {
 883         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 884
 885         set_bit(bit, &gl->gl_flags);
 886         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 887                 gl->gl_demote_state = state;
 888                 gl->gl_demote_time = jiffies;
 889         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 890                         gl->gl_demote_state != state) {
 891                 gl->gl_demote_state = LM_ST_UNLOCKED;
 892         }
 893         if (gl->gl_ops->go_callback)
 894                 gl->gl_ops->go_callback(gl);
 895         trace_gfs2_demote_rq(gl);
 896 }
 897
 898 /**
 899  * gfs2_glock_wait - wait on a glock acquisition
 900  * @gh: the glock holder
 901  *
 902  * Returns: 0 on success
 903  */
 904
 905 int gfs2_glock_wait(struct gfs2_holder *gh)
 906 {
 907         wait_on_holder(gh);
 908         return gh->gh_error;
 909 }
 910
 911 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 912 {
 913         struct va_format vaf;
 914         va_list args;
 915
 916         va_start(args, fmt);
 917
 918         if (seq) {
 919                 struct gfs2_glock_iter *gi = seq->private;
 920                 vsprintf(gi->string, fmt, args);
 921                 seq_printf(seq, gi->string);
 922         } else {
 923                 vaf.fmt = fmt;
 924                 vaf.va = &args;
 925
 926                 printk(KERN_ERR " %pV", &vaf);
 927         }
 928
 929         va_end(args);
 930 }
 931
 932 /**
 933  * add_to_queue - Add a holder to the wait queue (but look for recursion)
 934  * @gh: the holder structure to add
 935  *
 936  * Eventually we should move the recursive locking trap to a
 937  * debugging option or something like that. This is the fast
 938  * path and needs to have the minimum number of distractions.
 939  *
 940  */
 941
 942 static inline void add_to_queue(struct gfs2_holder *gh)
 943 __releases(&gl->gl_spin)
 944 __acquires(&gl->gl_spin)
 945 {
 946         struct gfs2_glock *gl = gh->gh_gl;
 947         struct gfs2_sbd *sdp = gl->gl_sbd;
 948         struct list_head *insert_pt = NULL;
 949         struct gfs2_holder *gh2;
 950         int try_lock = 0;
 951
 952         BUG_ON(gh->gh_owner_pid == NULL);
 953         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
 954                 BUG();
 955
 956         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 957                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 958                         try_lock = 1;
 959                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 960                         goto fail;
 961         }
 962
 963         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
 964                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
 965                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
 966                         goto trap_recursive;
 967                 if (try_lock &&
 968                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) &&
 969                     !may_grant(gl, gh)) {
 970 fail:
 971                         gh->gh_error = GLR_TRYFAILED;
 972                         gfs2_holder_wake(gh);
 973                         return;
 974                 }
 975                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
 976                         continue;
 977                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
 978                         insert_pt = &gh2->gh_list;
 979         }
 980         set_bit(GLF_QUEUED, &gl->gl_flags);
 981         trace_gfs2_glock_queue(gh, 1);
 982         if (likely(insert_pt == NULL)) {
 983                 list_add_tail(&gh->gh_list, &gl->gl_holders);
 984                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
 985                         goto do_cancel;
 986                 return;
 987         }
 988         list_add_tail(&gh->gh_list, insert_pt);
 989 do_cancel:
 990         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 991         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
 992                 spin_unlock(&gl->gl_spin);
 993                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
 994                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
 995                 spin_lock(&gl->gl_spin);
 996         }
 997         return;
 998
 999 trap_recursive:
1000         print_symbol(KERN_ERR "original: %s\n", gh2->gh_ip);
1001         printk(KERN_ERR "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1002         printk(KERN_ERR "lock type: %d req lock state : %d\n",
1003                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1004         print_symbol(KERN_ERR "new: %s\n", gh->gh_ip);
1005         printk(KERN_ERR "pid: %d\n", pid_nr(gh->gh_owner_pid));
1006         printk(KERN_ERR "lock type: %d req lock state : %d\n",
1007                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1008         __dump_glock(NULL, gl);
1009         BUG();
1010 }
1011
1012 /**
1013  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1014  * @gh: the holder structure
