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 <linux/rwsem.h>
  23 #include <asm/uaccess.h>
  24 #include <linux/seq_file.h>
  25 #include <linux/debugfs.h>
  26 #include <linux/kthread.h>
  27 #include <linux/freezer.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/jiffies.h>
  30
  31 #include "gfs2.h"
  32 #include "incore.h"
  33 #include "glock.h"
  34 #include "glops.h"
  35 #include "inode.h"
  36 #include "lops.h"
  37 #include "meta_io.h"
  38 #include "quota.h"
  39 #include "super.h"
  40 #include "util.h"
  41 #include "bmap.h"
  42
  43 struct gfs2_gl_hash_bucket {
  44         struct hlist_head hb_list;
  45 };
  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 gfs2_dump_lockstate(struct gfs2_sbd *sdp);
  57 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl);
  58 #define GLOCK_BUG_ON(gl,x) do { if (unlikely(x)) { __dump_glock(NULL, gl); BUG(); } } while(0)
  59 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  60
  61 static DECLARE_RWSEM(gfs2_umount_flush_sem);
  62 static struct dentry *gfs2_root;
  63 static struct workqueue_struct *glock_workqueue;
  64 static LIST_HEAD(lru_list);
  65 static atomic_t lru_count = ATOMIC_INIT(0);
  66 static DEFINE_SPINLOCK(lru_lock);
  67
  68 #define GFS2_GL_HASH_SHIFT      15
  69 #define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
  70 #define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
  71
  72 static struct gfs2_gl_hash_bucket gl_hash_table[GFS2_GL_HASH_SIZE];
  73 static struct dentry *gfs2_root;
  74
  75 /*
  76  * Despite what you might think, the numbers below are not arbitrary :-)
  77  * They are taken from the ipv4 routing hash code, which is well tested
  78  * and thus should be nearly optimal. Later on we might tweek the numbers
  79  * but for now this should be fine.
  80  *
  81  * The reason for putting the locks in a separate array from the list heads
  82  * is that we can have fewer locks than list heads and save memory. We use
  83  * the same hash function for both, but with a different hash mask.
  84  */
  85 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \
  86         defined(CONFIG_PROVE_LOCKING)
  87
  88 #ifdef CONFIG_LOCKDEP
  89 # define GL_HASH_LOCK_SZ        256
  90 #else
  91 # if NR_CPUS >= 32
  92 #  define GL_HASH_LOCK_SZ       4096
  93 # elif NR_CPUS >= 16
  94 #  define GL_HASH_LOCK_SZ       2048
  95 # elif NR_CPUS >= 8
  96 #  define GL_HASH_LOCK_SZ       1024
  97 # elif NR_CPUS >= 4
  98 #  define GL_HASH_LOCK_SZ       512
  99 # else
 100 #  define GL_HASH_LOCK_SZ       256
 101 # endif
 102 #endif
 103
 104 /* We never want more locks than chains */
 105 #if GFS2_GL_HASH_SIZE < GL_HASH_LOCK_SZ
 106 # undef GL_HASH_LOCK_SZ
 107 # define GL_HASH_LOCK_SZ GFS2_GL_HASH_SIZE
 108 #endif
 109
 110 static rwlock_t gl_hash_locks[GL_HASH_LOCK_SZ];
 111
 112 static inline rwlock_t *gl_lock_addr(unsigned int x)
 113 {
 114         return &gl_hash_locks[x & (GL_HASH_LOCK_SZ-1)];
 115 }
 116 #else /* not SMP, so no spinlocks required */
 117 static inline rwlock_t *gl_lock_addr(unsigned int x)
 118 {
 119         return NULL;
 120 }
 121 #endif
 122
 123 /**
 124  * gl_hash() - Turn glock number into hash bucket number
 125  * @lock: The glock number
 126  *
 127  * Returns: The number of the corresponding hash bucket
 128  */
 129
 130 static unsigned int gl_hash(const struct gfs2_sbd *sdp,
 131                             const struct lm_lockname *name)
 132 {
 133         unsigned int h;
 134
 135         h = jhash(&name->ln_number, sizeof(u64), 0);
 136         h = jhash(&name->ln_type, sizeof(unsigned int), h);
 137         h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
 138         h &= GFS2_GL_HASH_MASK;
 139
 140         return h;
 141 }
 142
 143 /**
 144  * glock_free() - Perform a few checks and then release struct gfs2_glock
 145  * @gl: The glock to release
 146  *
 147  * Also calls lock module to release its internal structure for this glock.
 148  *
 149  */
 150
 151 static void glock_free(struct gfs2_glock *gl)
 152 {
 153         struct gfs2_sbd *sdp = gl->gl_sbd;
 154         struct inode *aspace = gl->gl_aspace;
 155
 156         if (aspace)
 157                 gfs2_aspace_put(aspace);
 158
 159         sdp->sd_lockstruct.ls_ops->lm_put_lock(gfs2_glock_cachep, gl);
 160 }
 161
 162 /**
 163  * gfs2_glock_hold() - increment reference count on glock
 164  * @gl: The glock to hold
 165  *
 166  */
 167
 168 static void gfs2_glock_hold(struct gfs2_glock *gl)
 169 {
 170         GLOCK_BUG_ON(gl, atomic_read(&gl->gl_ref) == 0);
 171         atomic_inc(&gl->gl_ref);
 172 }
 173
 174 /**
 175  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
 176  * @gl: the glock
 177  *
 178  */
 179
 180 static void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
 181 {
 182         spin_lock(&lru_lock);
 183         if (list_empty(&gl->gl_lru) && gl->gl_state != LM_ST_UNLOCKED) {
 184                 list_add_tail(&gl->gl_lru, &lru_list);
 185                 atomic_inc(&lru_count);
 186         }
 187         spin_unlock(&lru_lock);
 188 }
 189
 190 /**
 191  * gfs2_glock_put() - Decrement reference count on glock
 192  * @gl: The glock to put
 193  *
 194  */
 195
 196 int gfs2_glock_put(struct gfs2_glock *gl)
 197 {
 198         int rv = 0;
 199
 200         write_lock(gl_lock_addr(gl->gl_hash));
 201         if (atomic_dec_and_test(&gl->gl_ref)) {
 202                 hlist_del(&gl->gl_list);
 203                 write_unlock(gl_lock_addr(gl->gl_hash));
 204                 spin_lock(&lru_lock);
 205                 if (!list_empty(&gl->gl_lru)) {
 206                         list_del_init(&gl->gl_lru);
 207                         atomic_dec(&lru_count);
 208                 }
 209                 spin_unlock(&lru_lock);
 210                 GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 211                 glock_free(gl);
 212                 rv = 1;
 213                 goto out;
 214         }
 215         /* 1 for being hashed, 1 for having state != LM_ST_UNLOCKED */
 216         if (atomic_read(&gl->gl_ref) == 2)
 217                 gfs2_glock_schedule_for_reclaim(gl);
 218         write_unlock(gl_lock_addr(gl->gl_hash));
 219 out:
 220         return rv;
 221 }
 222
 223 /**
 224  * search_bucket() - Find struct gfs2_glock by lock number
 225  * @bucket: the bucket to search
 226  * @name: The lock name
 227  *
 228  * Returns: NULL, or the struct gfs2_glock with the requested number
 229  */
 230
 231 static struct gfs2_glock *search_bucket(unsigned int hash,
 232                                         const struct gfs2_sbd *sdp,
 233                                         const struct lm_lockname *name)
 234 {
 235         struct gfs2_glock *gl;
 236         struct hlist_node *h;
 237
 238         hlist_for_each_entry(gl, h, &gl_hash_table[hash].hb_list, gl_list) {
 239                 if (!lm_name_equal(&gl->gl_name, name))
 240                         continue;
 241                 if (gl->gl_sbd != sdp)
 242                         continue;
 243
 244                 atomic_inc(&gl->gl_ref);
 245
 246                 return gl;
 247         }
 248
 249         return NULL;
 250 }
 251
 252 /**
 253  * may_grant - check if its ok to grant a new lock
 254  * @gl: The glock
 255  * @gh: The lock request which we wish to grant
 256  *
 257  * Returns: true if its ok to grant the lock
 258  */
 259
 260 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 261 {
 262         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 263         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 264              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 265                 return 0;
 266         if (gl->gl_state == gh->gh_state)
 267                 return 1;
 268         if (gh->gh_flags & GL_EXACT)
 269                 return 0;
 270         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 271                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 272                         return 1;
 273                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 274                         return 1;
 275         }
 276         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 277                 return 1;
 278         return 0;
 279 }
 280
 281 static void gfs2_holder_wake(struct gfs2_holder *gh)
 282 {
 283         clear_bit(HIF_WAIT, &gh->gh_iflags);
 284         smp_mb__after_clear_bit();
 285         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 286 }
 287
 288 /**
 289  * do_promote - promote as many requests as possible on the current queue
 290  * @gl: The glock
 291  *
 292  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 293  *          if a type specific operation is underway.
