git.oblomov.eu Git - linux-2.6/blob - 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 /**
 800  * gfs2_glock_holder_wait
 801  * @word: unused
 802  *
 803  * This function and gfs2_glock_demote_wait both show up in the WCHAN
 804  * field. Thus I've separated these otherwise identical functions in
 805  * order to be more informative to the user.
 806  */
 807
 808 static int gfs2_glock_holder_wait(void *word)
 809 {
 810         schedule();
 811         return 0;
 812 }
 813
 814 static int gfs2_glock_demote_wait(void *word)
 815 {
 816         schedule();
 817         return 0;
 818 }
 819
 820 static void wait_on_holder(struct gfs2_holder *gh)
 821 {
 822         might_sleep();
 823         wait_on_bit(&gh->gh_iflags, HIF_WAIT, gfs2_glock_holder_wait, TASK_UNINTERRUPTIBLE);
 824 }
 825
 826 static void wait_on_demote(struct gfs2_glock *gl)
 827 {
 828         might_sleep();
 829         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, gfs2_glock_demote_wait, TASK_UNINTERRUPTIBLE);
 830 }
 831
 832 /**
 833  * handle_callback - process a demote request
 834  * @gl: the glock
 835  * @state: the state the caller wants us to change to
 836  *
 837  * There are only two requests that we are going to see in actual
 838  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 839  */
 840
 841 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 842                             unsigned long delay)
 843 {
 844         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 845
 846         set_bit(bit, &gl->gl_flags);
 847         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 848                 gl->gl_demote_state = state;
 849                 gl->gl_demote_time = jiffies;
 850         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 851                         gl->gl_demote_state != state) {
 852                 gl->gl_demote_state = LM_ST_UNLOCKED;
 853         }
 854 }
 855
 856 /**
 857  * gfs2_glock_wait - wait on a glock acquisition
 858  * @gh: the glock holder
 859  *
 860  * Returns: 0 on success
 861  */
 862
 863 int gfs2_glock_wait(struct gfs2_holder *gh)
 864 {
 865         wait_on_holder(gh);
 866         return gh->gh_error;
 867 }
 868
 869 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 870 {
 871         va_list args;
 872
 873         va_start(args, fmt);
 874         if (seq) {
 875                 struct gfs2_glock_iter *gi = seq->private;
 876                 vsprintf(gi->string, fmt, args);
 877                 seq_printf(seq, gi->string);
 878         } else {
 879                 printk(KERN_ERR " ");
 880                 vprintk(fmt, args);
 881         }
 882         va_end(args);
 883 }
 884
 885 /**
 886  * add_to_queue - Add a holder to the wait queue (but look for recursion)
 887  * @gh: the holder structure to add
 888  *
 889  * Eventually we should move the recursive locking trap to a
 890  * debugging option or something like that. This is the fast
 891  * path and needs to have the minimum number of distractions.
 892  *
 893  */
 894
 895 static inline void add_to_queue(struct gfs2_holder *gh)
 896 __releases(&gl->gl_spin)
 897 __acquires(&gl->gl_spin)
 898 {
 899         struct gfs2_glock *gl = gh->gh_gl;
 900         struct gfs2_sbd *sdp = gl->gl_sbd;
 901         struct list_head *insert_pt = NULL;
 902         struct gfs2_holder *gh2;
 903         int try_lock = 0;
 904
 905         BUG_ON(gh->gh_owner_pid == NULL);
 906         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
 907                 BUG();
 908
 909         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 910                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 911                         try_lock = 1;
 912                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 913                         goto fail;
 914         }
 915
 916         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
 917                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
 918                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
 919                         goto trap_recursive;
 920                 if (try_lock &&
 921                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) &&
 922                     !may_grant(gl, gh)) {
 923 fail:
 924                         gh->gh_error = GLR_TRYFAILED;
 925                         gfs2_holder_wake(gh);
 926                         return;
 927                 }
 928                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
 929                         continue;
 930                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
 931                         insert_pt = &gh2->gh_list;
 932         }