1015  *
1016  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1017  *
1018  * Returns: 0, GLR_TRYFAILED, or errno on failure
1019  */
1020
1021 int gfs2_glock_nq(struct gfs2_holder *gh)
1022 {
1023         struct gfs2_glock *gl = gh->gh_gl;
1024         struct gfs2_sbd *sdp = gl->gl_sbd;
1025         int error = 0;
1026
1027         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1028                 return -EIO;
1029
1030         spin_lock(&gl->gl_spin);
1031         add_to_queue(gh);
1032         if ((LM_FLAG_NOEXP & gh->gh_flags) &&
1033             test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1034                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1035         run_queue(gl, 1);
1036         spin_unlock(&gl->gl_spin);
1037
1038         if (!(gh->gh_flags & GL_ASYNC))
1039                 error = gfs2_glock_wait(gh);
1040
1041         return error;
1042 }
1043
1044 /**
1045  * gfs2_glock_poll - poll to see if an async request has been completed
1046  * @gh: the holder
1047  *
1048  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1049  */
1050
1051 int gfs2_glock_poll(struct gfs2_holder *gh)
1052 {
1053         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1054 }
1055
1056 /**
1057  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1058  * @gh: the glock holder
1059  *
1060  */
1061
1062 void gfs2_glock_dq(struct gfs2_holder *gh)
1063 {
1064         struct gfs2_glock *gl = gh->gh_gl;
1065         const struct gfs2_glock_operations *glops = gl->gl_ops;
1066         unsigned delay = 0;
1067         int fast_path = 0;
1068
1069         spin_lock(&gl->gl_spin);
1070         if (gh->gh_flags & GL_NOCACHE)
1071                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1072
1073         list_del_init(&gh->gh_list);
1074         if (find_first_holder(gl) == NULL) {
1075                 if (glops->go_unlock) {
1076                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1077                         spin_unlock(&gl->gl_spin);
1078                         glops->go_unlock(gh);
1079                         spin_lock(&gl->gl_spin);
1080                         clear_bit(GLF_LOCK, &gl->gl_flags);
1081                 }
1082                 if (list_empty(&gl->gl_holders) &&
1083                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1084                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1085                         fast_path = 1;
1086         }
1087         __gfs2_glock_schedule_for_reclaim(gl);
1088         trace_gfs2_glock_queue(gh, 0);
1089         spin_unlock(&gl->gl_spin);
1090         if (likely(fast_path))
1091                 return;
1092
1093         gfs2_glock_hold(gl);
1094         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1095             !test_bit(GLF_DEMOTE, &gl->gl_flags))
1096                 delay = gl->gl_ops->go_min_hold_time;
1097         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1098                 gfs2_glock_put(gl);
1099 }
1100
1101 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1102 {
1103         struct gfs2_glock *gl = gh->gh_gl;
1104         gfs2_glock_dq(gh);
1105         wait_on_demote(gl);
1106 }
1107
1108 /**
1109  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1110  * @gh: the holder structure
1111  *
1112  */
1113
1114 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1115 {
1116         gfs2_glock_dq(gh);
1117         gfs2_holder_uninit(gh);
1118 }
1119
1120 /**
1121  * gfs2_glock_nq_num - acquire a glock based on lock number
1122  * @sdp: the filesystem
1123  * @number: the lock number
1124  * @glops: the glock operations for the type of glock
1125  * @state: the state to acquire the glock in
1126  * @flags: modifier flags for the aquisition
1127  * @gh: the struct gfs2_holder
1128  *
1129  * Returns: errno
1130  */
1131
1132 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1133                       const struct gfs2_glock_operations *glops,
1134                       unsigned int state, int flags, struct gfs2_holder *gh)
1135 {
1136         struct gfs2_glock *gl;
1137         int error;
1138
1139         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1140         if (!error) {
1141                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1142                 gfs2_glock_put(gl);
1143         }
1144
1145         return error;
1146 }
1147
1148 /**
1149  * glock_compare - Compare two struct gfs2_glock structures for sorting
1150  * @arg_a: the first structure
1151  * @arg_b: the second structure
1152  *
1153  */
1154
1155 static int glock_compare(const void *arg_a, const void *arg_b)
1156 {
1157         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1158         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1159         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1160         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1161
1162         if (a->ln_number > b->ln_number)
1163                 return 1;
1164         if (a->ln_number < b->ln_number)
1165                 return -1;
