fs/gfs2/glock.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2006 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/completion.h>
  14 #include <linux/buffer_head.h>
  15 #include <linux/delay.h>
  16 #include <linux/sort.h>
  17 #include <linux/jhash.h>
  18 #include <linux/kref.h>
  19 #include <linux/kallsyms.h>
  20 #include <linux/gfs2_ondisk.h>
  21 #include <asm/uaccess.h>
  22
  23 #include "gfs2.h"
  24 #include "lm_interface.h"
  25 #include "incore.h"
  26 #include "glock.h"
  27 #include "glops.h"
  28 #include "inode.h"
  29 #include "lm.h"
  30 #include "lops.h"
  31 #include "meta_io.h"
  32 #include "quota.h"
  33 #include "super.h"
  34 #include "util.h"
  35
  36 /*  Must be kept in sync with the beginning of struct gfs2_glock  */
  37 struct glock_plug {
  38         struct list_head gl_list;
  39         unsigned long gl_flags;
  40 };
  41
  42 struct greedy {
  43         struct gfs2_holder gr_gh;
  44         struct work_struct gr_work;
  45 };
  46
  47 struct gfs2_gl_hash_bucket {
  48         struct list_head hb_list;
  49 };
  50
  51 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  52
  53 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp);
  54 static int dump_glock(struct gfs2_glock *gl);
  55
  56 #define GFS2_GL_HASH_SHIFT      13
  57 #define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
  58 #define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
  59
  60 static struct gfs2_gl_hash_bucket gl_hash_table[GFS2_GL_HASH_SIZE];
  61
  62 /*
  63  * Despite what you might think, the numbers below are not arbitrary :-)
  64  * They are taken from the ipv4 routing hash code, which is well tested
  65  * and thus should be nearly optimal. Later on we might tweek the numbers
  66  * but for now this should be fine.
  67  *
  68  * The reason for putting the locks in a separate array from the list heads
  69  * is that we can have fewer locks than list heads and save memory. We use
  70  * the same hash function for both, but with a different hash mask.
  71  */
  72 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \
  73         defined(CONFIG_PROVE_LOCKING)
  74
  75 #ifdef CONFIG_LOCKDEP
  76 # define GL_HASH_LOCK_SZ        256
  77 #else
  78 # if NR_CPUS >= 32
  79 #  define GL_HASH_LOCK_SZ       4096
  80 # elif NR_CPUS >= 16
  81 #  define GL_HASH_LOCK_SZ       2048
  82 # elif NR_CPUS >= 8
  83 #  define GL_HASH_LOCK_SZ       1024
  84 # elif NR_CPUS >= 4
  85 #  define GL_HASH_LOCK_SZ       512
  86 # else
  87 #  define GL_HASH_LOCK_SZ       256
  88 # endif
  89 #endif
  90
  91 /* We never want more locks than chains */
  92 #if GFS2_GL_HASH_SIZE < GL_HASH_LOCK_SZ
  93 # undef GL_HASH_LOCK_SZ
  94 # define GL_HASH_LOCK_SZ GFS2_GL_HASH_SIZE
  95 #endif
  96
  97 static rwlock_t gl_hash_locks[GL_HASH_LOCK_SZ];
  98
  99 static inline rwlock_t *gl_lock_addr(unsigned int x)
 100 {
 101         return &gl_hash_locks[(x) & (GL_HASH_LOCK_SZ-1)];
 102 }
 103 #else /* not SMP, so no spinlocks required */
 104 static inline rwlock_t *gl_lock_addr(x)
 105 {
 106         return NULL;
 107 }
 108 #endif
 109
 110 /**
 111  * relaxed_state_ok - is a requested lock compatible with the current lock mode?
 112  * @actual: the current state of the lock
 113  * @requested: the lock state that was requested by the caller
 114  * @flags: the modifier flags passed in by the caller
 115  *
 116  * Returns: 1 if the locks are compatible, 0 otherwise
 117  */
 118
 119 static inline int relaxed_state_ok(unsigned int actual, unsigned requested,
 120                                    int flags)
 121 {
 122         if (actual == requested)
 123                 return 1;
 124
 125         if (flags & GL_EXACT)
 126                 return 0;
 127
 128         if (actual == LM_ST_EXCLUSIVE && requested == LM_ST_SHARED)
 129                 return 1;
 130
 131         if (actual != LM_ST_UNLOCKED && (flags & LM_FLAG_ANY))
 132                 return 1;
 133
 134         return 0;
 135 }
 136
 137 /**
 138  * gl_hash() - Turn glock number into hash bucket number
 139  * @lock: The glock number
 140  *
 141  * Returns: The number of the corresponding hash bucket
 142  */
 143
 144 static unsigned int gl_hash(const struct gfs2_sbd *sdp,
 145                             const struct lm_lockname *name)
 146 {
 147         unsigned int h;
 148
 149         h = jhash(&name->ln_number, sizeof(u64), 0);
 150         h = jhash(&name->ln_type, sizeof(unsigned int), h);
 151         h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
 152         h &= GFS2_GL_HASH_MASK;
 153
 154         return h;
 155 }
 156
 157 /**
 158  * glock_free() - Perform a few checks and then release struct gfs2_glock
 159  * @gl: The glock to release
 160  *
 161  * Also calls lock module to release its internal structure for this glock.
 162  *
 163  */
 164
 165 static void glock_free(struct gfs2_glock *gl)
 166 {
 167         struct gfs2_sbd *sdp = gl->gl_sbd;
 168         struct inode *aspace = gl->gl_aspace;
 169
 170         gfs2_lm_put_lock(sdp, gl->gl_lock);
 171
 172         if (aspace)
 173                 gfs2_aspace_put(aspace);
 174
 175         kmem_cache_free(gfs2_glock_cachep, gl);
 176 }
 177
 178 /**
 179  * gfs2_glock_hold() - increment reference count on glock
 180  * @gl: The glock to hold
 181  *
 182  */
 183
 184 void gfs2_glock_hold(struct gfs2_glock *gl)
 185 {
 186         kref_get(&gl->gl_ref);
 187 }
 188
 189 /* All work is done after the return from kref_put() so we
 190    can release the write_lock before the free. */
 191
 192 static void kill_glock(struct kref *kref)
 193 {
 194         struct gfs2_glock *gl = container_of(kref, struct gfs2_glock, gl_ref);
 195         struct gfs2_sbd *sdp = gl->gl_sbd;
 196
 197         gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
 198         gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
 199         gfs2_assert(sdp, list_empty(&gl->gl_holders));
 200         gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
 201         gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
 202         gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
 203 }
 204
 205 /**
 206  * gfs2_glock_put() - Decrement reference count on glock
 207  * @gl: The glock to put
 208  *
 209  */
 210
 211 int gfs2_glock_put(struct gfs2_glock *gl)
 212 {
 213         int rv = 0;
 214
 215         write_lock(gl_lock_addr(gl->gl_hash));
 216         if (kref_put(&gl->gl_ref, kill_glock)) {
 217                 list_del_init(&gl_hash_table[gl->gl_hash].hb_list);
 218                 write_unlock(gl_lock_addr(gl->gl_hash));
 219                 BUG_ON(spin_is_locked(&gl->gl_spin));
 220                 glock_free(gl);
 221                 rv = 1;
 222                 goto out;
 223         }
 224         write_unlock(gl_lock_addr(gl->gl_hash));
 225 out:
 226         return rv;
 227 }
 228
 229 /**
 230  * queue_empty - check to see if a glock's queue is empty
 231  * @gl: the glock
 232  * @head: the head of the queue to check
 233  *
 234  * This function protects the list in the event that a process already
 235  * has a holder on the list and is adding a second holder for itself.