 933         if (likely(insert_pt == NULL)) {
 934                 list_add_tail(&gh->gh_list, &gl->gl_holders);
 935                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
 936                         goto do_cancel;
 937                 return;
 938         }
 939         list_add_tail(&gh->gh_list, insert_pt);
 940 do_cancel:
 941         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 942         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
 943                 spin_unlock(&gl->gl_spin);
 944                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
 945                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
 946                 spin_lock(&gl->gl_spin);
 947         }
 948         return;
 949
 950 trap_recursive:
 951         print_symbol(KERN_ERR "original: %s\n", gh2->gh_ip);
 952         printk(KERN_ERR "pid: %d\n", pid_nr(gh2->gh_owner_pid));
 953         printk(KERN_ERR "lock type: %d req lock state : %d\n",
 954                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
 955         print_symbol(KERN_ERR "new: %s\n", gh->gh_ip);
 956         printk(KERN_ERR "pid: %d\n", pid_nr(gh->gh_owner_pid));
 957         printk(KERN_ERR "lock type: %d req lock state : %d\n",
 958                gh->gh_gl->gl_name.ln_type, gh->gh_state);
 959         __dump_glock(NULL, gl);
 960         BUG();
 961 }
 962
 963 /**
 964  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
 965  * @gh: the holder structure
 966  *
 967  * if (gh->gh_flags & GL_ASYNC), this never returns an error
 968  *
 969  * Returns: 0, GLR_TRYFAILED, or errno on failure
 970  */
 971
 972 int gfs2_glock_nq(struct gfs2_holder *gh)
 973 {
 974         struct gfs2_glock *gl = gh->gh_gl;
 975         struct gfs2_sbd *sdp = gl->gl_sbd;
 976         int error = 0;
 977
 978         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 979                 return -EIO;
 980
 981         spin_lock(&gl->gl_spin);
 982         add_to_queue(gh);
 983         run_queue(gl, 1);
 984         spin_unlock(&gl->gl_spin);
 985
 986         if (!(gh->gh_flags & GL_ASYNC))
 987                 error = gfs2_glock_wait(gh);
 988
 989         return error;
 990 }
 991
 992 /**
 993  * gfs2_glock_poll - poll to see if an async request has been completed
 994  * @gh: the holder
 995  *
 996  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
 997  */
 998
 999 int gfs2_glock_poll(struct gfs2_holder *gh)
1000 {
1001         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1002 }
1003
1004 /**
1005  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1006  * @gh: the glock holder
1007  *
1008  */
1009
1010 void gfs2_glock_dq(struct gfs2_holder *gh)
1011 {
1012         struct gfs2_glock *gl = gh->gh_gl;
1013         const struct gfs2_glock_operations *glops = gl->gl_ops;
1014         unsigned delay = 0;
1015         int fast_path = 0;
1016
1017         spin_lock(&gl->gl_spin);
1018         if (gh->gh_flags & GL_NOCACHE)
1019                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1020
1021         list_del_init(&gh->gh_list);
1022         if (find_first_holder(gl) == NULL) {
1023                 if (glops->go_unlock) {
1024                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1025                         spin_unlock(&gl->gl_spin);
1026                         glops->go_unlock(gh);
1027                         spin_lock(&gl->gl_spin);
1028                         clear_bit(GLF_LOCK, &gl->gl_flags);
1029                 }
1030                 if (list_empty(&gl->gl_holders) &&
1031                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1032                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1033                         fast_path = 1;
1034         }
1035         spin_unlock(&gl->gl_spin);
1036         if (likely(fast_path))
1037                 return;
1038
1039         gfs2_glock_hold(gl);
1040         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1041             !test_bit(GLF_DEMOTE, &gl->gl_flags))
1042                 delay = gl->gl_ops->go_min_hold_time;
1043         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1044                 gfs2_glock_put(gl);
1045 }
1046
1047 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1048 {
1049         struct gfs2_glock *gl = gh->gh_gl;
1050         gfs2_glock_dq(gh);
1051         wait_on_demote(gl);
1052 }
1053
1054 /**
1055  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1056  * @gh: the holder structure
1057  *
1058  */
1059
1060 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1061 {
1062         gfs2_glock_dq(gh);
1063         gfs2_holder_uninit(gh);
1064 }
1065
1066 /**
1067  * gfs2_glock_nq_num - acquire a glock based on lock number