1166         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1167         return 0;
1168 }
1169
1170 /**
1171  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1172  * @num_gh: the number of structures
1173  * @ghs: an array of struct gfs2_holder structures
1174  *
1175  * Returns: 0 on success (all glocks acquired),
1176  *          errno on failure (no glocks acquired)
1177  */
1178
1179 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1180                      struct gfs2_holder **p)
1181 {
1182         unsigned int x;
1183         int error = 0;
1184
1185         for (x = 0; x < num_gh; x++)
1186                 p[x] = &ghs[x];
1187
1188         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1189
1190         for (x = 0; x < num_gh; x++) {
1191                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1192
1193                 error = gfs2_glock_nq(p[x]);
1194                 if (error) {
1195                         while (x--)
1196                                 gfs2_glock_dq(p[x]);
1197                         break;
1198                 }
1199         }
1200
1201         return error;
1202 }
1203
1204 /**
1205  * gfs2_glock_nq_m - acquire multiple glocks
1206  * @num_gh: the number of structures
1207  * @ghs: an array of struct gfs2_holder structures
1208  *
1209  *
1210  * Returns: 0 on success (all glocks acquired),
1211  *          errno on failure (no glocks acquired)
1212  */
1213
1214 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1215 {
1216         struct gfs2_holder *tmp[4];
1217         struct gfs2_holder **pph = tmp;
1218         int error = 0;
1219
1220         switch(num_gh) {
1221         case 0:
1222                 return 0;
1223         case 1:
1224                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1225                 return gfs2_glock_nq(ghs);
1226         default:
1227                 if (num_gh <= 4)
1228                         break;
1229                 pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
1230                 if (!pph)
1231                         return -ENOMEM;
1232         }
1233
1234         error = nq_m_sync(num_gh, ghs, pph);
1235
1236         if (pph != tmp)
1237                 kfree(pph);
1238
1239         return error;
1240 }
1241
1242 /**
1243  * gfs2_glock_dq_m - release multiple glocks
1244  * @num_gh: the number of structures
1245  * @ghs: an array of struct gfs2_holder structures
1246  *
1247  */
1248
1249 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1250 {
1251         while (num_gh--)
1252                 gfs2_glock_dq(&ghs[num_gh]);
1253 }
1254
1255 /**
1256  * gfs2_glock_dq_uninit_m - release multiple glocks
1257  * @num_gh: the number of structures
1258  * @ghs: an array of struct gfs2_holder structures
1259  *
1260  */
1261
1262 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1263 {
1264         while (num_gh--)
1265                 gfs2_glock_dq_uninit(&ghs[num_gh]);
1266 }
1267
1268 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1269 {
1270         unsigned long delay = 0;
1271         unsigned long holdtime;
1272         unsigned long now = jiffies;
1273
1274         gfs2_glock_hold(gl);
1275         holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
1276         if (test_bit(GLF_QUEUED, &gl->gl_flags)) {
1277                 if (time_before(now, holdtime))
1278                         delay = holdtime - now;
1279                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1280                         delay = gl->gl_ops->go_min_hold_time;
1281         }
1282
1283         spin_lock(&gl->gl_spin);
1284         handle_callback(gl, state, delay);
1285         spin_unlock(&gl->gl_spin);
1286         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1287                 gfs2_glock_put(gl);
1288 }
1289
1290 /**
1291  * gfs2_should_freeze - Figure out if glock should be frozen
1292  * @gl: The glock in question
1293  *
1294  * Glocks are not frozen if (a) the result of the dlm operation is
1295  * an error, (b) the locking operation was an unlock operation or
1296  * (c) if there is a "noexp" flagged request anywhere in the queue
1297  *
1298  * Returns: 1 if freezing should occur, 0 otherwise
1299  */
1300
1301 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1302 {
1303         const struct gfs2_holder *gh;
1304
1305         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1306                 return 0;
1307         if (gl->gl_target == LM_ST_UNLOCKED)
1308                 return 0;
1309
1310         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1311                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1312                         continue;
1313                 if (LM_FLAG_NOEXP & gh->gh_flags)
1314                         return 0;
1315         }
1316
1317         return 1;
1318 }
1319
1320 /**
1321  * gfs2_glock_complete - Callback used by locking
1322  * @gl: Pointer to the glock
1323  * @ret: The return value from the dlm
1324  *
1325  * The gl_reply field is under the gl_spin lock so that it is ok
1326  * to use a bitfield shared with other glock state fields.