 294  */
 295
 296 static int do_promote(struct gfs2_glock *gl)
 297 __releases(&gl->gl_spin)
 298 __acquires(&gl->gl_spin)
 299 {
 300         const struct gfs2_glock_operations *glops = gl->gl_ops;
 301         struct gfs2_holder *gh, *tmp;
 302         int ret;
 303
 304 restart:
 305         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 306                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 307                         continue;
 308                 if (may_grant(gl, gh)) {
 309                         if (gh->gh_list.prev == &gl->gl_holders &&
 310                             glops->go_lock) {
 311                                 spin_unlock(&gl->gl_spin);
 312                                 /* FIXME: eliminate this eventually */
 313                                 ret = glops->go_lock(gh);
 314                                 spin_lock(&gl->gl_spin);
 315                                 if (ret) {
 316                                         if (ret == 1)
 317                                                 return 2;
 318                                         gh->gh_error = ret;
 319                                         list_del_init(&gh->gh_list);
 320                                         gfs2_holder_wake(gh);
 321                                         goto restart;
 322                                 }
 323                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 324                                 gfs2_holder_wake(gh);
 325                                 goto restart;
 326                         }
 327                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 328                         gfs2_holder_wake(gh);
 329                         continue;
 330                 }
 331                 if (gh->gh_list.prev == &gl->gl_holders)
 332                         return 1;
 333                 break;
 334         }
 335         return 0;
 336 }
 337
 338 /**
 339  * do_error - Something unexpected has happened during a lock request
 340  *
 341  */
 342
 343 static inline void do_error(struct gfs2_glock *gl, const int ret)
 344 {
 345         struct gfs2_holder *gh, *tmp;
 346
 347         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 348                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 349                         continue;
 350                 if (ret & LM_OUT_ERROR)
 351                         gh->gh_error = -EIO;
 352                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 353                         gh->gh_error = GLR_TRYFAILED;
 354                 else
 355                         continue;
 356                 list_del_init(&gh->gh_list);
 357                 gfs2_holder_wake(gh);
 358         }
 359 }
 360
 361 /**
 362  * find_first_waiter - find the first gh that's waiting for the glock
 363  * @gl: the glock
 364  */
 365
 366 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 367 {
 368         struct gfs2_holder *gh;
 369
 370         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 371                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 372                         return gh;
 373         }
 374         return NULL;
 375 }
 376
 377 /**
 378  * state_change - record that the glock is now in a different state
 379  * @gl: the glock
 380  * @new_state the new state
 381  *
 382  */
 383
 384 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 385 {
 386         int held1, held2;
 387
 388         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 389         held2 = (new_state != LM_ST_UNLOCKED);
 390
 391         if (held1 != held2) {
 392                 if (held2)
 393                         gfs2_glock_hold(gl);
 394                 else
 395                         gfs2_glock_put(gl);
 396         }
 397
 398         gl->gl_state = new_state;
 399         gl->gl_tchange = jiffies;
 400 }
 401
 402 static void gfs2_demote_wake(struct gfs2_glock *gl)
 403 {
 404         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 405         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 406         smp_mb__after_clear_bit();
 407         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 408 }
 409
 410 /**
 411  * finish_xmote - The DLM has replied to one of our lock requests
 412  * @gl: The glock
 413  * @ret: The status from the DLM
 414  *
 415  */
 416
 417 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 418 {
 419         const struct gfs2_glock_operations *glops = gl->gl_ops;
 420         struct gfs2_holder *gh;
 421         unsigned state = ret & LM_OUT_ST_MASK;
 422         int rv;
 423
 424         spin_lock(&gl->gl_spin);
 425         state_change(gl, state);
 426         gh = find_first_waiter(gl);
 427
 428         /* Demote to UN request arrived during demote to SH or DF */
 429         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 430             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 431                 gl->gl_target = LM_ST_UNLOCKED;
 432
 433         /* Check for state != intended state */
 434         if (unlikely(state != gl->gl_target)) {
 435                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 436                         /* move to back of queue and try next entry */
 437                         if (ret & LM_OUT_CANCELED) {
 438                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 439                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 440                                 gh = find_first_waiter(gl);
 441                                 gl->gl_target = gh->gh_state;
 442                                 goto retry;
 443                         }
 444                         /* Some error or failed "try lock" - report it */
 445                         if ((ret & LM_OUT_ERROR) ||
 446                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 447                                 gl->gl_target = gl->gl_state;
 448                                 do_error(gl, ret);
 449                                 goto out;
 450                         }
 451                 }
 452                 switch(state) {
 453                 /* Unlocked due to conversion deadlock, try again */
 454                 case LM_ST_UNLOCKED:
 455 retry:
 456                         do_xmote(gl, gh, gl->gl_target);
 457                         break;
 458                 /* Conversion fails, unlock and try again */
 459                 case LM_ST_SHARED:
 460                 case LM_ST_DEFERRED:
 461                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 462                         break;
 463                 default: /* Everything else */
 464                         printk(KERN_ERR "GFS2: wanted %u got %u\n", gl->gl_target, state);
 465                         GLOCK_BUG_ON(gl, 1);
 466                 }
 467                 spin_unlock(&gl->gl_spin);
 468                 gfs2_glock_put(gl);
 469                 return;
 470         }
 471
 472         /* Fast path - we got what we asked for */
 473         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 474                 gfs2_demote_wake(gl);