1068  * @sdp: the filesystem
1069  * @number: the lock number
1070  * @glops: the glock operations for the type of glock
1071  * @state: the state to acquire the glock in
1072  * @flags: modifier flags for the aquisition
1073  * @gh: the struct gfs2_holder
1074  *
1075  * Returns: errno
1076  */
1077
1078 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1079                       const struct gfs2_glock_operations *glops,
1080                       unsigned int state, int flags, struct gfs2_holder *gh)
1081 {
1082         struct gfs2_glock *gl;
1083         int error;
1084
1085         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1086         if (!error) {
1087                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1088                 gfs2_glock_put(gl);
1089         }
1090
1091         return error;
1092 }
1093
1094 /**
1095  * glock_compare - Compare two struct gfs2_glock structures for sorting
1096  * @arg_a: the first structure
1097  * @arg_b: the second structure
1098  *
1099  */
1100
1101 static int glock_compare(const void *arg_a, const void *arg_b)
1102 {
1103         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1104         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1105         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1106         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1107
1108         if (a->ln_number > b->ln_number)
1109                 return 1;
1110         if (a->ln_number < b->ln_number)
1111                 return -1;
1112         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1113         return 0;
1114 }
1115
1116 /**
1117  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1118  * @num_gh: the number of structures
1119  * @ghs: an array of struct gfs2_holder structures
1120  *
1121  * Returns: 0 on success (all glocks acquired),
1122  *          errno on failure (no glocks acquired)
1123  */
1124
1125 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1126                      struct gfs2_holder **p)
1127 {
1128         unsigned int x;
1129         int error = 0;
1130
1131         for (x = 0; x < num_gh; x++)
1132                 p[x] = &ghs[x];
1133
1134         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1135
1136         for (x = 0; x < num_gh; x++) {
1137                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1138
1139                 error = gfs2_glock_nq(p[x]);
1140                 if (error) {
1141                         while (x--)
1142                                 gfs2_glock_dq(p[x]);
1143                         break;
1144                 }
1145         }
1146
1147         return error;
1148 }
1149
1150 /**
1151  * gfs2_glock_nq_m - acquire multiple glocks
1152  * @num_gh: the number of structures
1153  * @ghs: an array of struct gfs2_holder structures
1154  *
1155  *
1156  * Returns: 0 on success (all glocks acquired),
1157  *          errno on failure (no glocks acquired)
1158  */
1159
1160 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1161 {
1162         struct gfs2_holder *tmp[4];
1163         struct gfs2_holder **pph = tmp;
1164         int error = 0;
1165
1166         switch(num_gh) {
1167         case 0:
1168                 return 0;
1169         case 1:
1170                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1171                 return gfs2_glock_nq(ghs);
1172         default:
1173                 if (num_gh <= 4)
1174                         break;
1175                 pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
1176                 if (!pph)
1177                         return -ENOMEM;
1178         }
1179
1180         error = nq_m_sync(num_gh, ghs, pph);
1181
1182         if (pph != tmp)
1183                 kfree(pph);
1184
1185         return error;
1186 }
1187
1188 /**
1189  * gfs2_glock_dq_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_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(&ghs[x]);
1201 }
1202
1203 /**
1204  * gfs2_glock_dq_uninit_m - release multiple glocks
1205  * @num_gh: the number of structures
1206  * @ghs: an array of struct gfs2_holder structures
1207  *
1208  */
1209
1210 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1211 {
1212         unsigned int x;
1213
1214         for (x = 0; x < num_gh; x++)
1215                 gfs2_glock_dq_uninit(&ghs[x]);
1216 }
1217
1218 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1219 {
1220         unsigned long delay = 0;
1221         unsigned long holdtime;
1222         unsigned long now = jiffies;
1223
1224         gfs2_glock_hold(gl);
1225         holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
1226         if (time_before(now, holdtime))