 236  * The glmutex lock is what generally prevents processes from working
 237  * on the same glock at once, but the special case of adding a second
 238  * holder for yourself ("recursive" locking) doesn't involve locking
 239  * glmutex, making the spin lock necessary.
 240  *
 241  * Returns: 1 if the queue is empty
 242  */
 243
 244 static inline int queue_empty(struct gfs2_glock *gl, struct list_head *head)
 245 {
 246         int empty;
 247         spin_lock(&gl->gl_spin);
 248         empty = list_empty(head);
 249         spin_unlock(&gl->gl_spin);
 250         return empty;
 251 }
 252
 253 /**
 254  * search_bucket() - Find struct gfs2_glock by lock number
 255  * @bucket: the bucket to search
 256  * @name: The lock name
 257  *
 258  * Returns: NULL, or the struct gfs2_glock with the requested number
 259  */
 260
 261 static struct gfs2_glock *search_bucket(unsigned int hash,
 262                                         const struct gfs2_sbd *sdp,
 263                                         const struct lm_lockname *name)
 264 {
 265         struct gfs2_glock *gl;
 266
 267         list_for_each_entry(gl, &gl_hash_table[hash].hb_list, gl_list) {
 268                 if (test_bit(GLF_PLUG, &gl->gl_flags))
 269                         continue;
 270                 if (!lm_name_equal(&gl->gl_name, name))
 271                         continue;
 272                 if (gl->gl_sbd != sdp)
 273                         continue;
 274
 275                 kref_get(&gl->gl_ref);
 276
 277                 return gl;
 278         }
 279
 280         return NULL;
 281 }
 282
 283 /**
 284  * gfs2_glock_find() - Find glock by lock number
 285  * @sdp: The GFS2 superblock
 286  * @name: The lock name
 287  *
 288  * Returns: NULL, or the struct gfs2_glock with the requested number
 289  */
 290
 291 static struct gfs2_glock *gfs2_glock_find(const struct gfs2_sbd *sdp,
 292                                           const struct lm_lockname *name)
 293 {
 294         unsigned int hash = gl_hash(sdp, name);
 295         struct gfs2_glock *gl;
 296
 297         read_lock(gl_lock_addr(hash));
 298         gl = search_bucket(hash, sdp, name);
 299         read_unlock(gl_lock_addr(hash));
 300
 301         return gl;
 302 }
 303
 304 /**
 305  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 306  * @sdp: The GFS2 superblock
 307  * @number: the lock number
 308  * @glops: The glock_operations to use
 309  * @create: If 0, don't create the glock if it doesn't exist
 310  * @glp: the glock is returned here
 311  *
 312  * This does not lock a glock, just finds/creates structures for one.
 313  *
 314  * Returns: errno
 315  */
 316
 317 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 318                    const struct gfs2_glock_operations *glops, int create,
 319                    struct gfs2_glock **glp)
 320 {
 321         struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
 322         struct gfs2_glock *gl, *tmp;
 323         unsigned int hash = gl_hash(sdp, &name);
 324         int error;
 325
 326         read_lock(gl_lock_addr(hash));
 327         gl = search_bucket(hash, sdp, &name);
 328         read_unlock(gl_lock_addr(hash));
 329
 330         if (gl || !create) {
 331                 *glp = gl;
 332                 return 0;
 333         }
 334
 335         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
 336         if (!gl)
 337                 return -ENOMEM;
 338
 339         gl->gl_flags = 0;
 340         gl->gl_name = name;
 341         kref_init(&gl->gl_ref);
 342         gl->gl_state = LM_ST_UNLOCKED;
 343         gl->gl_hash = hash;
 344         gl->gl_owner = NULL;
 345         gl->gl_ip = 0;
 346         gl->gl_ops = glops;
 347         gl->gl_req_gh = NULL;
 348         gl->gl_req_bh = NULL;
 349         gl->gl_vn = 0;
 350         gl->gl_stamp = jiffies;
 351         gl->gl_object = NULL;
 352         gl->gl_sbd = sdp;
 353         gl->gl_aspace = NULL;
 354         lops_init_le(&gl->gl_le, &gfs2_glock_lops);
 355
 356         /* If this glock protects actual on-disk data or metadata blocks,
 357            create a VFS inode to manage the pages/buffers holding them. */
 358         if (glops == &gfs2_inode_glops || glops == &gfs2_rgrp_glops) {
 359                 gl->gl_aspace = gfs2_aspace_get(sdp);
 360                 if (!gl->gl_aspace) {
 361                         error = -ENOMEM;
 362                         goto fail;
 363                 }
 364         }
 365
 366         error = gfs2_lm_get_lock(sdp, &name, &gl->gl_lock);
 367         if (error)
 368                 goto fail_aspace;
 369
 370         write_lock(gl_lock_addr(hash));
 371         tmp = search_bucket(hash, sdp, &name);
 372         if (tmp) {
 373                 write_unlock(gl_lock_addr(hash));
 374                 glock_free(gl);
 375                 gl = tmp;
 376         } else {
 377                 list_add_tail(&gl->gl_list, &gl_hash_table[hash].hb_list);
 378                 write_unlock(gl_lock_addr(hash));
 379         }
 380
 381         *glp = gl;
 382
 383         return 0;
 384
 385 fail_aspace:
 386         if (gl->gl_aspace)
 387                 gfs2_aspace_put(gl->gl_aspace);
 388 fail:
 389         kmem_cache_free(gfs2_glock_cachep, gl);
 390         return error;
 391 }
 392
 393 /**
 394  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 395  * @gl: the glock
 396  * @state: the state we're requesting
 397  * @flags: the modifier flags
 398  * @gh: the holder structure
 399  *
 400  */
 401
 402 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
 403                       struct gfs2_holder *gh)
 404 {
 405         INIT_LIST_HEAD(&gh->gh_list);
 406         gh->gh_gl = gl;
 407         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 408         gh->gh_owner = current;
 409         gh->gh_state = state;
 410         gh->gh_flags = flags;
 411         gh->gh_error = 0;
 412         gh->gh_iflags = 0;
 413         init_completion(&gh->gh_wait);
 414
 415         if (gh->gh_state == LM_ST_EXCLUSIVE)
 416                 gh->gh_flags |= GL_LOCAL_EXCL;
 417
 418         gfs2_glock_hold(gl);
 419 }
 420
 421 /**
 422  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 423  * @state: the state we're requesting
 424  * @flags: the modifier flags
 425  * @gh: the holder structure
 426  *
 427  * Don't mess with the glock.