1327  */
1328
1329 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1330 {
1331         struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
1332
1333         spin_lock(&gl->gl_spin);
1334         gl->gl_reply = ret;
1335
1336         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_flags))) {
1337                 if (gfs2_should_freeze(gl)) {
1338                         set_bit(GLF_FROZEN, &gl->gl_flags);
1339                         spin_unlock(&gl->gl_spin);
1340                         return;
1341                 }
1342         }
1343
1344         spin_unlock(&gl->gl_spin);
1345         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1346         smp_wmb();
1347         gfs2_glock_hold(gl);
1348         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1349                 gfs2_glock_put(gl);
1350 }
1351
1352
1353 static int gfs2_shrink_glock_memory(struct shrinker *shrink, int nr, gfp_t gfp_mask)
1354 {
1355         struct gfs2_glock *gl;
1356         int may_demote;
1357         int nr_skipped = 0;
1358         LIST_HEAD(skipped);
1359
1360         if (nr == 0)
1361                 goto out;
1362
1363         if (!(gfp_mask & __GFP_FS))
1364                 return -1;
1365
1366         spin_lock(&lru_lock);
1367         while(nr && !list_empty(&lru_list)) {
1368                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1369                 list_del_init(&gl->gl_lru);
1370                 atomic_dec(&lru_count);
1371
1372                 /* Test for being demotable */
1373                 if (!test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1374                         gfs2_glock_hold(gl);
1375                         spin_unlock(&lru_lock);
1376                         spin_lock(&gl->gl_spin);
1377                         may_demote = demote_ok(gl);
1378                         if (may_demote) {
1379                                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1380                                 nr--;
1381                         }
1382                         clear_bit(GLF_LOCK, &gl->gl_flags);
1383                         smp_mb__after_clear_bit();
1384                         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1385                                 gfs2_glock_put_nolock(gl);
1386                         spin_unlock(&gl->gl_spin);
1387                         spin_lock(&lru_lock);
1388                         continue;
1389                 }
1390                 nr_skipped++;
1391                 list_add(&gl->gl_lru, &skipped);
1392         }
1393         list_splice(&skipped, &lru_list);
1394         atomic_add(nr_skipped, &lru_count);
1395         spin_unlock(&lru_lock);
1396 out:
1397         return (atomic_read(&lru_count) / 100) * sysctl_vfs_cache_pressure;
1398 }
1399
1400 static struct shrinker glock_shrinker = {
1401         .shrink = gfs2_shrink_glock_memory,
1402         .seeks = DEFAULT_SEEKS,
1403 };
1404
1405 /**
1406  * examine_bucket - Call a function for glock in a hash bucket
1407  * @examiner: the function
1408  * @sdp: the filesystem
1409  * @bucket: the bucket
1410  *
1411  */
1412
1413 static void examine_bucket(glock_examiner examiner, const struct gfs2_sbd *sdp,
1414                           unsigned int hash)
1415 {
1416         struct gfs2_glock *gl;
1417         struct hlist_bl_head *head = &gl_hash_table[hash];
1418         struct hlist_bl_node *pos;
1419
1420         rcu_read_lock();
1421         hlist_bl_for_each_entry_rcu(gl, pos, head, gl_list) {
1422                 if ((gl->gl_sbd == sdp) && atomic_read(&gl->gl_ref))
1423                         examiner(gl);
1424         }
1425         rcu_read_unlock();
1426         cond_resched();
1427 }
1428
1429 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1430 {
1431         unsigned x;
1432
1433         for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
1434                 examine_bucket(examiner, sdp, x);
1435 }
1436
1437
1438 /**
1439  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1440  * @gl: The glock to thaw
1441  *
1442  * N.B. When we freeze a glock, we leave a ref to the glock outstanding,
1443  * so this has to result in the ref count being dropped by one.