1227                 delay = holdtime - now;
1228         if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1229                 delay = gl->gl_ops->go_min_hold_time;
1230
1231         spin_lock(&gl->gl_spin);
1232         handle_callback(gl, state, delay);
1233         spin_unlock(&gl->gl_spin);
1234         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1235                 gfs2_glock_put(gl);
1236 }
1237
1238 /**
1239  * gfs2_glock_complete - Callback used by locking
1240  * @gl: Pointer to the glock
1241  * @ret: The return value from the dlm
1242  *
1243  */
1244
1245 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1246 {
1247         struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
1248         gl->gl_reply = ret;
1249         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_flags))) {
1250                 struct gfs2_holder *gh;
1251                 spin_lock(&gl->gl_spin);
1252                 gh = find_first_waiter(gl);
1253                 if ((!(gh && (gh->gh_flags & LM_FLAG_NOEXP)) &&
1254                      (gl->gl_target != LM_ST_UNLOCKED)) ||
1255                     ((ret & ~LM_OUT_ST_MASK) != 0))
1256                         set_bit(GLF_FROZEN, &gl->gl_flags);
1257                 spin_unlock(&gl->gl_spin);
1258                 if (test_bit(GLF_FROZEN, &gl->gl_flags))
1259                         return;
1260         }
1261         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1262         gfs2_glock_hold(gl);
1263         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1264                 gfs2_glock_put(gl);
1265 }
1266
1267 /**
1268  * demote_ok - Check to see if it's ok to unlock a glock
1269  * @gl: the glock
1270  *
1271  * Returns: 1 if it's ok
1272  */
1273
1274 static int demote_ok(const struct gfs2_glock *gl)
1275 {
1276         const struct gfs2_glock_operations *glops = gl->gl_ops;
1277
1278         if (gl->gl_state == LM_ST_UNLOCKED)
1279                 return 0;
1280         if (!list_empty(&gl->gl_holders))
1281                 return 0;
1282         if (glops->go_demote_ok)
1283                 return glops->go_demote_ok(gl);
1284         return 1;
1285 }
1286
1287
1288 static int gfs2_shrink_glock_memory(int nr, gfp_t gfp_mask)
1289 {
1290         struct gfs2_glock *gl;
1291         int may_demote;
1292         int nr_skipped = 0;
1293         int got_ref = 0;
1294         LIST_HEAD(skipped);
1295
1296         if (nr == 0)
1297                 goto out;
1298
1299         if (!(gfp_mask & __GFP_FS))
1300                 return -1;
1301
1302         spin_lock(&lru_lock);
1303         while(nr && !list_empty(&lru_list)) {
1304                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1305                 list_del_init(&gl->gl_lru);
1306                 atomic_dec(&lru_count);
1307
1308                 /* Test for being demotable */
1309                 if (!test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1310                         gfs2_glock_hold(gl);
1311                         got_ref = 1;
1312                         spin_unlock(&lru_lock);
1313                         spin_lock(&gl->gl_spin);
1314                         may_demote = demote_ok(gl);
1315                         spin_unlock(&gl->gl_spin);
1316                         clear_bit(GLF_LOCK, &gl->gl_flags);
1317                         if (may_demote) {
1318                                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1319                                 nr--;
1320                                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1321                                         gfs2_glock_put(gl);
1322                                 got_ref = 0;
1323                         }
1324                         spin_lock(&lru_lock);
1325                         if (may_demote)
1326                                 continue;
1327                 }
1328                 if (list_empty(&gl->gl_lru) &&
1329                     (atomic_read(&gl->gl_ref) <= (2 + got_ref))) {
1330                         nr_skipped++;
1331                         list_add(&gl->gl_lru, &skipped);
1332                 }
1333                 if (got_ref) {
1334                         spin_unlock(&lru_lock);
1335                         gfs2_glock_put(gl);
1336                         spin_lock(&lru_lock);
1337                         got_ref = 0;
1338                 }
1339         }
1340         list_splice(&skipped, &lru_list);
1341         atomic_add(nr_skipped, &lru_count);
1342         spin_unlock(&lru_lock);
1343 out:
1344         return (atomic_read(&lru_count) / 100) * sysctl_vfs_cache_pressure;
1345 }
1346
1347 static struct shrinker glock_shrinker = {
1348         .shrink = gfs2_shrink_glock_memory,