 428  *
 429  */
 430
 431 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
 432 {
 433         gh->gh_state = state;
 434         gh->gh_flags = flags;
 435         if (gh->gh_state == LM_ST_EXCLUSIVE)
 436                 gh->gh_flags |= GL_LOCAL_EXCL;
 437
 438         gh->gh_iflags &= 1 << HIF_ALLOCED;
 439         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 440 }
 441
 442 /**
 443  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 444  * @gh: the holder structure
 445  *
 446  */
 447
 448 void gfs2_holder_uninit(struct gfs2_holder *gh)
 449 {
 450         gfs2_glock_put(gh->gh_gl);
 451         gh->gh_gl = NULL;
 452         gh->gh_ip = 0;
 453 }
 454
 455 /**
 456  * gfs2_holder_get - get a struct gfs2_holder structure
 457  * @gl: the glock
 458  * @state: the state we're requesting
 459  * @flags: the modifier flags
 460  * @gfp_flags:
 461  *
 462  * Figure out how big an impact this function has.  Either:
 463  * 1) Replace it with a cache of structures hanging off the struct gfs2_sbd
 464  * 2) Leave it like it is
 465  *
 466  * Returns: the holder structure, NULL on ENOMEM
 467  */
 468
 469 static struct gfs2_holder *gfs2_holder_get(struct gfs2_glock *gl,
 470                                            unsigned int state,
 471                                            int flags, gfp_t gfp_flags)
 472 {
 473         struct gfs2_holder *gh;
 474
 475         gh = kmalloc(sizeof(struct gfs2_holder), gfp_flags);
 476         if (!gh)
 477                 return NULL;
 478
 479         gfs2_holder_init(gl, state, flags, gh);
 480         set_bit(HIF_ALLOCED, &gh->gh_iflags);
 481         gh->gh_ip = (unsigned long)__builtin_return_address(0);
 482         return gh;
 483 }
 484
 485 /**
 486  * gfs2_holder_put - get rid of a struct gfs2_holder structure
 487  * @gh: the holder structure
 488  *
 489  */
 490
 491 static void gfs2_holder_put(struct gfs2_holder *gh)
 492 {
 493         gfs2_holder_uninit(gh);
 494         kfree(gh);
 495 }
 496
 497 /**
 498  * rq_mutex - process a mutex request in the queue
 499  * @gh: the glock holder
 500  *
 501  * Returns: 1 if the queue is blocked
 502  */
 503
 504 static int rq_mutex(struct gfs2_holder *gh)
 505 {
 506         struct gfs2_glock *gl = gh->gh_gl;
 507
 508         list_del_init(&gh->gh_list);
 509         /*  gh->gh_error never examined.  */
 510         set_bit(GLF_LOCK, &gl->gl_flags);
 511         complete(&gh->gh_wait);
 512
 513         return 1;
 514 }
 515
 516 /**
 517  * rq_promote - process a promote request in the queue
 518  * @gh: the glock holder
 519  *
 520  * Acquire a new inter-node lock, or change a lock state to more restrictive.
 521  *
 522  * Returns: 1 if the queue is blocked
 523  */
 524
 525 static int rq_promote(struct gfs2_holder *gh)
 526 {
 527         struct gfs2_glock *gl = gh->gh_gl;
 528         struct gfs2_sbd *sdp = gl->gl_sbd;
 529         const struct gfs2_glock_operations *glops = gl->gl_ops;
 530
 531         if (!relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 532                 if (list_empty(&gl->gl_holders)) {
 533                         gl->gl_req_gh = gh;
 534                         set_bit(GLF_LOCK, &gl->gl_flags);
 535                         spin_unlock(&gl->gl_spin);
 536
 537                         if (atomic_read(&sdp->sd_reclaim_count) >
 538                             gfs2_tune_get(sdp, gt_reclaim_limit) &&
 539                             !(gh->gh_flags & LM_FLAG_PRIORITY)) {
 540                                 gfs2_reclaim_glock(sdp);
 541                                 gfs2_reclaim_glock(sdp);
 542                         }
 543
 544                         glops->go_xmote_th(gl, gh->gh_state, gh->gh_flags);
 545                         spin_lock(&gl->gl_spin);
 546                 }
 547                 return 1;
 548         }
 549
 550         if (list_empty(&gl->gl_holders)) {
 551                 set_bit(HIF_FIRST, &gh->gh_iflags);
 552                 set_bit(GLF_LOCK, &gl->gl_flags);
 553         } else {
 554                 struct gfs2_holder *next_gh;
 555                 if (gh->gh_flags & GL_LOCAL_EXCL)
 556                         return 1;
 557                 next_gh = list_entry(gl->gl_holders.next, struct gfs2_holder,
 558                                      gh_list);
 559                 if (next_gh->gh_flags & GL_LOCAL_EXCL)
 560                          return 1;
 561         }
 562
 563         list_move_tail(&gh->gh_list, &gl->gl_holders);
 564         gh->gh_error = 0;
 565         set_bit(HIF_HOLDER, &gh->gh_iflags);
 566
 567         complete(&gh->gh_wait);
 568
 569         return 0;
 570 }
 571
 572 /**
 573  * rq_demote - process a demote request in the queue
 574  * @gh: the glock holder
 575  *
 576  * Returns: 1 if the queue is blocked
 577  */
 578
 579 static int rq_demote(struct gfs2_holder *gh)
 580 {
 581         struct gfs2_glock *gl = gh->gh_gl;
 582         const struct gfs2_glock_operations *glops = gl->gl_ops;
 583
 584         if (!list_empty(&gl->gl_holders))
 585                 return 1;
 586
 587         if (gl->gl_state == gh->gh_state || gl->gl_state == LM_ST_UNLOCKED) {
 588                 list_del_init(&gh->gh_list);
 589                 gh->gh_error = 0;
 590                 spin_unlock(&gl->gl_spin);
 591                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 592                         gfs2_holder_put(gh);
 593                 else
 594                         complete(&gh->gh_wait);
 595                 spin_lock(&gl->gl_spin);