1444  */
1445
1446 static void thaw_glock(struct gfs2_glock *gl)
1447 {
1448         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1449                 return;
1450         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1451         gfs2_glock_hold(gl);
1452         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1453                 gfs2_glock_put(gl);
1454 }
1455
1456 /**
1457  * clear_glock - look at a glock and see if we can free it from glock cache
1458  * @gl: the glock to look at
1459  *
1460  */
1461
1462 static void clear_glock(struct gfs2_glock *gl)
1463 {
1464         spin_lock(&lru_lock);
1465         if (!list_empty(&gl->gl_lru)) {
1466                 list_del_init(&gl->gl_lru);
1467                 atomic_dec(&lru_count);
1468         }
1469         spin_unlock(&lru_lock);
1470
1471         spin_lock(&gl->gl_spin);
1472         if (gl->gl_state != LM_ST_UNLOCKED)
1473                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1474         spin_unlock(&gl->gl_spin);
1475         gfs2_glock_hold(gl);
1476         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1477                 gfs2_glock_put(gl);
1478 }
1479
1480 /**
1481  * gfs2_glock_thaw - Thaw any frozen glocks
1482  * @sdp: The super block
1483  *
1484  */
1485
1486 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1487 {
1488         glock_hash_walk(thaw_glock, sdp);
1489 }
1490
1491 static int dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1492 {
1493         int ret;
1494         spin_lock(&gl->gl_spin);
1495         ret = __dump_glock(seq, gl);
1496         spin_unlock(&gl->gl_spin);
1497         return ret;
1498 }
1499
1500 static void dump_glock_func(struct gfs2_glock *gl)
1501 {
1502         dump_glock(NULL, gl);
1503 }
1504
1505 /**
1506  * gfs2_gl_hash_clear - Empty out the glock hash table
1507  * @sdp: the filesystem
1508  * @wait: wait until it's all gone
1509  *
1510  * Called when unmounting the filesystem.
1511  */
1512
1513 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1514 {
1515         glock_hash_walk(clear_glock, sdp);
1516         flush_workqueue(glock_workqueue);
1517         wait_event(sdp->sd_glock_wait, atomic_read(&sdp->sd_glock_disposal) == 0);
1518         glock_hash_walk(dump_glock_func, sdp);
1519 }
1520
1521 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1522 {
1523         struct gfs2_glock *gl = ip->i_gl;
1524         int ret;
1525
1526         ret = gfs2_truncatei_resume(ip);
1527         gfs2_assert_withdraw(gl->gl_sbd, ret == 0);
1528
1529         spin_lock(&gl->gl_spin);
1530         clear_bit(GLF_LOCK, &gl->gl_flags);
1531         run_queue(gl, 1);
1532         spin_unlock(&gl->gl_spin);
1533 }
1534
1535 static const char *state2str(unsigned state)
1536 {
1537         switch(state) {
1538         case LM_ST_UNLOCKED:
1539                 return "UN";
1540         case LM_ST_SHARED:
1541                 return "SH";
1542         case LM_ST_DEFERRED:
1543                 return "DF";
1544         case LM_ST_EXCLUSIVE:
1545                 return "EX";
1546         }
1547         return "??";
1548 }
1549
1550 static const char *hflags2str(char *buf, unsigned flags, unsigned long iflags)
1551 {
1552         char *p = buf;
1553         if (flags & LM_FLAG_TRY)
1554                 *p++ = 't';
1555         if (flags & LM_FLAG_TRY_1CB)
1556                 *p++ = 'T';
1557         if (flags & LM_FLAG_NOEXP)
1558                 *p++ = 'e';
1559         if (flags & LM_FLAG_ANY)
1560                 *p++ = 'A';
1561         if (flags & LM_FLAG_PRIORITY)
1562                 *p++ = 'p';
1563         if (flags & GL_ASYNC)
1564                 *p++ = 'a';
1565         if (flags & GL_EXACT)
1566                 *p++ = 'E';
1567         if (flags & GL_NOCACHE)
1568                 *p++ = 'c';
1569         if (test_bit(HIF_HOLDER, &iflags))
1570                 *p++ = 'H';
1571         if (test_bit(HIF_WAIT, &iflags))
1572                 *p++ = 'W';
1573         if (test_bit(HIF_FIRST, &iflags))
1574                 *p++ = 'F';
1575         *p = 0;
1576         return buf;
1577 }
1578
1579 /**
1580  * dump_holder - print information about a glock holder
1581  * @seq: the seq_file struct
1582  * @gh: the glock holder
1583  *
1584  * Returns: 0 on success, -ENOBUFS when we run out of space
1585  */
1586
1587 static int dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1588 {
1589         struct task_struct *gh_owner = NULL;
1590         char flags_buf[32];
1591
1592         if (gh->gh_owner_pid)
1593                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1594         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1595                        state2str(gh->gh_state),
1596                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1597                        gh->gh_error,