1349         .seeks = DEFAULT_SEEKS,
1350 };
1351
1352 /**
1353  * examine_bucket - Call a function for glock in a hash bucket
1354  * @examiner: the function
1355  * @sdp: the filesystem
1356  * @bucket: the bucket
1357  *
1358  * Returns: 1 if the bucket has entries
1359  */
1360
1361 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
1362                           unsigned int hash)
1363 {
1364         struct gfs2_glock *gl, *prev = NULL;
1365         int has_entries = 0;
1366         struct hlist_head *head = &gl_hash_table[hash].hb_list;
1367
1368         read_lock(gl_lock_addr(hash));
1369         /* Can't use hlist_for_each_entry - don't want prefetch here */
1370         if (hlist_empty(head))
1371                 goto out;
1372         gl = list_entry(head->first, struct gfs2_glock, gl_list);
1373         while(1) {
1374                 if (!sdp || gl->gl_sbd == sdp) {
1375                         gfs2_glock_hold(gl);
1376                         read_unlock(gl_lock_addr(hash));
1377                         if (prev)
1378                                 gfs2_glock_put(prev);
1379                         prev = gl;
1380                         examiner(gl);
1381                         has_entries = 1;
1382                         read_lock(gl_lock_addr(hash));
1383                 }
1384                 if (gl->gl_list.next == NULL)
1385                         break;
1386                 gl = list_entry(gl->gl_list.next, struct gfs2_glock, gl_list);
1387         }
1388 out:
1389         read_unlock(gl_lock_addr(hash));
1390         if (prev)
1391                 gfs2_glock_put(prev);
1392         cond_resched();
1393         return has_entries;
1394 }
1395
1396
1397 /**
1398  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1399  * @gl: The glock to thaw
1400  *
1401  * N.B. When we freeze a glock, we leave a ref to the glock outstanding,
1402  * so this has to result in the ref count being dropped by one.
1403  */
1404
1405 static void thaw_glock(struct gfs2_glock *gl)
1406 {
1407         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1408                 return;
1409         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1410         gfs2_glock_hold(gl);
1411         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1412                 gfs2_glock_put(gl);
1413 }
1414
1415 /**
1416  * clear_glock - look at a glock and see if we can free it from glock cache
1417  * @gl: the glock to look at
1418  *
1419  */
1420
1421 static void clear_glock(struct gfs2_glock *gl)
1422 {
1423         spin_lock(&lru_lock);
1424         if (!list_empty(&gl->gl_lru)) {
1425                 list_del_init(&gl->gl_lru);
1426                 atomic_dec(&lru_count);
1427         }
1428         spin_unlock(&lru_lock);
1429
1430         spin_lock(&gl->gl_spin);
1431         if (find_first_holder(gl) == NULL && gl->gl_state != LM_ST_UNLOCKED)
1432                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1433         spin_unlock(&gl->gl_spin);
1434         gfs2_glock_hold(gl);
1435         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1436                 gfs2_glock_put(gl);
1437 }
1438
1439 /**
1440  * gfs2_glock_thaw - Thaw any frozen glocks
1441  * @sdp: The super block
1442  *
1443  */
1444
1445 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1446 {
1447         unsigned x;
1448
1449         for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
1450                 examine_bucket(thaw_glock, sdp, x);
1451 }
1452
1453 /**
1454  * gfs2_gl_hash_clear - Empty out the glock hash table
1455  * @sdp: the filesystem
1456  * @wait: wait until it's all gone
1457  *
1458  * Called when unmounting the filesystem.
1459  */
1460
1461 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1462 {
1463         unsigned long t;
1464         unsigned int x;
1465         int cont;
1466
1467         t = jiffies;
1468
1469         for (;;) {
1470                 cont = 0;
1471                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
1472                         if (examine_bucket(clear_glock, sdp, x))
1473                                 cont = 1;
1474                 }
1475
1476                 if (!cont)
1477                         break;
1478
1479                 if (time_after_eq(jiffies,
1480                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
1481                         fs_warn(sdp, "Unmount seems to be stalled. "
1482                                      "Dumping lock state...\n");
1483                         gfs2_dump_lockstate(sdp);
1484                         t = jiffies;
1485                 }
1486
1487                 down_write(&gfs2_umount_flush_sem);
1488                 invalidate_inodes(sdp->sd_vfs);