 596         } else {
 597                 gl->gl_req_gh = gh;
 598                 set_bit(GLF_LOCK, &gl->gl_flags);
 599                 spin_unlock(&gl->gl_spin);
 600
 601                 if (gh->gh_state == LM_ST_UNLOCKED ||
 602                     gl->gl_state != LM_ST_EXCLUSIVE)
 603                         glops->go_drop_th(gl);
 604                 else
 605                         glops->go_xmote_th(gl, gh->gh_state, gh->gh_flags);
 606
 607                 spin_lock(&gl->gl_spin);
 608         }
 609
 610         return 0;
 611 }
 612
 613 /**
 614  * rq_greedy - process a queued request to drop greedy status
 615  * @gh: the glock holder
 616  *
 617  * Returns: 1 if the queue is blocked
 618  */
 619
 620 static int rq_greedy(struct gfs2_holder *gh)
 621 {
 622         struct gfs2_glock *gl = gh->gh_gl;
 623
 624         list_del_init(&gh->gh_list);
 625         /*  gh->gh_error never examined.  */
 626         clear_bit(GLF_GREEDY, &gl->gl_flags);
 627         spin_unlock(&gl->gl_spin);
 628
 629         gfs2_holder_uninit(gh);
 630         kfree(container_of(gh, struct greedy, gr_gh));
 631
 632         spin_lock(&gl->gl_spin);
 633
 634         return 0;
 635 }
 636
 637 /**
 638  * run_queue - process holder structures on a glock
 639  * @gl: the glock
 640  *
 641  */
 642 static void run_queue(struct gfs2_glock *gl)
 643 {
 644         struct gfs2_holder *gh;
 645         int blocked = 1;
 646
 647         for (;;) {
 648                 if (test_bit(GLF_LOCK, &gl->gl_flags))
 649                         break;
 650
 651                 if (!list_empty(&gl->gl_waiters1)) {
 652                         gh = list_entry(gl->gl_waiters1.next,
 653                                         struct gfs2_holder, gh_list);
 654
 655                         if (test_bit(HIF_MUTEX, &gh->gh_iflags))
 656                                 blocked = rq_mutex(gh);
 657                         else
 658                                 gfs2_assert_warn(gl->gl_sbd, 0);
 659
 660                 } else if (!list_empty(&gl->gl_waiters2) &&
 661                            !test_bit(GLF_SKIP_WAITERS2, &gl->gl_flags)) {
 662                         gh = list_entry(gl->gl_waiters2.next,
 663                                         struct gfs2_holder, gh_list);
 664
 665                         if (test_bit(HIF_DEMOTE, &gh->gh_iflags))
 666                                 blocked = rq_demote(gh);
 667                         else if (test_bit(HIF_GREEDY, &gh->gh_iflags))
 668                                 blocked = rq_greedy(gh);
 669                         else
 670                                 gfs2_assert_warn(gl->gl_sbd, 0);
 671
 672                 } else if (!list_empty(&gl->gl_waiters3)) {
 673                         gh = list_entry(gl->gl_waiters3.next,
 674                                         struct gfs2_holder, gh_list);
 675
 676                         if (test_bit(HIF_PROMOTE, &gh->gh_iflags))
 677                                 blocked = rq_promote(gh);
 678                         else
 679                                 gfs2_assert_warn(gl->gl_sbd, 0);
 680
 681                 } else
 682                         break;
 683
 684                 if (blocked)
 685                         break;
 686         }
 687 }
 688
 689 /**
 690  * gfs2_glmutex_lock - acquire a local lock on a glock
 691  * @gl: the glock
 692  *
 693  * Gives caller exclusive access to manipulate a glock structure.
 694  */
 695
 696 static void gfs2_glmutex_lock(struct gfs2_glock *gl)
 697 {
 698         struct gfs2_holder gh;
 699
 700         gfs2_holder_init(gl, 0, 0, &gh);
 701         set_bit(HIF_MUTEX, &gh.gh_iflags);
 702
 703         spin_lock(&gl->gl_spin);
 704         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
 705                 list_add_tail(&gh.gh_list, &gl->gl_waiters1);
 706         } else {
 707                 gl->gl_owner = current;
 708                 gl->gl_ip = (unsigned long)__builtin_return_address(0);
 709                 complete(&gh.gh_wait);
 710         }
 711         spin_unlock(&gl->gl_spin);
 712
 713         wait_for_completion(&gh.gh_wait);
 714         gfs2_holder_uninit(&gh);
 715 }
 716
 717 /**
 718  * gfs2_glmutex_trylock - try to acquire a local lock on a glock
 719  * @gl: the glock
 720  *
 721  * Returns: 1 if the glock is acquired
 722  */
 723
 724 static int gfs2_glmutex_trylock(struct gfs2_glock *gl)
 725 {
 726         int acquired = 1;
 727
 728         spin_lock(&gl->gl_spin);
 729         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
 730                 acquired = 0;
 731         } else {
 732                 gl->gl_owner = current;
 733                 gl->gl_ip = (unsigned long)__builtin_return_address(0);
 734         }
 735         spin_unlock(&gl->gl_spin);
 736
 737         return acquired;
 738 }
 739
 740 /**
 741  * gfs2_glmutex_unlock - release a local lock on a glock
 742  * @gl: the glock
 743  *
 744  */
 745
 746 static void gfs2_glmutex_unlock(struct gfs2_glock *gl)
 747 {
 748         spin_lock(&gl->gl_spin);
 749         clear_bit(GLF_LOCK, &gl->gl_flags);
 750         gl->gl_owner = NULL;
 751         gl->gl_ip = 0;
 752         run_queue(gl);
 753         BUG_ON(!spin_is_locked(&gl->gl_spin));
 754         spin_unlock(&gl->gl_spin);
 755 }
 756
 757 /**
 758  * handle_callback - add a demote request to a lock's queue
 759  * @gl: the glock
 760  * @state: the state the caller wants us to change to
 761  *
 762  * Note: This may fail sliently if we are out of memory.