1598                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1599                        gh_owner ? gh_owner->comm : "(ended)",
1600                        (void *)gh->gh_ip);
1601         return 0;
1602 }
1603
1604 static const char *gflags2str(char *buf, const unsigned long *gflags)
1605 {
1606         char *p = buf;
1607         if (test_bit(GLF_LOCK, gflags))
1608                 *p++ = 'l';
1609         if (test_bit(GLF_DEMOTE, gflags))
1610                 *p++ = 'D';
1611         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1612                 *p++ = 'd';
1613         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1614                 *p++ = 'p';
1615         if (test_bit(GLF_DIRTY, gflags))
1616                 *p++ = 'y';
1617         if (test_bit(GLF_LFLUSH, gflags))
1618                 *p++ = 'f';
1619         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1620                 *p++ = 'i';
1621         if (test_bit(GLF_REPLY_PENDING, gflags))
1622                 *p++ = 'r';
1623         if (test_bit(GLF_INITIAL, gflags))
1624                 *p++ = 'I';
1625         if (test_bit(GLF_FROZEN, gflags))
1626                 *p++ = 'F';
1627         if (test_bit(GLF_QUEUED, gflags))
1628                 *p++ = 'q';
1629         *p = 0;
1630         return buf;
1631 }
1632
1633 /**
1634  * __dump_glock - print information about a glock
1635  * @seq: The seq_file struct
1636  * @gl: the glock
1637  *
1638  * The file format is as follows:
1639  * One line per object, capital letters are used to indicate objects
1640  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1641  * other objects are indented by a single space and follow the glock to
1642  * which they are related. Fields are indicated by lower case letters
1643  * followed by a colon and the field value, except for strings which are in
1644  * [] so that its possible to see if they are composed of spaces for
1645  * example. The field's are n = number (id of the object), f = flags,
1646  * t = type, s = state, r = refcount, e = error, p = pid.
1647  *
1648  * Returns: 0 on success, -ENOBUFS when we run out of space
1649  */
1650
1651 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1652 {
1653         const struct gfs2_glock_operations *glops = gl->gl_ops;
1654         unsigned long long dtime;
1655         const struct gfs2_holder *gh;
1656         char gflags_buf[32];
1657         int error = 0;
1658
1659         dtime = jiffies - gl->gl_demote_time;
1660         dtime *= 1000000/HZ; /* demote time in uSec */
1661         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1662                 dtime = 0;
1663         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d r:%d\n",
1664                   state2str(gl->gl_state),
1665                   gl->gl_name.ln_type,
1666                   (unsigned long long)gl->gl_name.ln_number,
1667                   gflags2str(gflags_buf, &gl->gl_flags),
1668                   state2str(gl->gl_target),
1669                   state2str(gl->gl_demote_state), dtime,
1670                   atomic_read(&gl->gl_ail_count),
1671                   atomic_read(&gl->gl_ref));
1672
1673         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1674                 error = dump_holder(seq, gh);
1675                 if (error)
1676                         goto out;
1677         }
1678         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1679                 error = glops->go_dump(seq, gl);
1680 out:
1681         return error;
1682 }
1683
1684
1685
1686
1687 int __init gfs2_glock_init(void)
1688 {
1689         unsigned i;
1690         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
1691                 INIT_HLIST_BL_HEAD(&gl_hash_table[i]);
1692         }
1693
1694         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1695                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1696         if (IS_ERR(glock_workqueue))
1697                 return PTR_ERR(glock_workqueue);
1698         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1699                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1700                                                 0);
1701         if (IS_ERR(gfs2_delete_workqueue)) {
1702                 destroy_workqueue(glock_workqueue);
1703                 return PTR_ERR(gfs2_delete_workqueue);
1704         }
1705
1706         register_shrinker(&glock_shrinker);
1707
1708         return 0;
1709 }
1710
1711 void gfs2_glock_exit(void)
1712 {
1713         unregister_shrinker(&glock_shrinker);
1714         destroy_workqueue(glock_workqueue);
1715         destroy_workqueue(gfs2_delete_workqueue);
1716 }
1717
1718 static inline struct gfs2_glock *glock_hash_chain(unsigned hash)