1489                 up_write(&gfs2_umount_flush_sem);
1490                 msleep(10);
1491         }
1492 }
1493
1494 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1495 {
1496         struct gfs2_glock *gl = ip->i_gl;
1497         int ret;
1498
1499         ret = gfs2_truncatei_resume(ip);
1500         gfs2_assert_withdraw(gl->gl_sbd, ret == 0);
1501
1502         spin_lock(&gl->gl_spin);
1503         clear_bit(GLF_LOCK, &gl->gl_flags);
1504         run_queue(gl, 1);
1505         spin_unlock(&gl->gl_spin);
1506 }
1507
1508 static const char *state2str(unsigned state)
1509 {
1510         switch(state) {
1511         case LM_ST_UNLOCKED:
1512                 return "UN";
1513         case LM_ST_SHARED:
1514                 return "SH";
1515         case LM_ST_DEFERRED:
1516                 return "DF";
1517         case LM_ST_EXCLUSIVE:
1518                 return "EX";
1519         }
1520         return "??";
1521 }
1522
1523 static const char *hflags2str(char *buf, unsigned flags, unsigned long iflags)
1524 {
1525         char *p = buf;
1526         if (flags & LM_FLAG_TRY)
1527                 *p++ = 't';
1528         if (flags & LM_FLAG_TRY_1CB)
1529                 *p++ = 'T';
1530         if (flags & LM_FLAG_NOEXP)
1531                 *p++ = 'e';
1532         if (flags & LM_FLAG_ANY)
1533                 *p++ = 'A';
1534         if (flags & LM_FLAG_PRIORITY)
1535                 *p++ = 'p';
1536         if (flags & GL_ASYNC)
1537                 *p++ = 'a';
1538         if (flags & GL_EXACT)
1539                 *p++ = 'E';
1540         if (flags & GL_NOCACHE)
1541                 *p++ = 'c';
1542         if (test_bit(HIF_HOLDER, &iflags))
1543                 *p++ = 'H';
1544         if (test_bit(HIF_WAIT, &iflags))
1545                 *p++ = 'W';
1546         if (test_bit(HIF_FIRST, &iflags))
1547                 *p++ = 'F';
1548         *p = 0;
1549         return buf;
1550 }
1551
1552 /**
1553  * dump_holder - print information about a glock holder
1554  * @seq: the seq_file struct
1555  * @gh: the glock holder
1556  *
1557  * Returns: 0 on success, -ENOBUFS when we run out of space
1558  */
1559
1560 static int dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1561 {
1562         struct task_struct *gh_owner = NULL;
1563         char buffer[KSYM_SYMBOL_LEN];
1564         char flags_buf[32];
1565
1566         sprint_symbol(buffer, gh->gh_ip);
1567         if (gh->gh_owner_pid)
1568                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1569         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %s\n",
1570                   state2str(gh->gh_state),
1571                   hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1572                   gh->gh_error,
1573                   gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1574                   gh_owner ? gh_owner->comm : "(ended)", buffer);
1575         return 0;
1576 }
1577
1578 static const char *gflags2str(char *buf, const unsigned long *gflags)
1579 {
1580         char *p = buf;
1581         if (test_bit(GLF_LOCK, gflags))
1582                 *p++ = 'l';
1583         if (test_bit(GLF_DEMOTE, gflags))
1584                 *p++ = 'D';
1585         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1586                 *p++ = 'd';
1587         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1588                 *p++ = 'p';
1589         if (test_bit(GLF_DIRTY, gflags))
1590                 *p++ = 'y';
1591         if (test_bit(GLF_LFLUSH, gflags))
1592                 *p++ = 'f';
1593         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1594                 *p++ = 'i';
1595         if (test_bit(GLF_REPLY_PENDING, gflags))
1596                 *p++ = 'r';
1597         if (test_bit(GLF_INITIAL, gflags))
1598                 *p++ = 'I';
1599         if (test_bit(GLF_FROZEN, gflags))
1600                 *p++ = 'F';
1601         *p = 0;
1602         return buf;
1603 }
1604
1605 /**
1606  * __dump_glock - print information about a glock
1607  * @seq: The seq_file struct
1608  * @gl: the glock
1609  *
1610  * The file format is as follows:
1611  * One line per object, capital letters are used to indicate objects
1612  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1613  * other objects are indented by a single space and follow the glock to
1614  * which they are related. Fields are indicated by lower case letters
1615  * followed by a colon and the field value, except for strings which are in
1616  * [] so that its possible to see if they are composed of spaces for
1617  * example. The field's are n = number (id of the object), f = flags,
1618  * t = type, s = state, r = refcount, e = error, p = pid.