 763  */
 764
 765 static void handle_callback(struct gfs2_glock *gl, unsigned int state)
 766 {
 767         struct gfs2_holder *gh, *new_gh = NULL;
 768
 769 restart:
 770         spin_lock(&gl->gl_spin);
 771
 772         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
 773                 if (test_bit(HIF_DEMOTE, &gh->gh_iflags) &&
 774                     gl->gl_req_gh != gh) {
 775                         if (gh->gh_state != state)
 776                                 gh->gh_state = LM_ST_UNLOCKED;
 777                         goto out;
 778                 }
 779         }
 780
 781         if (new_gh) {
 782                 list_add_tail(&new_gh->gh_list, &gl->gl_waiters2);
 783                 new_gh = NULL;
 784         } else {
 785                 spin_unlock(&gl->gl_spin);
 786
 787                 new_gh = gfs2_holder_get(gl, state, LM_FLAG_TRY, GFP_KERNEL);
 788                 if (!new_gh)
 789                         return;
 790                 set_bit(HIF_DEMOTE, &new_gh->gh_iflags);
 791                 set_bit(HIF_DEALLOC, &new_gh->gh_iflags);
 792
 793                 goto restart;
 794         }
 795
 796 out:
 797         spin_unlock(&gl->gl_spin);
 798
 799         if (new_gh)
 800                 gfs2_holder_put(new_gh);
 801 }
 802
 803 void gfs2_glock_inode_squish(struct inode *inode)
 804 {
 805         struct gfs2_holder gh;
 806         struct gfs2_glock *gl = GFS2_I(inode)->i_gl;
 807         gfs2_holder_init(gl, LM_ST_UNLOCKED, 0, &gh);
 808         set_bit(HIF_DEMOTE, &gh.gh_iflags);
 809         spin_lock(&gl->gl_spin);
 810         gfs2_assert(inode->i_sb->s_fs_info, list_empty(&gl->gl_holders));
 811         list_add_tail(&gh.gh_list, &gl->gl_waiters2);
 812         run_queue(gl);
 813         spin_unlock(&gl->gl_spin);
 814         wait_for_completion(&gh.gh_wait);
 815         gfs2_holder_uninit(&gh);
 816 }
 817
 818 /**
 819  * state_change - record that the glock is now in a different state
 820  * @gl: the glock
 821  * @new_state the new state
 822  *
 823  */
 824
 825 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 826 {
 827         int held1, held2;
 828
 829         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 830         held2 = (new_state != LM_ST_UNLOCKED);
 831
 832         if (held1 != held2) {
 833                 if (held2)
 834                         gfs2_glock_hold(gl);
 835                 else
 836                         gfs2_glock_put(gl);
 837         }
 838
 839         gl->gl_state = new_state;
 840 }
 841
 842 /**
 843  * xmote_bh - Called after the lock module is done acquiring a lock
 844  * @gl: The glock in question
 845  * @ret: the int returned from the lock module
 846  *
 847  */
 848
 849 static void xmote_bh(struct gfs2_glock *gl, unsigned int ret)
 850 {
 851         struct gfs2_sbd *sdp = gl->gl_sbd;
 852         const struct gfs2_glock_operations *glops = gl->gl_ops;
 853         struct gfs2_holder *gh = gl->gl_req_gh;
 854         int prev_state = gl->gl_state;
 855         int op_done = 1;
 856
 857         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 858         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 859         gfs2_assert_warn(sdp, !(ret & LM_OUT_ASYNC));
 860
 861         state_change(gl, ret & LM_OUT_ST_MASK);
 862
 863         if (prev_state != LM_ST_UNLOCKED && !(ret & LM_OUT_CACHEABLE)) {
 864                 if (glops->go_inval)
 865                         glops->go_inval(gl, DIO_METADATA | DIO_DATA);
 866         } else if (gl->gl_state == LM_ST_DEFERRED) {
 867                 /* We might not want to do this here.
 868                    Look at moving to the inode glops. */
 869                 if (glops->go_inval)
 870                         glops->go_inval(gl, DIO_DATA);
 871         }
 872
 873         /*  Deal with each possible exit condition  */
 874
 875         if (!gh)
 876                 gl->gl_stamp = jiffies;
 877         else if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
 878                 spin_lock(&gl->gl_spin);
 879                 list_del_init(&gh->gh_list);
 880                 gh->gh_error = -EIO;
 881                 spin_unlock(&gl->gl_spin);
 882         } else if (test_bit(HIF_DEMOTE, &gh->gh_iflags)) {
 883                 spin_lock(&gl->gl_spin);
 884                 list_del_init(&gh->gh_list);
 885                 if (gl->gl_state == gh->gh_state ||
 886                     gl->gl_state == LM_ST_UNLOCKED) {
 887                         gh->gh_error = 0;
 888                 } else {
 889                         if (gfs2_assert_warn(sdp, gh->gh_flags &
 890                                         (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) == -1)
 891                                 fs_warn(sdp, "ret = 0x%.8X\n", ret);
 892                         gh->gh_error = GLR_TRYFAILED;
 893                 }
 894                 spin_unlock(&gl->gl_spin);
 895
 896                 if (ret & LM_OUT_CANCELED)
 897                         handle_callback(gl, LM_ST_UNLOCKED);
 898
 899         } else if (ret & LM_OUT_CANCELED) {
 900                 spin_lock(&gl->gl_spin);
 901                 list_del_init(&gh->gh_list);
 902                 gh->gh_error = GLR_CANCELED;
 903                 spin_unlock(&gl->gl_spin);
 904
 905         } else if (relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
 906                 spin_lock(&gl->gl_spin);
 907                 list_move_tail(&gh->gh_list, &gl->gl_holders);
 908                 gh->gh_error = 0;
 909                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 910                 spin_unlock(&gl->gl_spin);
 911
 912                 set_bit(HIF_FIRST, &gh->gh_iflags);
 913
 914                 op_done = 0;
 915
 916         } else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
 917                 spin_lock(&gl->gl_spin);
 918                 list_del_init(&gh->gh_list);
 919                 gh->gh_error = GLR_TRYFAILED;
 920                 spin_unlock(&gl->gl_spin);
 921
 922         } else {
 923                 if (gfs2_assert_withdraw(sdp, 0) == -1)
 924                         fs_err(sdp, "ret = 0x%.8X\n", ret);
 925         }
 926
 927         if (glops->go_xmote_bh)
 928                 glops->go_xmote_bh(gl);
 929
 930         if (op_done) {
 931                 spin_lock(&gl->gl_spin);
 932                 gl->gl_req_gh = NULL;
 933                 gl->gl_req_bh = NULL;
 934                 clear_bit(GLF_LOCK, &gl->gl_flags);
 935                 run_queue(gl);
 936                 spin_unlock(&gl->gl_spin);
 937         }
 938
 939         gfs2_glock_put(gl);
 940
 941         if (gh) {
 942                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
 943                         gfs2_holder_put(gh);
 944                 else
 945                         complete(&gh->gh_wait);
 946         }
 947 }
 948
 949 /**
 950  * gfs2_glock_xmote_th - Call into the lock module to acquire or change a glock
 951  * @gl: The glock in question
 952  * @state: the requested state
 953  * @flags: modifier flags to the lock call
 954  *
 955  */
 956
 957 void gfs2_glock_xmote_th(struct gfs2_glock *gl, unsigned int state, int flags)
 958 {