1719 {
1720         return hlist_bl_entry(hlist_bl_first_rcu(&gl_hash_table[hash]),
1721                               struct gfs2_glock, gl_list);
1722 }
1723
1724 static inline struct gfs2_glock *glock_hash_next(struct gfs2_glock *gl)
1725 {
1726         return hlist_bl_entry(rcu_dereference(gl->gl_list.next),
1727                               struct gfs2_glock, gl_list);
1728 }
1729
1730 static int gfs2_glock_iter_next(struct gfs2_glock_iter *gi)
1731 {
1732         struct gfs2_glock *gl;
1733
1734         do {
1735                 gl = gi->gl;
1736                 if (gl) {
1737                         gi->gl = glock_hash_next(gl);
1738                 } else {
1739                         gi->gl = glock_hash_chain(gi->hash);
1740                 }
1741                 while (gi->gl == NULL) {
1742                         gi->hash++;
1743                         if (gi->hash >= GFS2_GL_HASH_SIZE) {
1744                                 rcu_read_unlock();
1745                                 return 1;
1746                         }
1747                         gi->gl = glock_hash_chain(gi->hash);
1748                 }
1749         /* Skip entries for other sb and dead entries */
1750         } while (gi->sdp != gi->gl->gl_sbd || atomic_read(&gi->gl->gl_ref) == 0);
1751
1752         return 0;
1753 }
1754
1755 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1756 {
1757         struct gfs2_glock_iter *gi = seq->private;
1758         loff_t n = *pos;
1759
1760         gi->hash = 0;
1761         rcu_read_lock();
1762
1763         do {
1764                 if (gfs2_glock_iter_next(gi))
1765                         return NULL;
1766         } while (n--);
1767
1768         return gi->gl;
1769 }
1770
1771 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1772                                  loff_t *pos)
1773 {
1774         struct gfs2_glock_iter *gi = seq->private;
1775
1776         (*pos)++;
1777
1778         if (gfs2_glock_iter_next(gi))
1779                 return NULL;
1780
1781         return gi->gl;
1782 }
1783
1784 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1785 {
1786         struct gfs2_glock_iter *gi = seq->private;
1787
1788         if (gi->gl)
1789                 rcu_read_unlock();
1790         gi->gl = NULL;
1791 }
1792
1793 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
1794 {
1795         return dump_glock(seq, iter_ptr);
1796 }
1797
1798 static const struct seq_operations gfs2_glock_seq_ops = {
1799         .start = gfs2_glock_seq_start,
1800         .next  = gfs2_glock_seq_next,
1801         .stop  = gfs2_glock_seq_stop,
1802         .show  = gfs2_glock_seq_show,
1803 };
1804
1805 static int gfs2_debugfs_open(struct inode *inode, struct file *file)
1806 {
1807         int ret = seq_open_private(file, &gfs2_glock_seq_ops,
1808                                    sizeof(struct gfs2_glock_iter));
1809         if (ret == 0) {
1810                 struct seq_file *seq = file->private_data;
1811                 struct gfs2_glock_iter *gi = seq->private;
1812                 gi->sdp = inode->i_private;
1813         }
1814         return ret;
1815 }
1816
1817 static const struct file_operations gfs2_debug_fops = {
1818         .owner   = THIS_MODULE,
1819         .open    = gfs2_debugfs_open,
1820         .read    = seq_read,
1821         .llseek  = seq_lseek,
1822         .release = seq_release_private,
1823 };
1824
1825 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
1826 {
1827         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
1828         if (!sdp->debugfs_dir)
1829                 return -ENOMEM;
1830         sdp->debugfs_dentry_glocks = debugfs_create_file("glocks",
1831                                                          S_IFREG | S_IRUGO,
1832                                                          sdp->debugfs_dir, sdp,
1833                                                          &gfs2_debug_fops);
1834         if (!sdp->debugfs_dentry_glocks)
1835                 return -ENOMEM;
1836
1837         return 0;
1838 }
1839
1840 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
1841 {
1842         if (sdp && sdp->debugfs_dir) {
1843                 if (sdp->debugfs_dentry_glocks) {
1844                         debugfs_remove(sdp->debugfs_dentry_glocks);
1845                         sdp->debugfs_dentry_glocks = NULL;
1846                 }
1847                 debugfs_remove(sdp->debugfs_dir);
1848                 sdp->debugfs_dir = NULL;
1849         }
1850 }
1851
1852 int gfs2_register_debugfs(void)
1853 {
1854         gfs2_root = debugfs_create_dir("gfs2", NULL);
1855         return gfs2_root ? 0 : -ENOMEM;
1856 }
1857
1858 void gfs2_unregister_debugfs(void)
1859 {
1860         debugfs_remove(gfs2_root);
1861         gfs2_root = NULL;
1862 }