 475         if (state != LM_ST_UNLOCKED) {
 476                 if (glops->go_xmote_bh) {
 477                         spin_unlock(&gl->gl_spin);
 478                         rv = glops->go_xmote_bh(gl, gh);
 479                         if (rv == -EAGAIN)
 480                                 return;
 481                         spin_lock(&gl->gl_spin);
 482                         if (rv) {
 483                                 do_error(gl, rv);
 484                                 goto out;
 485                         }
 486                 }
 487                 rv = do_promote(gl);
 488                 if (rv == 2)
 489                         goto out_locked;
 490         }
 491 out:
 492         clear_bit(GLF_LOCK, &gl->gl_flags);
 493 out_locked:
 494         spin_unlock(&gl->gl_spin);
 495         gfs2_glock_put(gl);
 496 }
 497
 498 static unsigned int gfs2_lm_lock(struct gfs2_sbd *sdp, void *lock,
 499                                  unsigned int req_state,
 500                                  unsigned int flags)
 501 {
 502         int ret = LM_OUT_ERROR;
 503
 504         if (!sdp->sd_lockstruct.ls_ops->lm_lock)
 505                 return req_state == LM_ST_UNLOCKED ? 0 : req_state;
 506
 507         if (likely(!test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 508                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(lock,
 509                                                          req_state, flags);
 510         return ret;
 511 }
 512
 513 /**
 514  * do_xmote - Calls the DLM to change the state of a lock
 515  * @gl: The lock state
 516  * @gh: The holder (only for promotes)
 517  * @target: The target lock state
 518  *
 519  */
 520
 521 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 522 __releases(&gl->gl_spin)
 523 __acquires(&gl->gl_spin)
 524 {
 525         const struct gfs2_glock_operations *glops = gl->gl_ops;
 526         struct gfs2_sbd *sdp = gl->gl_sbd;
 527         unsigned int lck_flags = gh ? gh->gh_flags : 0;
 528         int ret;
 529
 530         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 531                       LM_FLAG_PRIORITY);
 532         BUG_ON(gl->gl_state == target);
 533         BUG_ON(gl->gl_state == gl->gl_target);
 534         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 535             glops->go_inval) {
 536                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 537                 do_error(gl, 0); /* Fail queued try locks */
 538         }
 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         ret = gfs2_lm_lock(sdp, gl, target, lck_flags);
 552
 553         if (!(ret & LM_OUT_ASYNC)) {
 554                 finish_xmote(gl, ret);
 555                 gfs2_glock_hold(gl);
 556                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
 557                         gfs2_glock_put(gl);
 558         } else {
 559                 GLOCK_BUG_ON(gl, ret != LM_OUT_ASYNC);
 560         }
 561         spin_lock(&gl->gl_spin);
 562 }
 563
 564 /**
 565  * find_first_holder - find the first "holder" gh
 566  * @gl: the glock
 567  */
 568
 569 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 570 {
 571         struct gfs2_holder *gh;
 572
 573         if (!list_empty(&gl->gl_holders)) {
 574                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 575                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 576                         return gh;
 577         }
 578         return NULL;
 579 }
 580
 581 /**
 582  * run_queue - do all outstanding tasks related to a glock
 583  * @gl: The glock in question
 584  * @nonblock: True if we must not block in run_queue
 585  *
 586  */
 587
 588 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 589 __releases(&gl->gl_spin)
 590 __acquires(&gl->gl_spin)
 591 {
 592         struct gfs2_holder *gh = NULL;
 593         int ret;
 594
 595         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 596                 return;
 597
 598         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 599
 600         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 601             gl->gl_demote_state != gl->gl_state) {
 602                 if (find_first_holder(gl))
 603                         goto out_unlock;
 604                 if (nonblock)
 605                         goto out_sched;
 606                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 607                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 608                 gl->gl_target = gl->gl_demote_state;
 609         } else {
 610                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 611                         gfs2_demote_wake(gl);
 612                 ret = do_promote(gl);
 613                 if (ret == 0)
 614                         goto out_unlock;
 615                 if (ret == 2)
 616                         goto out;
 617                 gh = find_first_waiter(gl);
 618                 gl->gl_target = gh->gh_state;
 619                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 620                         do_error(gl, 0); /* Fail queued try locks */
 621         }
 622         do_xmote(gl, gh, gl->gl_target);
 623 out:
 624         return;
 625
 626 out_sched:
 627         gfs2_glock_hold(gl);
 628         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
 629                 gfs2_glock_put(gl);
 630 out_unlock:
 631         clear_bit(GLF_LOCK, &gl->gl_flags);
 632         goto out;
 633 }
 634
 635 static void glock_work_func(struct work_struct *work)
 636 {
 637         unsigned long delay = 0;
 638         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 639
 640         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags))
 641                 finish_xmote(gl, gl->gl_reply);
 642         down_read(&gfs2_umount_flush_sem);
 643         spin_lock(&gl->gl_spin);
 644         if (test_and_clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 645             gl->gl_state != LM_ST_UNLOCKED &&
 646             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 647                 unsigned long holdtime, now = jiffies;
 648                 holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
 649                 if (time_before(now, holdtime))
 650                         delay = holdtime - now;
 651                 set_bit(delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE, &gl->gl_flags);
 652         }
 653         run_queue(gl, 0);
 654         spin_unlock(&gl->gl_spin);
 655         up_read(&gfs2_umount_flush_sem);
 656         if (!delay ||
 657             queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
 658                 gfs2_glock_put(gl);
 659 }
 660
 661 /**
 662  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 663  * @sdp: The GFS2 superblock
 664  * @number: the lock number
 665  * @glops: The glock_operations to use
 666  * @create: If 0, don't create the glock if it doesn't exist
 667  * @glp: the glock is returned here
 668  *
 669  * This does not lock a glock, just finds/creates structures for one.