 959         struct gfs2_sbd *sdp = gl->gl_sbd;
 960         const struct gfs2_glock_operations *glops = gl->gl_ops;
 961         int lck_flags = flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB |
 962                                  LM_FLAG_NOEXP | LM_FLAG_ANY |
 963                                  LM_FLAG_PRIORITY);
 964         unsigned int lck_ret;
 965
 966         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
 967         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
 968         gfs2_assert_warn(sdp, state != LM_ST_UNLOCKED);
 969         gfs2_assert_warn(sdp, state != gl->gl_state);
 970
 971         if (gl->gl_state == LM_ST_EXCLUSIVE && glops->go_sync)
 972                 glops->go_sync(gl, DIO_METADATA | DIO_DATA | DIO_RELEASE);
 973
 974         gfs2_glock_hold(gl);
 975         gl->gl_req_bh = xmote_bh;
 976
 977         lck_ret = gfs2_lm_lock(sdp, gl->gl_lock, gl->gl_state, state, lck_flags);
 978
 979         if (gfs2_assert_withdraw(sdp, !(lck_ret & LM_OUT_ERROR)))
 980                 return;
 981
 982         if (lck_ret & LM_OUT_ASYNC)
 983                 gfs2_assert_warn(sdp, lck_ret == LM_OUT_ASYNC);
 984         else
 985                 xmote_bh(gl, lck_ret);
 986 }
 987
 988 /**
 989  * drop_bh - Called after a lock module unlock completes
 990  * @gl: the glock
 991  * @ret: the return status
 992  *
 993  * Doesn't wake up the process waiting on the struct gfs2_holder (if any)
 994  * Doesn't drop the reference on the glock the top half took out
 995  *
 996  */
 997
 998 static void drop_bh(struct gfs2_glock *gl, unsigned int ret)
 999 {
1000         struct gfs2_sbd *sdp = gl->gl_sbd;
1001         const struct gfs2_glock_operations *glops = gl->gl_ops;
1002         struct gfs2_holder *gh = gl->gl_req_gh;
1003
1004         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1005
1006         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1007         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1008         gfs2_assert_warn(sdp, !ret);
1009
1010         state_change(gl, LM_ST_UNLOCKED);
1011
1012         if (glops->go_inval)
1013                 glops->go_inval(gl, DIO_METADATA | DIO_DATA);
1014
1015         if (gh) {
1016                 spin_lock(&gl->gl_spin);
1017                 list_del_init(&gh->gh_list);
1018                 gh->gh_error = 0;
1019                 spin_unlock(&gl->gl_spin);
1020         }
1021
1022         if (glops->go_drop_bh)
1023                 glops->go_drop_bh(gl);
1024
1025         spin_lock(&gl->gl_spin);
1026         gl->gl_req_gh = NULL;
1027         gl->gl_req_bh = NULL;
1028         clear_bit(GLF_LOCK, &gl->gl_flags);
1029         run_queue(gl);
1030         spin_unlock(&gl->gl_spin);
1031
1032         gfs2_glock_put(gl);
1033
1034         if (gh) {
1035                 if (test_bit(HIF_DEALLOC, &gh->gh_iflags))
1036                         gfs2_holder_put(gh);
1037                 else
1038                         complete(&gh->gh_wait);
1039         }
1040 }
1041
1042 /**
1043  * gfs2_glock_drop_th - call into the lock module to unlock a lock
1044  * @gl: the glock
1045  *
1046  */
1047
1048 void gfs2_glock_drop_th(struct gfs2_glock *gl)
1049 {
1050         struct gfs2_sbd *sdp = gl->gl_sbd;
1051         const struct gfs2_glock_operations *glops = gl->gl_ops;
1052         unsigned int ret;
1053
1054         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1055         gfs2_assert_warn(sdp, queue_empty(gl, &gl->gl_holders));
1056         gfs2_assert_warn(sdp, gl->gl_state != LM_ST_UNLOCKED);
1057
1058         if (gl->gl_state == LM_ST_EXCLUSIVE && glops->go_sync)
1059                 glops->go_sync(gl, DIO_METADATA | DIO_DATA | DIO_RELEASE);
1060
1061         gfs2_glock_hold(gl);
1062         gl->gl_req_bh = drop_bh;
1063
1064         ret = gfs2_lm_unlock(sdp, gl->gl_lock, gl->gl_state);
1065
1066         if (gfs2_assert_withdraw(sdp, !(ret & LM_OUT_ERROR)))
1067                 return;
1068
1069         if (!ret)
1070                 drop_bh(gl, ret);
1071         else
1072                 gfs2_assert_warn(sdp, ret == LM_OUT_ASYNC);
1073 }
1074
1075 /**
1076  * do_cancels - cancel requests for locks stuck waiting on an expire flag
1077  * @gh: the LM_FLAG_PRIORITY holder waiting to acquire the lock
1078  *
1079  * Don't cancel GL_NOCANCEL requests.
1080  */
1081
1082 static void do_cancels(struct gfs2_holder *gh)
1083 {
1084         struct gfs2_glock *gl = gh->gh_gl;
1085
1086         spin_lock(&gl->gl_spin);
1087
1088         while (gl->gl_req_gh != gh &&
1089                !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1090                !list_empty(&gh->gh_list)) {
1091                 if (gl->gl_req_bh && !(gl->gl_req_gh &&
1092                                      (gl->gl_req_gh->gh_flags & GL_NOCANCEL))) {
1093                         spin_unlock(&gl->gl_spin);
1094                         gfs2_lm_cancel(gl->gl_sbd, gl->gl_lock);
1095                         msleep(100);
1096                         spin_lock(&gl->gl_spin);
1097                 } else {
1098                         spin_unlock(&gl->gl_spin);
1099                         msleep(100);
1100                         spin_lock(&gl->gl_spin);
1101                 }
1102         }
1103
1104         spin_unlock(&gl->gl_spin);
1105 }
1106
1107 /**
1108  * glock_wait_internal - wait on a glock acquisition
1109  * @gh: the glock holder
1110  *
1111  * Returns: 0 on success
1112  */
1113
1114 static int glock_wait_internal(struct gfs2_holder *gh)
1115 {
1116         struct gfs2_glock *gl = gh->gh_gl;
1117         struct gfs2_sbd *sdp = gl->gl_sbd;
1118         const struct gfs2_glock_operations *glops = gl->gl_ops;
1119
1120         if (test_bit(HIF_ABORTED, &gh->gh_iflags))
1121                 return -EIO;
1122
1123         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1124                 spin_lock(&gl->gl_spin);
1125                 if (gl->gl_req_gh != gh &&
1126                     !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1127                     !list_empty(&gh->gh_list)) {
1128                         list_del_init(&gh->gh_list);
1129                         gh->gh_error = GLR_TRYFAILED;
1130                         run_queue(gl);
1131                         spin_unlock(&gl->gl_spin);
1132                         return gh->gh_error;
1133                 }
1134                 spin_unlock(&gl->gl_spin);
1135         }
1136
1137         if (gh->gh_flags & LM_FLAG_PRIORITY)
1138                 do_cancels(gh);
1139
1140         wait_for_completion(&gh->gh_wait);
1141
1142         if (gh->gh_error)
1143                 return gh->gh_error;
1144
1145         gfs2_assert_withdraw(sdp, test_bit(HIF_HOLDER, &gh->gh_iflags));
1146         gfs2_assert_withdraw(sdp, relaxed_state_ok(gl->gl_state, gh->gh_state,
1147                                                    gh->gh_flags));
1148
1149         if (test_bit(HIF_FIRST, &gh->gh_iflags)) {
1150                 gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1151
1152                 if (glops->go_lock) {
1153                         gh->gh_error = glops->go_lock(gh);
1154                         if (gh->gh_error) {
1155                                 spin_lock(&gl->gl_spin);