1619  *
1620  * Returns: 0 on success, -ENOBUFS when we run out of space
1621  */
1622
1623 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1624 {
1625         const struct gfs2_glock_operations *glops = gl->gl_ops;
1626         unsigned long long dtime;
1627         const struct gfs2_holder *gh;
1628         char gflags_buf[32];
1629         int error = 0;
1630
1631         dtime = jiffies - gl->gl_demote_time;
1632         dtime *= 1000000/HZ; /* demote time in uSec */
1633         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1634                 dtime = 0;
1635         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llu f:%s t:%s d:%s/%llu a:%d r:%d\n",
1636                   state2str(gl->gl_state),
1637                   gl->gl_name.ln_type,
1638                   (unsigned long long)gl->gl_name.ln_number,
1639                   gflags2str(gflags_buf, &gl->gl_flags),
1640                   state2str(gl->gl_target),
1641                   state2str(gl->gl_demote_state), dtime,
1642                   atomic_read(&gl->gl_ail_count),
1643                   atomic_read(&gl->gl_ref));
1644
1645         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1646                 error = dump_holder(seq, gh);
1647                 if (error)
1648                         goto out;
1649         }
1650         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1651                 error = glops->go_dump(seq, gl);
1652 out:
1653         return error;
1654 }
1655
1656 static int dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1657 {
1658         int ret;
1659         spin_lock(&gl->gl_spin);
1660         ret = __dump_glock(seq, gl);
1661         spin_unlock(&gl->gl_spin);
1662         return ret;
1663 }
1664
1665 /**
1666  * gfs2_dump_lockstate - print out the current lockstate
1667  * @sdp: the filesystem
1668  * @ub: the buffer to copy the information into
1669  *
1670  * If @ub is NULL, dump the lockstate to the console.
1671  *
1672  */
1673
1674 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
1675 {
1676         struct gfs2_glock *gl;
1677         struct hlist_node *h;
1678         unsigned int x;
1679         int error = 0;
1680
1681         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
1682
1683                 read_lock(gl_lock_addr(x));
1684
1685                 hlist_for_each_entry(gl, h, &gl_hash_table[x].hb_list, gl_list) {
1686                         if (gl->gl_sbd != sdp)
1687                                 continue;
1688
1689                         error = dump_glock(NULL, gl);
1690                         if (error)
1691                                 break;
1692                 }
1693
1694                 read_unlock(gl_lock_addr(x));
1695
1696                 if (error)
1697                         break;
1698         }
1699
1700
1701         return error;
1702 }
1703
1704
1705 int __init gfs2_glock_init(void)
1706 {
1707         unsigned i;
1708         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
1709                 INIT_HLIST_HEAD(&gl_hash_table[i].hb_list);
1710         }
1711 #ifdef GL_HASH_LOCK_SZ
1712         for(i = 0; i < GL_HASH_LOCK_SZ; i++) {
1713                 rwlock_init(&gl_hash_locks[i]);
1714         }
1715 #endif
1716
1717         glock_workqueue = create_workqueue("glock_workqueue");
1718         if (IS_ERR(glock_workqueue))
1719                 return PTR_ERR(glock_workqueue);
1720
1721         register_shrinker(&glock_shrinker);
1722
1723         return 0;
1724 }
1725
1726 void gfs2_glock_exit(void)
1727 {
1728         unregister_shrinker(&glock_shrinker);
1729         destroy_workqueue(glock_workqueue);
1730 }
1731
1732 static int gfs2_glock_iter_next(struct gfs2_glock_iter *gi)
1733 {
1734         struct gfs2_glock *gl;
1735
1736 restart:
1737         read_lock(gl_lock_addr(gi->hash));
1738         gl = gi->gl;
1739         if (gl) {
1740                 gi->gl = hlist_entry(gl->gl_list.next,
1741                                      struct gfs2_glock, gl_list);
1742         } else {
1743                 gi->gl = hlist_entry(gl_hash_table[gi->hash].hb_list.first,
1744                                      struct gfs2_glock, gl_list);
1745         }
1746         if (gi->gl)
1747                 gfs2_glock_hold(gi->gl);
1748         read_unlock(gl_lock_addr(gi->hash));
1749         if (gl)
1750                 gfs2_glock_put(gl);