 670  *
 671  * Returns: errno
 672  */
 673
 674 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 675                    const struct gfs2_glock_operations *glops, int create,
 676                    struct gfs2_glock **glp)
 677 {
 678         struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
 679         struct gfs2_glock *gl, *tmp;
 680         unsigned int hash = gl_hash(sdp, &name);
 681         int error;
 682
 683         read_lock(gl_lock_addr(hash));
 684         gl = search_bucket(hash, sdp, &name);
 685         read_unlock(gl_lock_addr(hash));
 686
 687         *glp = gl;
 688         if (gl)
 689                 return 0;
 690         if (!create)
 691                 return -ENOENT;
 692
 693         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
 694         if (!gl)
 695                 return -ENOMEM;
 696
 697         gl->gl_flags = 0;
 698         gl->gl_name = name;
 699         atomic_set(&gl->gl_ref, 1);
 700         gl->gl_state = LM_ST_UNLOCKED;
 701         gl->gl_target = LM_ST_UNLOCKED;
 702         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 703         gl->gl_hash = hash;
 704         gl->gl_ops = glops;
 705         snprintf(gl->gl_strname, GDLM_STRNAME_BYTES, "%8x%16llx", name.ln_type, (unsigned long long)number);
 706         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 707         gl->gl_lksb.sb_lvbptr = gl->gl_lvb;
 708         gl->gl_tchange = jiffies;
 709         gl->gl_object = NULL;
 710         gl->gl_sbd = sdp;
 711         gl->gl_aspace = NULL;
 712         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 713
 714         /* If this glock protects actual on-disk data or metadata blocks,
 715            create a VFS inode to manage the pages/buffers holding them. */
 716         if (glops == &gfs2_inode_glops || glops == &gfs2_rgrp_glops) {
 717                 gl->gl_aspace = gfs2_aspace_get(sdp);
 718                 if (!gl->gl_aspace) {
 719                         error = -ENOMEM;
 720                         goto fail;
 721                 }
 722         }
 723
 724         write_lock(gl_lock_addr(hash));
 725         tmp = search_bucket(hash, sdp, &name);
 726         if (tmp) {
 727                 write_unlock(gl_lock_addr(hash));
 728                 glock_free(gl);
 729                 gl = tmp;
 730         } else {
 731                 hlist_add_head(&gl->gl_list, &gl_hash_table[hash].hb_list);
 732                 write_unlock(gl_lock_addr(hash));
 733         }
 734
 735         *glp = gl;
 736
 737         return 0;
 738
 739 fail:
 740         kmem_cache_free(gfs2_glock_cachep, gl);
 741         return error;
 742 }
 743
 744 /**
 745  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 746  * @gl: the glock
 747  * @state: the state we're requesting
 748  * @flags: the modifier flags
 749  * @gh: the holder structure
 750  *
 751  */
 752
 753 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
 754                       struct gfs2_holder *gh)
 755 {
 756         INIT_LIST_HEAD(&gh->gh_list);
 757         gh->gh_gl = gl;
 758         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 759         gh->gh_owner_pid = get_pid(task_pid(current));
 760         gh->gh_state = state;
 761         gh->gh_flags = flags;
 762         gh->gh_error = 0;
 763         gh->gh_iflags = 0;
 764         gfs2_glock_hold(gl);
 765 }
 766
 767 /**
 768  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 769  * @state: the state we're requesting
 770  * @flags: the modifier flags
 771  * @gh: the holder structure
 772  *
 773  * Don't mess with the glock.
 774  *
 775  */
 776
 777 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
 778 {
 779         gh->gh_state = state;
 780         gh->gh_flags = flags;
 781         gh->gh_iflags = 0;
 782         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 783 }
 784
 785 /**
 786  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 787  * @gh: the holder structure
 788  *
 789  */
 790
 791 void gfs2_holder_uninit(struct gfs2_holder *gh)
 792 {
 793         put_pid(gh->gh_owner_pid);
 794         gfs2_glock_put(gh->gh_gl);
 795         gh->gh_gl = NULL;
 796         gh->gh_ip = 0;
 797 }
 798
 799 static int just_schedule(void *word)
 800 {
 801         schedule();
 802         return 0;
 803 }
 804
 805 static void wait_on_holder(struct gfs2_holder *gh)
 806 {
 807         might_sleep();
 808         wait_on_bit(&gh->gh_iflags, HIF_WAIT, just_schedule, TASK_UNINTERRUPTIBLE);
 809 }
 810
 811 static void wait_on_demote(struct gfs2_glock *gl)
 812 {
 813         might_sleep();
 814         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, just_schedule, TASK_UNINTERRUPTIBLE);
 815 }
 816
 817 /**
 818  * handle_callback - process a demote request
 819  * @gl: the glock
 820  * @state: the state the caller wants us to change to
 821  *
 822  * There are only two requests that we are going to see in actual
 823  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 824  */
 825
 826 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 827                             unsigned long delay)
 828 {
 829         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 830
 831         set_bit(bit, &gl->gl_flags);
 832         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 833                 gl->gl_demote_state = state;
 834                 gl->gl_demote_time = jiffies;
 835         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 836                         gl->gl_demote_state != state) {
 837                 gl->gl_demote_state = LM_ST_UNLOCKED;
 838         }
 839 }
 840
 841 /**
 842  * gfs2_glock_wait - wait on a glock acquisition
 843  * @gh: the glock holder
 844  *
 845  * Returns: 0 on success
 846  */
 847
 848 int gfs2_glock_wait(struct gfs2_holder *gh)
 849 {
 850         wait_on_holder(gh);
 851         return gh->gh_error;
 852 }
 853
 854 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 855 {
 856         va_list args;
 857
 858         va_start(args, fmt);
 859         if (seq) {
 860                 struct gfs2_glock_iter *gi = seq->private;
 861                 vsprintf(gi->string, fmt, args);
 862                 seq_printf(seq, gi->string);
 863         } else {
 864                 printk(KERN_ERR " ");
 865                 vprintk(fmt, args);
 866         }
 867         va_end(args);
 868 }
 869
 870 /**
 871  * add_to_queue - Add a holder to the wait queue (but look for recursion)
 872  * @gh: the holder structure to add
 873  *
 874  * Eventually we should move the recursive locking trap to a
 875  * debugging option or something like that. This is the fast
 876  * path and needs to have the minimum number of distractions.
 877  *
 878  */
 879
 880 static inline void add_to_queue(struct gfs2_holder *gh)
 881 __releases(&gl->gl_spin)
 882 __acquires(&gl->gl_spin)
 883 {
 884         struct gfs2_glock *gl = gh->gh_gl;
 885         struct gfs2_sbd *sdp = gl->gl_sbd;
 886         struct list_head *insert_pt = NULL;
 887         struct gfs2_holder *gh2;
 888         int try_lock = 0;
 889
 890         BUG_ON(gh->gh_owner_pid == NULL);
 891         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
 892                 BUG();
 893
 894         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 895                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 896                         try_lock = 1;
 897                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 898                         goto fail;
 899         }
 900
 901         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
 902                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
 903                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
 904                         goto trap_recursive;
 905                 if (try_lock &&
 906                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) &&
 907                     !may_grant(gl, gh)) {
 908 fail:
 909                         gh->gh_error = GLR_TRYFAILED;
 910                         gfs2_holder_wake(gh);
 911                         return;
 912                 }
 913                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
 914                         continue;
 915                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
 916                         insert_pt = &gh2->gh_list;
 917         }
 918         if (likely(insert_pt == NULL)) {
 919                 list_add_tail(&gh->gh_list, &gl->gl_holders);
 920                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
 921                         goto do_cancel;
 922                 return;
 923         }
 924         list_add_tail(&gh->gh_list, insert_pt);
 925 do_cancel:
 926         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 927         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