1156                                 list_del_init(&gh->gh_list);
1157                                 spin_unlock(&gl->gl_spin);
1158                         }
1159                 }
1160
1161                 spin_lock(&gl->gl_spin);
1162                 gl->gl_req_gh = NULL;
1163                 gl->gl_req_bh = NULL;
1164                 clear_bit(GLF_LOCK, &gl->gl_flags);
1165                 run_queue(gl);
1166                 spin_unlock(&gl->gl_spin);
1167         }
1168
1169         return gh->gh_error;
1170 }
1171
1172 static inline struct gfs2_holder *
1173 find_holder_by_owner(struct list_head *head, struct task_struct *owner)
1174 {
1175         struct gfs2_holder *gh;
1176
1177         list_for_each_entry(gh, head, gh_list) {
1178                 if (gh->gh_owner == owner)
1179                         return gh;
1180         }
1181
1182         return NULL;
1183 }
1184
1185 /**
1186  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1187  * @gh: the holder structure to add
1188  *
1189  */
1190
1191 static void add_to_queue(struct gfs2_holder *gh)
1192 {
1193         struct gfs2_glock *gl = gh->gh_gl;
1194         struct gfs2_holder *existing;
1195
1196         BUG_ON(!gh->gh_owner);
1197
1198         existing = find_holder_by_owner(&gl->gl_holders, gh->gh_owner);
1199         if (existing) {
1200                 print_symbol(KERN_WARNING "original: %s\n", existing->gh_ip);
1201                 printk(KERN_INFO "pid : %d\n", existing->gh_owner->pid);
1202                 printk(KERN_INFO "lock type : %d lock state : %d\n",
1203                                 existing->gh_gl->gl_name.ln_type, existing->gh_gl->gl_state);
1204                 print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
1205                 printk(KERN_INFO "pid : %d\n", gh->gh_owner->pid);
1206                 printk(KERN_INFO "lock type : %d lock state : %d\n",
1207                                 gl->gl_name.ln_type, gl->gl_state);
1208                 BUG();
1209         }
1210
1211         existing = find_holder_by_owner(&gl->gl_waiters3, gh->gh_owner);
1212         if (existing) {
1213                 print_symbol(KERN_WARNING "original: %s\n", existing->gh_ip);
1214                 print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
1215                 BUG();
1216         }
1217
1218         if (gh->gh_flags & LM_FLAG_PRIORITY)
1219                 list_add(&gh->gh_list, &gl->gl_waiters3);
1220         else
1221                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1222 }
1223
1224 /**
1225  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1226  * @gh: the holder structure
1227  *
1228  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1229  *
1230  * Returns: 0, GLR_TRYFAILED, or errno on failure
1231  */
1232
1233 int gfs2_glock_nq(struct gfs2_holder *gh)
1234 {
1235         struct gfs2_glock *gl = gh->gh_gl;
1236         struct gfs2_sbd *sdp = gl->gl_sbd;
1237         int error = 0;
1238
1239 restart:
1240         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
1241                 set_bit(HIF_ABORTED, &gh->gh_iflags);
1242                 return -EIO;
1243         }
1244
1245         set_bit(HIF_PROMOTE, &gh->gh_iflags);
1246
1247         spin_lock(&gl->gl_spin);
1248         add_to_queue(gh);
1249         run_queue(gl);
1250         spin_unlock(&gl->gl_spin);
1251
1252         if (!(gh->gh_flags & GL_ASYNC)) {
1253                 error = glock_wait_internal(gh);
1254                 if (error == GLR_CANCELED) {
1255                         msleep(100);
1256                         goto restart;
1257                 }
1258         }
1259
1260         clear_bit(GLF_PREFETCH, &gl->gl_flags);
1261
1262         if (error == GLR_TRYFAILED && (gh->gh_flags & GL_DUMP))
1263                 dump_glock(gl);
1264
1265         return error;
1266 }
1267
1268 /**
1269  * gfs2_glock_poll - poll to see if an async request has been completed
1270  * @gh: the holder
1271  *
1272  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1273  */
1274
1275 int gfs2_glock_poll(struct gfs2_holder *gh)
1276 {
1277         struct gfs2_glock *gl = gh->gh_gl;
1278         int ready = 0;
1279
1280         spin_lock(&gl->gl_spin);
1281
1282         if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1283                 ready = 1;
1284         else if (list_empty(&gh->gh_list)) {
1285                 if (gh->gh_error == GLR_CANCELED) {
1286                         spin_unlock(&gl->gl_spin);
1287                         msleep(100);
1288                         if (gfs2_glock_nq(gh))
1289                                 return 1;
1290                         return 0;
1291                 } else
1292                         ready = 1;
1293         }
1294
1295         spin_unlock(&gl->gl_spin);
1296
1297         return ready;
1298 }
1299
1300 /**
1301  * gfs2_glock_wait - wait for a lock acquisition that ended in a GLR_ASYNC
1302  * @gh: the holder structure
1303  *
1304  * Returns: 0, GLR_TRYFAILED, or errno on failure
1305  */
1306
1307 int gfs2_glock_wait(struct gfs2_holder *gh)
1308 {
1309         int error;
1310
1311         error = glock_wait_internal(gh);
1312         if (error == GLR_CANCELED) {
1313                 msleep(100);
1314                 gh->gh_flags &= ~GL_ASYNC;
1315                 error = gfs2_glock_nq(gh);
1316         }
1317
1318         return error;
1319 }
1320
1321 /**
1322  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1323  * @gh: the glock holder
1324  *
1325  */
1326
1327 void gfs2_glock_dq(struct gfs2_holder *gh)
1328 {
1329         struct gfs2_glock *gl = gh->gh_gl;
1330         const struct gfs2_glock_operations *glops = gl->gl_ops;
1331
1332         if (gh->gh_flags & GL_NOCACHE)
1333                 handle_callback(gl, LM_ST_UNLOCKED);
1334
1335         gfs2_glmutex_lock(gl);
1336
1337         spin_lock(&gl->gl_spin);
1338         list_del_init(&gh->gh_list);
1339
1340         if (list_empty(&gl->gl_holders)) {
1341                 spin_unlock(&gl->gl_spin);
1342
1343                 if (glops->go_unlock)
1344                         glops->go_unlock(gh);
1345
1346                 gl->gl_stamp = jiffies;
1347
1348                 spin_lock(&gl->gl_spin);
1349         }
1350
1351         clear_bit(GLF_LOCK, &gl->gl_flags);
1352         run_queue(gl);
1353         spin_unlock(&gl->gl_spin);
1354 }
1355
1356 /**
1357  * gfs2_glock_prefetch - Try to prefetch a glock
1358  * @gl: the glock
1359  * @state: the state to prefetch in
1360  * @flags: flags passed to go_xmote_th()
1361  *
1362  */
1363
1364 static void gfs2_glock_prefetch(struct gfs2_glock *gl, unsigned int state,
1365                                 int flags)
1366 {
1367         const struct gfs2_glock_operations *glops = gl->gl_ops;
1368
1369         spin_lock(&gl->gl_spin);
1370
1371         if (test_bit(GLF_LOCK, &gl->gl_flags) || !list_empty(&gl->gl_holders) ||
1372             !list_empty(&gl->gl_waiters1) || !list_empty(&gl->gl_waiters2) ||
1373             !list_empty(&gl->gl_waiters3) ||
1374             relaxed_state_ok(gl->gl_state, state, flags)) {
1375                 spin_unlock(&gl->gl_spin);
1376                 return;
1377         }
1378
1379         set_bit(GLF_PREFETCH, &gl->gl_flags);