1751         while (gi->gl == NULL) {
1752                 gi->hash++;
1753                 if (gi->hash >= GFS2_GL_HASH_SIZE)
1754                         return 1;
1755                 read_lock(gl_lock_addr(gi->hash));
1756                 gi->gl = hlist_entry(gl_hash_table[gi->hash].hb_list.first,
1757                                      struct gfs2_glock, gl_list);
1758                 if (gi->gl)
1759                         gfs2_glock_hold(gi->gl);
1760                 read_unlock(gl_lock_addr(gi->hash));
1761         }
1762
1763         if (gi->sdp != gi->gl->gl_sbd)
1764                 goto restart;
1765
1766         return 0;
1767 }
1768
1769 static void gfs2_glock_iter_free(struct gfs2_glock_iter *gi)
1770 {
1771         if (gi->gl)
1772                 gfs2_glock_put(gi->gl);
1773         gi->gl = NULL;
1774 }
1775
1776 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1777 {
1778         struct gfs2_glock_iter *gi = seq->private;
1779         loff_t n = *pos;
1780
1781         gi->hash = 0;
1782
1783         do {
1784                 if (gfs2_glock_iter_next(gi)) {
1785                         gfs2_glock_iter_free(gi);
1786                         return NULL;
1787                 }
1788         } while (n--);
1789
1790         return gi->gl;
1791 }
1792
1793 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1794                                  loff_t *pos)
1795 {
1796         struct gfs2_glock_iter *gi = seq->private;
1797
1798         (*pos)++;
1799
1800         if (gfs2_glock_iter_next(gi)) {
1801                 gfs2_glock_iter_free(gi);
1802                 return NULL;
1803         }
1804
1805         return gi->gl;
1806 }
1807
1808 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1809 {
1810         struct gfs2_glock_iter *gi = seq->private;
1811         gfs2_glock_iter_free(gi);
1812 }
1813
1814 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
1815 {
1816         return dump_glock(seq, iter_ptr);
1817 }
1818
1819 static const struct seq_operations gfs2_glock_seq_ops = {
1820         .start = gfs2_glock_seq_start,
1821         .next  = gfs2_glock_seq_next,
1822         .stop  = gfs2_glock_seq_stop,
1823         .show  = gfs2_glock_seq_show,
1824 };
1825
1826 static int gfs2_debugfs_open(struct inode *inode, struct file *file)
1827 {
1828         int ret = seq_open_private(file, &gfs2_glock_seq_ops,
1829                                    sizeof(struct gfs2_glock_iter));
1830         if (ret == 0) {
1831                 struct seq_file *seq = file->private_data;
1832                 struct gfs2_glock_iter *gi = seq->private;
1833                 gi->sdp = inode->i_private;
1834         }
1835         return ret;
1836 }
1837
1838 static const struct file_operations gfs2_debug_fops = {
1839         .owner   = THIS_MODULE,
1840         .open    = gfs2_debugfs_open,
1841         .read    = seq_read,
1842         .llseek  = seq_lseek,
1843         .release = seq_release_private,
1844 };
1845
1846 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
1847 {
1848         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
1849         if (!sdp->debugfs_dir)
1850                 return -ENOMEM;
1851         sdp->debugfs_dentry_glocks = debugfs_create_file("glocks",
1852                                                          S_IFREG | S_IRUGO,
1853                                                          sdp->debugfs_dir, sdp,
1854                                                          &gfs2_debug_fops);
1855         if (!sdp->debugfs_dentry_glocks)
1856                 return -ENOMEM;
1857
1858         return 0;
1859 }
1860
1861 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
1862 {
1863         if (sdp && sdp->debugfs_dir) {
1864                 if (sdp->debugfs_dentry_glocks) {
1865                         debugfs_remove(sdp->debugfs_dentry_glocks);
1866                         sdp->debugfs_dentry_glocks = NULL;
1867                 }
1868                 debugfs_remove(sdp->debugfs_dir);
1869                 sdp->debugfs_dir = NULL;
1870         }
1871 }
1872
1873 int gfs2_register_debugfs(void)
1874 {
1875         gfs2_root = debugfs_create_dir("gfs2", NULL);
1876         return gfs2_root ? 0 : -ENOMEM;
1877 }
1878
1879 void gfs2_unregister_debugfs(void)
1880 {
1881         debugfs_remove(gfs2_root);
1882         gfs2_root = NULL;
1883 }