 928                 spin_unlock(&gl->gl_spin);
 929                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
 930                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
 931                 spin_lock(&gl->gl_spin);
 932         }
 933         return;
 934
 935 trap_recursive:
 936         print_symbol(KERN_ERR "original: %s\n", gh2->gh_ip);
 937         printk(KERN_ERR "pid: %d\n", pid_nr(gh2->gh_owner_pid));
 938         printk(KERN_ERR "lock type: %d req lock state : %d\n",
 939                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
 940         print_symbol(KERN_ERR "new: %s\n", gh->gh_ip);
 941         printk(KERN_ERR "pid: %d\n", pid_nr(gh->gh_owner_pid));
 942         printk(KERN_ERR "lock type: %d req lock state : %d\n",
 943                gh->gh_gl->gl_name.ln_type, gh->gh_state);
 944         __dump_glock(NULL, gl);
 945         BUG();
 946 }
 947
 948 /**
 949  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
 950  * @gh: the holder structure
 951  *
 952  * if (gh->gh_flags & GL_ASYNC), this never returns an error
 953  *
 954  * Returns: 0, GLR_TRYFAILED, or errno on failure
 955  */
 956
 957 int gfs2_glock_nq(struct gfs2_holder *gh)
 958 {
 959         struct gfs2_glock *gl = gh->gh_gl;
 960         struct gfs2_sbd *sdp = gl->gl_sbd;
 961         int error = 0;
 962
 963         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 964                 return -EIO;
 965
 966         spin_lock(&gl->gl_spin);
 967         add_to_queue(gh);
 968         run_queue(gl, 1);
 969         spin_unlock(&gl->gl_spin);
 970
 971         if (!(gh->gh_flags & GL_ASYNC))
 972                 error = gfs2_glock_wait(gh);
 973
 974         return error;
 975 }
 976
 977 /**
 978  * gfs2_glock_poll - poll to see if an async request has been completed
 979  * @gh: the holder
 980  *
 981  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
 982  */
 983
 984 int gfs2_glock_poll(struct gfs2_holder *gh)
 985 {
 986         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
 987 }
 988
 989 /**
 990  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
 991  * @gh: the glock holder
 992  *
 993  */
 994
 995 void gfs2_glock_dq(struct gfs2_holder *gh)
 996 {
 997         struct gfs2_glock *gl = gh->gh_gl;
 998         const struct gfs2_glock_operations *glops = gl->gl_ops;
 999         unsigned delay = 0;
1000         int fast_path = 0;
1001
1002         spin_lock(&gl->gl_spin);
1003         if (gh->gh_flags & GL_NOCACHE)
1004                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1005
1006         list_del_init(&gh->gh_list);
1007         if (find_first_holder(gl) == NULL) {
1008                 if (glops->go_unlock) {
1009                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1010                         spin_unlock(&gl->gl_spin);
1011                         glops->go_unlock(gh);
1012                         spin_lock(&gl->gl_spin);
1013                         clear_bit(GLF_LOCK, &gl->gl_flags);
1014                 }
1015                 if (list_empty(&gl->gl_holders) &&
1016                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1017                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1018                         fast_path = 1;
1019         }
1020         spin_unlock(&gl->gl_spin);
1021         if (likely(fast_path))
1022                 return;
1023
1024         gfs2_glock_hold(gl);
1025         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1026             !test_bit(GLF_DEMOTE, &gl->gl_flags))
1027                 delay = gl->gl_ops->go_min_hold_time;
1028         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1029                 gfs2_glock_put(gl);
1030 }
1031
1032 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1033 {
1034         struct gfs2_glock *gl = gh->gh_gl;
1035         gfs2_glock_dq(gh);
1036         wait_on_demote(gl);
1037 }
1038
1039 /**
1040  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1041  * @gh: the holder structure
1042  *
1043  */
1044
1045 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1046 {
1047         gfs2_glock_dq(gh);
1048         gfs2_holder_uninit(gh);
1049 }
1050
1051 /**
1052  * gfs2_glock_nq_num - acquire a glock based on lock number
1053  * @sdp: the filesystem
1054  * @number: the lock number
1055  * @glops: the glock operations for the type of glock
1056  * @state: the state to acquire the glock in
1057  * @flags: modifier flags for the aquisition
1058  * @gh: the struct gfs2_holder
1059  *
1060  * Returns: errno
1061  */
1062
1063 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1064                       const struct gfs2_glock_operations *glops,
1065                       unsigned int state, int flags, struct gfs2_holder *gh)
1066 {
1067         struct gfs2_glock *gl;
1068         int error;
1069
1070         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1071         if (!error) {
1072                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1073                 gfs2_glock_put(gl);
1074         }
1075
1076         return error;
1077 }
1078
1079 /**
1080  * glock_compare - Compare two struct gfs2_glock structures for sorting
1081  * @arg_a: the first structure
1082  * @arg_b: the second structure
1083  *
1084  */
1085
1086 static int glock_compare(const void *arg_a, const void *arg_b)
1087 {
1088         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1089         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1090         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1091         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1092
1093         if (a->ln_number > b->ln_number)
1094                 return 1;
1095         if (a->ln_number < b->ln_number)
1096                 return -1;
1097         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1098         return 0;
1099 }
1100
1101 /**
1102  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1103  * @num_gh: the number of structures
1104  * @ghs: an array of struct gfs2_holder structures
1105  *
1106  * Returns: 0 on success (all glocks acquired),
1107  *          errno on failure (no glocks acquired)
1108  */
1109
1110 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1111                      struct gfs2_holder **p)
1112 {
1113         unsigned int x;
1114         int error = 0;
1115
1116         for (x = 0; x < num_gh; x++)
1117                 p[x] = &ghs[x];
1118
1119         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1120
1121         for (x = 0; x < num_gh; x++) {
1122                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1123
1124                 error = gfs2_glock_nq(p[x]);
1125                 if (error) {
1126                         while (x--)
1127                                 gfs2_glock_dq(p[x]);
1128                         break;
1129                 }
1130         }
1131
1132         return error;
1133 }
1134
1135 /**
1136  * gfs2_glock_nq_m - acquire multiple glocks
1137  * @num_gh: the number of structures
1138  * @ghs: an array of struct gfs2_holder structures
1139  *
1140  *
1141  * Returns: 0 on success (all glocks acquired),
1142  *          errno on failure (no glocks acquired)
1143  */
1144
1145 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1146 {
1147         struct gfs2_holder *tmp[4];
1148         struct gfs2_holder **pph = tmp;
1149         int error = 0;
1150
1151         switch(num_gh) {
1152         case 0:
1153                 return 0;
1154         case 1:
1155                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1156                 return gfs2_glock_nq(ghs);
1157         default:
1158                 if (num_gh <= 4)
1159                         break;
1160                 pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
1161                 if (!pph)
1162                         return -ENOMEM;
1163         }
1164
1165         error = nq_m_sync(num_gh, ghs, pph);
1166
1167         if (pph != tmp)
1168                 kfree(pph);
1169
1170         return error;
1171 }
1172
1173 /**
1174  * gfs2_glock_dq_m - release multiple glocks
1175  * @num_gh: the number of structures
1176  * @ghs: an array of struct gfs2_holder structures
1177  *
1178  */
1179
1180 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1181 {
1182         unsigned int x;
1183
1184         for (x = 0; x < num_gh; x++)
1185                 gfs2_glock_dq(&ghs[x]);
1186 }
1187
1188 /**
1189  * gfs2_glock_dq_uninit_m - release multiple glocks
1190  * @num_gh: the number of structures
1191  * @ghs: an array of struct gfs2_holder structures
1192  *
1193  */
1194
1195 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1196 {
1197         unsigned int x;
1198
1199         for (x = 0; x < num_gh; x++)
1200                 gfs2_glock_dq_uninit(&ghs[x]);
1201 }
1202
1203 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1204 {
1205         unsigned long delay = 0;
1206         unsigned long holdtime;
1207         unsigned long now = jiffies;
1208
1209         gfs2_glock_hold(gl);
1210         holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
1211         if (time_before(now, holdtime))
1212                 delay = holdtime - now;
1213         if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1214                 delay = gl->gl_ops->go_min_hold_time;
1215
1216         spin_lock(&gl->gl_spin);
1217         handle_callback(gl, state, delay);
1218         spin_unlock(&gl->gl_spin);
1219         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1220                 gfs2_glock_put(gl);
1221 }
1222
1223 /**
1224  * gfs2_glock_complete - Callback used by locking
1225  * @gl: Pointer to the glock
1226  * @ret: The return value from the dlm
1227  *
1228  */
1229
1230 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1231 {
1232         struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
1233         gl->gl_reply = ret;
1234         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_flags))) {
1235                 struct gfs2_holder *gh;
1236                 spin_lock(&gl->gl_spin);
1237                 gh = find_first_waiter(gl);
1238                 if ((!(gh && (gh->gh_flags & LM_FLAG_NOEXP)) &&
1239                      (gl->gl_target != LM_ST_UNLOCKED)) ||
1240                     ((ret & ~LM_OUT_ST_MASK) != 0))
1241                         set_bit(GLF_FROZEN, &gl->gl_flags);
1242                 spin_unlock(&gl->gl_spin);
1243                 if (test_bit(GLF_FROZEN, &gl->gl_flags))
1244                         return;
1245         }
1246         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1247         gfs2_glock_hold(gl);
1248         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1249                 gfs2_glock_put(gl);
1250 }
1251
1252 /**
1253  * demote_ok - Check to see if it's ok to unlock a glock
1254  * @gl: the glock
1255  *
1256  * Returns: 1 if it's ok
1257  */
1258
1259 static int demote_ok(const struct gfs2_glock *gl)
1260 {
1261         const struct gfs2_glock_operations *glops = gl->gl_ops;
1262
1263         if (gl->gl_state == LM_ST_UNLOCKED)
1264                 return 0;
1265         if (!list_empty(&gl->gl_holders))
1266                 return 0;
1267         if (glops->go_demote_ok)
1268                 return glops->go_demote_ok(gl);
1269         return 1;
1270 }
1271
1272
1273 static int gfs2_shrink_glock_memory(int nr, gfp_t gfp_mask)
1274 {
1275         struct gfs2_glock *gl;
1276         int may_demote;
1277         int nr_skipped = 0;
1278         int got_ref = 0;
1279         LIST_HEAD(skipped);
1280
1281         if (nr == 0)
1282                 goto out;
1283
1284         if (!(gfp_mask & __GFP_FS))
1285                 return -1;
1286
1287         spin_lock(&lru_lock);
1288         while(nr && !list_empty(&lru_list)) {
1289                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1290                 list_del_init(&gl->gl_lru);
1291                 atomic_dec(&lru_count);
1292
1293                 /* Test for being demotable */
1294                 if (!test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1295                         gfs2_glock_hold(gl);
1296                         got_ref = 1;
1297                         spin_unlock(&lru_lock);
1298                         spin_lock(&gl->gl_spin);
1299                         may_demote = demote_ok(gl);
1300                         spin_unlock(&gl->gl_spin);
1301                         clear_bit(GLF_LOCK, &gl->gl_flags);
1302                         if (may_demote) {
1303                                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1304                                 nr--;
1305                                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1306                                         gfs2_glock_put(gl);
1307                         }
1308                         spin_lock(&lru_lock);
1309                         if (may_demote)
1310                                 continue;
1311                 }
1312                 if (list_empty(&gl->gl_lru) &&
1313                     (atomic_read(&gl->gl_ref) <= (2 + got_ref))) {
1314                         nr_skipped++;
1315                         list_add(&gl->gl_lru, &skipped);
1316                 }
1317                 if (got_ref) {
1318                         spin_unlock(&lru_lock);
1319                         gfs2_glock_put(gl);
1320                         spin_lock(&lru_lock);
1321                         got_ref = 0;
1322                 }
1323         }
1324         list_splice(&skipped, &lru_list);
1325         atomic_add(nr_skipped, &lru_count);
1326         spin_unlock(&lru_lock);
1327 out:
1328         return (atomic_read(&lru_count) / 100) * sysctl_vfs_cache_pressure;
1329 }
1330
1331 static struct shrinker glock_shrinker = {
1332         .shrink = gfs2_shrink_glock_memory,
1333         .seeks = DEFAULT_SEEKS,
1334 };
1335
1336 /**
1337  * examine_bucket - Call a function for glock in a hash bucket
1338  * @examiner: the function
1339  * @sdp: the filesystem
1340  * @bucket: the bucket
1341  *
1342  * Returns: 1 if the bucket has entries
1343  */
1344
1345 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
1346                           unsigned int hash)
1347 {
1348         struct gfs2_glock *gl, *prev = NULL;
1349         int has_entries = 0;
1350         struct hlist_head *head = &gl_hash_table[hash].hb_list;
1351
1352         read_lock(gl_lock_addr(hash));
1353         /* Can't use hlist_for_each_entry - don't want prefetch here */
1354         if (hlist_empty(head))
1355                 goto out;
1356         gl = list_entry(head->first, struct gfs2_glock, gl_list);
1357         while(1) {
1358                 if (!sdp || gl->gl_sbd == sdp) {
1359                         gfs2_glock_hold(gl);
1360                         read_unlock(gl_lock_addr(hash));
1361                         if (prev)
1362                                 gfs2_glock_put(prev);
1363                         prev = gl;
1364                         examiner(gl);
1365                         has_entries = 1;
1366                         read_lock(gl_lock_addr(hash));
1367                 }
1368                 if (gl->gl_list.next == NULL)
1369                         break;
1370                 gl = list_entry(gl->gl_list.next, struct gfs2_glock, gl_list);
1371         }
1372 out:
1373         read_unlock(gl_lock_addr(hash));
1374         if (prev)
1375                 gfs2_glock_put(prev);
1376         cond_resched();
1377         return has_entries;
1378 }
1379
1380
1381 /**
1382  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1383  * @gl: The glock to thaw
1384  *
1385  * N.B. When we freeze a glock, we leave a ref to the glock outstanding,
1386  * so this has to result in the ref count being dropped by one.
1387  */
1388
1389 static void thaw_glock(struct gfs2_glock *gl)
1390 {
1391         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1392                 return;
1393         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1394         gfs2_glock_hold(gl);
1395         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1396                 gfs2_glock_put(gl);
1397 }
1398
1399 /**
1400  * clear_glock - look at a glock and see if we can free it from glock cache
1401  * @gl: the glock to look at
1402  *
1403  */
1404
1405 static void clear_glock(struct gfs2_glock *gl)
1406 {
1407         spin_lock(&lru_lock);
1408         if (!list_empty(&gl->gl_lru)) {
1409                 list_del_init(&gl->gl_lru);
1410                 atomic_dec(&lru_count);
1411         }
1412         spin_unlock(&lru_lock);
1413
1414         spin_lock(&gl->gl_spin);
1415         if (find_first_holder(gl) == NULL && gl->gl_state != LM_ST_UNLOCKED)
1416                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1417         spin_unlock(&gl->gl_spin);
1418         gfs2_glock_hold(gl);
1419         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1420                 gfs2_glock_put(gl);
1421 }
1422
1423 /**
1424  * gfs2_glock_thaw - Thaw any frozen glocks
1425  * @sdp: The super block
1426  *
1427  */
1428
1429 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1430 {
1431         unsigned x;
1432
1433         for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
1434                 examine_bucket(thaw_glock, sdp, x);
1435 }
1436
1437 /**
1438  * gfs2_gl_hash_clear - Empty out the glock hash table
1439  * @sdp: the filesystem
1440  * @wait: wait until it's all gone
1441  *
1442  * Called when unmounting the filesystem.
1443  */
1444
1445 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1446 {
1447         unsigned long t;
1448         unsigned int x;
1449         int cont;
1450
1451         t = jiffies;
1452
1453         for (;;) {
1454                 cont = 0;
1455                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
1456                         if (examine_bucket(clear_glock, sdp, x))
1457                                 cont = 1;
1458                 }
1459
1460                 if (!cont)
1461                         break;
1462
1463                 if (time_after_eq(jiffies,
1464                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
1465                         fs_warn(sdp, "Unmount seems to be stalled. "
1466                                      "Dumping lock state...\n");
1467                         gfs2_dump_lockstate(sdp);
1468                         t = jiffies;
1469                 }
1470
1471                 down_write(&gfs2_umount_flush_sem);
1472                 invalidate_inodes(sdp->sd_vfs);
1473                 up_write(&gfs2_umount_flush_sem);
1474                 msleep(10);
1475         }
1476 }
1477
1478 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1479 {
1480         struct gfs2_glock *gl = ip->i_gl;
1481         int ret;
1482
1483         ret = gfs2_truncatei_resume(ip);
1484         gfs2_assert_withdraw(gl->gl_sbd, ret == 0);
1485
1486         spin_lock(&gl->gl_spin);
1487         clear_bit(GLF_LOCK, &gl->gl_flags);
1488         run_queue(gl, 1);
1489         spin_unlock(&gl->gl_spin);
1490 }
1491
1492 static const char *state2str(unsigned state)
1493 {
1494         switch(state) {
1495         case LM_ST_UNLOCKED:
1496                 return "UN";
1497         case LM_ST_SHARED:
1498                 return "SH";
1499         case LM_ST_DEFERRED:
1500                 return "DF";
1501         case LM_ST_EXCLUSIVE:
1502                 return "EX";
1503         }
1504         return "??";
1505 }
1506
1507 static const char *hflags2str(char *buf, unsigned flags, unsigned long iflags)
1508 {
1509         char *p = buf;
1510         if (flags & LM_FLAG_TRY)
1511                 *p++ = 't';
1512         if (flags & LM_FLAG_TRY_1CB)
1513                 *p++ = 'T';
1514         if (flags & LM_FLAG_NOEXP)
1515                 *p++ = 'e';
1516         if (flags & LM_FLAG_ANY)
1517                 *p++ = 'A';
1518         if (flags & LM_FLAG_PRIORITY)
1519                 *p++ = 'p';
1520         if (flags & GL_ASYNC)
1521                 *p++ = 'a';
1522         if (flags & GL_EXACT)
1523                 *p++ = 'E';
1524         if (flags & GL_NOCACHE)
1525                 *p++ = 'c';
1526         if (test_bit(HIF_HOLDER, &iflags))
1527                 *p++ = 'H';
1528         if (test_bit(HIF_WAIT, &iflags))
1529                 *p++ = 'W';
1530         if (test_bit(HIF_FIRST, &iflags))
1531                 *p++ = 'F';
1532         *p = 0;
1533         return buf;
1534 }
1535
1536 /**
1537  * dump_holder - print information about a glock holder
1538  * @seq: the seq_file struct
1539  * @gh: the glock holder
1540  *
1541  * Returns: 0 on success, -ENOBUFS when we run out of space
1542  */
1543
1544 static int dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1545 {
1546         struct task_struct *gh_owner = NULL;
1547         char buffer[KSYM_SYMBOL_LEN];
1548         char flags_buf[32];
1549
1550         sprint_symbol(buffer, gh->gh_ip);
1551         if (gh->gh_owner_pid)
1552                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1553         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %s\n",
1554                   state2str(gh->gh_state),
1555                   hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1556                   gh->gh_error,
1557                   gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1558                   gh_owner ? gh_owner->comm : "(ended)", buffer);
1559         return 0;
1560 }
1561
1562 static const char *gflags2str(char *buf, const unsigned long *gflags)
1563 {
1564         char *p = buf;
1565         if (test_bit(GLF_LOCK, gflags))
1566                 *p++ = 'l';
1567         if (test_bit(GLF_DEMOTE, gflags))
1568                 *p++ = 'D';
1569         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1570                 *p++ = 'd';
1571         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1572                 *p++ = 'p';
1573         if (test_bit(GLF_DIRTY, gflags))
1574                 *p++ = 'y';
1575         if (test_bit(GLF_LFLUSH, gflags))
1576                 *p++ = 'f';
1577         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1578                 *p++ = 'i';
1579         if (test_bit(GLF_REPLY_PENDING, gflags))
1580                 *p++ = 'r';
1581         if (test_bit(GLF_INITIAL, gflags))
1582                 *p++ = 'I';
1583         if (test_bit(GLF_FROZEN, gflags))
1584                 *p++ = 'F';
1585         *p = 0;
1586         return buf;
1587 }
1588
1589 /**
1590  * __dump_glock - print information about a glock
1591  * @seq: The seq_file struct
1592  * @gl: the glock
1593  *
1594  * The file format is as follows:
1595  * One line per object, capital letters are used to indicate objects
1596  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1597  * other objects are indented by a single space and follow the glock to
1598  * which they are related. Fields are indicated by lower case letters
1599  * followed by a colon and the field value, except for strings which are in
1600  * [] so that its possible to see if they are composed of spaces for
1601  * example. The field's are n = number (id of the object), f = flags,
1602  * t = type, s = state, r = refcount, e = error, p = pid.
1603  *
1604  * Returns: 0 on success, -ENOBUFS when we run out of space
1605  */
1606
1607 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1608 {
1609         const struct gfs2_glock_operations *glops = gl->gl_ops;
1610         unsigned long long dtime;
1611         const struct gfs2_holder *gh;
1612         char gflags_buf[32];
1613         int error = 0;
1614
1615         dtime = jiffies - gl->gl_demote_time;
1616         dtime *= 1000000/HZ; /* demote time in uSec */
1617         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1618                 dtime = 0;
1619         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llu f:%s t:%s d:%s/%llu a:%d r:%d\n",
1620                   state2str(gl->gl_state),
1621                   gl->gl_name.ln_type,
1622                   (unsigned long long)gl->gl_name.ln_number,
1623                   gflags2str(gflags_buf, &gl->gl_flags),
1624                   state2str(gl->gl_target),
1625                   state2str(gl->gl_demote_state), dtime,
1626                   atomic_read(&gl->gl_ail_count),
1627                   atomic_read(&gl->gl_ref));
1628
1629         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1630                 error = dump_holder(seq, gh);
1631                 if (error)
1632                         goto out;
1633         }
1634         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1635                 error = glops->go_dump(seq, gl);
1636 out:
1637         return error;
1638 }
1639
1640 static int dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1641 {
1642         int ret;
1643         spin_lock(&gl->gl_spin);
1644         ret = __dump_glock(seq, gl);
1645         spin_unlock(&gl->gl_spin);
1646         return ret;
1647 }
1648
1649 /**
1650  * gfs2_dump_lockstate - print out the current lockstate
1651  * @sdp: the filesystem
1652  * @ub: the buffer to copy the information into
1653  *
1654  * If @ub is NULL, dump the lockstate to the console.
1655  *
1656  */
1657
1658 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
1659 {
1660         struct gfs2_glock *gl;
1661         struct hlist_node *h;
1662         unsigned int x;
1663         int error = 0;
1664
1665         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
1666
1667                 read_lock(gl_lock_addr(x));
1668
1669                 hlist_for_each_entry(gl, h, &gl_hash_table[x].hb_list, gl_list) {
1670                         if (gl->gl_sbd != sdp)
1671                                 continue;
1672
1673                         error = dump_glock(NULL, gl);
1674                         if (error)
1675                                 break;
1676                 }
1677
1678                 read_unlock(gl_lock_addr(x));
1679
1680                 if (error)
1681                         break;
1682         }
1683
1684
1685         return error;
1686 }
1687
1688
1689 int __init gfs2_glock_init(void)
1690 {
1691         unsigned i;
1692         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
1693                 INIT_HLIST_HEAD(&gl_hash_table[i].hb_list);
1694         }
1695 #ifdef GL_HASH_LOCK_SZ
1696         for(i = 0; i < GL_HASH_LOCK_SZ; i++) {
1697                 rwlock_init(&gl_hash_locks[i]);
1698         }
1699 #endif
1700
1701         glock_workqueue = create_workqueue("glock_workqueue");
1702         if (IS_ERR(glock_workqueue))
1703                 return PTR_ERR(glock_workqueue);
1704
1705         register_shrinker(&glock_shrinker);
1706
1707         return 0;
1708 }
1709
1710 void gfs2_glock_exit(void)
1711 {
1712         unregister_shrinker(&glock_shrinker);
1713         destroy_workqueue(glock_workqueue);
1714 }
1715
1716 static int gfs2_glock_iter_next(struct gfs2_glock_iter *gi)
1717 {
1718         struct gfs2_glock *gl;
1719
1720 restart:
1721         read_lock(gl_lock_addr(gi->hash));
1722         gl = gi->gl;
1723         if (gl) {
1724                 gi->gl = hlist_entry(gl->gl_list.next,
1725                                      struct gfs2_glock, gl_list);
1726         } else {
1727                 gi->gl = hlist_entry(gl_hash_table[gi->hash].hb_list.first,
1728                                      struct gfs2_glock, gl_list);
1729         }
1730         if (gi->gl)
1731                 gfs2_glock_hold(gi->gl);
1732         read_unlock(gl_lock_addr(gi->hash));
1733         if (gl)
1734                 gfs2_glock_put(gl);
1735         while (gi->gl == NULL) {
1736                 gi->hash++;
1737                 if (gi->hash >= GFS2_GL_HASH_SIZE)
1738                         return 1;
1739                 read_lock(gl_lock_addr(gi->hash));
1740                 gi->gl = hlist_entry(gl_hash_table[gi->hash].hb_list.first,
1741                                      struct gfs2_glock, gl_list);
1742                 if (gi->gl)
1743                         gfs2_glock_hold(gi->gl);
1744                 read_unlock(gl_lock_addr(gi->hash));
1745         }
1746
1747         if (gi->sdp != gi->gl->gl_sbd)
1748                 goto restart;
1749
1750         return 0;
1751 }
1752
1753 static void gfs2_glock_iter_free(struct gfs2_glock_iter *gi)
1754 {
1755         if (gi->gl)
1756                 gfs2_glock_put(gi->gl);
1757         gi->gl = NULL;
1758 }
1759
1760 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1761 {
1762         struct gfs2_glock_iter *gi = seq->private;
1763         loff_t n = *pos;
1764
1765         gi->hash = 0;
1766
1767         do {
1768                 if (gfs2_glock_iter_next(gi)) {
1769                         gfs2_glock_iter_free(gi);
1770                         return NULL;
1771                 }
1772         } while (n--);
1773
1774         return gi->gl;
1775 }
1776
1777 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1778                                  loff_t *pos)
1779 {
1780         struct gfs2_glock_iter *gi = seq->private;
1781
1782         (*pos)++;
1783
1784         if (gfs2_glock_iter_next(gi)) {
1785                 gfs2_glock_iter_free(gi);
1786                 return NULL;
1787         }
1788
1789         return gi->gl;
1790 }
1791
1792 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1793 {
1794         struct gfs2_glock_iter *gi = seq->private;
1795         gfs2_glock_iter_free(gi);
1796 }
1797
1798 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
1799 {
1800         return dump_glock(seq, iter_ptr);
1801 }
1802
1803 static const struct seq_operations gfs2_glock_seq_ops = {
1804         .start = gfs2_glock_seq_start,
1805         .next  = gfs2_glock_seq_next,
1806         .stop  = gfs2_glock_seq_stop,
1807         .show  = gfs2_glock_seq_show,
1808 };
1809
1810 static int gfs2_debugfs_open(struct inode *inode, struct file *file)
1811 {
1812         int ret = seq_open_private(file, &gfs2_glock_seq_ops,
1813                                    sizeof(struct gfs2_glock_iter));
1814         if (ret == 0) {
1815                 struct seq_file *seq = file->private_data;
1816                 struct gfs2_glock_iter *gi = seq->private;
1817                 gi->sdp = inode->i_private;
1818         }
1819         return ret;
1820 }
1821
1822 static const struct file_operations gfs2_debug_fops = {
1823         .owner   = THIS_MODULE,
1824         .open    = gfs2_debugfs_open,
1825         .read    = seq_read,
1826         .llseek  = seq_lseek,
1827         .release = seq_release_private,
1828 };
1829
1830 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
1831 {
1832         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
1833         if (!sdp->debugfs_dir)
1834                 return -ENOMEM;
1835         sdp->debugfs_dentry_glocks = debugfs_create_file("glocks",
1836                                                          S_IFREG | S_IRUGO,
1837                                                          sdp->debugfs_dir, sdp,
1838                                                          &gfs2_debug_fops);
1839         if (!sdp->debugfs_dentry_glocks)
1840                 return -ENOMEM;
1841
1842         return 0;
1843 }
1844
1845 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
1846 {
1847         if (sdp && sdp->debugfs_dir) {
1848                 if (sdp->debugfs_dentry_glocks) {
1849                         debugfs_remove(sdp->debugfs_dentry_glocks);
1850                         sdp->debugfs_dentry_glocks = NULL;
1851                 }
1852                 debugfs_remove(sdp->debugfs_dir);
1853                 sdp->debugfs_dir = NULL;
1854         }
1855 }
1856
1857 int gfs2_register_debugfs(void)
1858 {
1859         gfs2_root = debugfs_create_dir("gfs2", NULL);
1860         return gfs2_root ? 0 : -ENOMEM;
1861 }
1862
1863 void gfs2_unregister_debugfs(void)
1864 {
1865         debugfs_remove(gfs2_root);
1866         gfs2_root = NULL;
1867 }