1380         set_bit(GLF_LOCK, &gl->gl_flags);
1381         spin_unlock(&gl->gl_spin);
1382
1383         glops->go_xmote_th(gl, state, flags);
1384 }
1385
1386 static void greedy_work(void *data)
1387 {
1388         struct greedy *gr = data;
1389         struct gfs2_holder *gh = &gr->gr_gh;
1390         struct gfs2_glock *gl = gh->gh_gl;
1391         const struct gfs2_glock_operations *glops = gl->gl_ops;
1392
1393         clear_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1394
1395         if (glops->go_greedy)
1396                 glops->go_greedy(gl);
1397
1398         spin_lock(&gl->gl_spin);
1399
1400         if (list_empty(&gl->gl_waiters2)) {
1401                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1402                 spin_unlock(&gl->gl_spin);
1403                 gfs2_holder_uninit(gh);
1404                 kfree(gr);
1405         } else {
1406                 gfs2_glock_hold(gl);
1407                 list_add_tail(&gh->gh_list, &gl->gl_waiters2);
1408                 run_queue(gl);
1409                 spin_unlock(&gl->gl_spin);
1410                 gfs2_glock_put(gl);
1411         }
1412 }
1413
1414 /**
1415  * gfs2_glock_be_greedy -
1416  * @gl:
1417  * @time:
1418  *
1419  * Returns: 0 if go_greedy will be called, 1 otherwise
1420  */
1421
1422 int gfs2_glock_be_greedy(struct gfs2_glock *gl, unsigned int time)
1423 {
1424         struct greedy *gr;
1425         struct gfs2_holder *gh;
1426
1427         if (!time || gl->gl_sbd->sd_args.ar_localcaching ||
1428             test_and_set_bit(GLF_GREEDY, &gl->gl_flags))
1429                 return 1;
1430
1431         gr = kmalloc(sizeof(struct greedy), GFP_KERNEL);
1432         if (!gr) {
1433                 clear_bit(GLF_GREEDY, &gl->gl_flags);
1434                 return 1;
1435         }
1436         gh = &gr->gr_gh;
1437
1438         gfs2_holder_init(gl, 0, 0, gh);
1439         set_bit(HIF_GREEDY, &gh->gh_iflags);
1440         INIT_WORK(&gr->gr_work, greedy_work, gr);
1441
1442         set_bit(GLF_SKIP_WAITERS2, &gl->gl_flags);
1443         schedule_delayed_work(&gr->gr_work, time);
1444
1445         return 0;
1446 }
1447
1448 /**
1449  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1450  * @gh: the holder structure
1451  *
1452  */
1453
1454 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1455 {
1456         gfs2_glock_dq(gh);
1457         gfs2_holder_uninit(gh);
1458 }
1459
1460 /**
1461  * gfs2_glock_nq_num - acquire a glock based on lock number
1462  * @sdp: the filesystem
1463  * @number: the lock number
1464  * @glops: the glock operations for the type of glock
1465  * @state: the state to acquire the glock in
1466  * @flags: modifier flags for the aquisition
1467  * @gh: the struct gfs2_holder
1468  *
1469  * Returns: errno
1470  */
1471
1472 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1473                       const struct gfs2_glock_operations *glops,
1474                       unsigned int state, int flags, struct gfs2_holder *gh)
1475 {
1476         struct gfs2_glock *gl;
1477         int error;
1478
1479         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1480         if (!error) {
1481                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1482                 gfs2_glock_put(gl);
1483         }
1484
1485         return error;
1486 }
1487
1488 /**
1489  * glock_compare - Compare two struct gfs2_glock structures for sorting
1490  * @arg_a: the first structure
1491  * @arg_b: the second structure
1492  *
1493  */
1494
1495 static int glock_compare(const void *arg_a, const void *arg_b)
1496 {
1497         struct gfs2_holder *gh_a = *(struct gfs2_holder **)arg_a;
1498         struct gfs2_holder *gh_b = *(struct gfs2_holder **)arg_b;
1499         struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1500         struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1501         int ret = 0;
1502
1503         if (a->ln_number > b->ln_number)
1504                 ret = 1;
1505         else if (a->ln_number < b->ln_number)
1506                 ret = -1;
1507         else {
1508                 if (gh_a->gh_state == LM_ST_SHARED &&
1509                     gh_b->gh_state == LM_ST_EXCLUSIVE)
1510                         ret = 1;
1511                 else if (!(gh_a->gh_flags & GL_LOCAL_EXCL) &&
1512                          (gh_b->gh_flags & GL_LOCAL_EXCL))
1513                         ret = 1;
1514         }
1515
1516         return ret;
1517 }
1518
1519 /**
1520  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1521  * @num_gh: the number of structures
1522  * @ghs: an array of struct gfs2_holder structures
1523  *
1524  * Returns: 0 on success (all glocks acquired),
1525  *          errno on failure (no glocks acquired)
1526  */
1527
1528 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1529                      struct gfs2_holder **p)
1530 {
1531         unsigned int x;
1532         int error = 0;
1533
1534         for (x = 0; x < num_gh; x++)
1535                 p[x] = &ghs[x];
1536
1537         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1538
1539         for (x = 0; x < num_gh; x++) {
1540                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1541
1542                 error = gfs2_glock_nq(p[x]);
1543                 if (error) {
1544                         while (x--)
1545                                 gfs2_glock_dq(p[x]);
1546                         break;
1547                 }
1548         }
1549
1550         return error;
1551 }
1552
1553 /**
1554  * gfs2_glock_nq_m - acquire multiple glocks
1555  * @num_gh: the number of structures
1556  * @ghs: an array of struct gfs2_holder structures
1557  *
1558  * Figure out how big an impact this function has.  Either:
1559  * 1) Replace this code with code that calls gfs2_glock_prefetch()
1560  * 2) Forget async stuff and just call nq_m_sync()
1561  * 3) Leave it like it is
1562  *
1563  * Returns: 0 on success (all glocks acquired),
1564  *          errno on failure (no glocks acquired)
1565  */
1566
1567 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1568 {
1569         int *e;
1570         unsigned int x;
1571         int borked = 0, serious = 0;
1572         int error = 0;
1573
1574         if (!num_gh)
1575                 return 0;
1576
1577         if (num_gh == 1) {
1578                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1579                 return gfs2_glock_nq(ghs);
1580         }
1581
1582         e = kcalloc(num_gh, sizeof(struct gfs2_holder *), GFP_KERNEL);
1583         if (!e)
1584                 return -ENOMEM;
1585
1586         for (x = 0; x < num_gh; x++) {
1587                 ghs[x].gh_flags |= LM_FLAG_TRY | GL_ASYNC;
1588                 error = gfs2_glock_nq(&ghs[x]);
1589                 if (error) {
1590                         borked = 1;
1591                         serious = error;
1592                         num_gh = x;
1593                         break;
1594                 }
1595         }
1596
1597         for (x = 0; x < num_gh; x++) {
1598                 error = e[x] = glock_wait_internal(&ghs[x]);
1599                 if (error) {
1600                         borked = 1;
1601                         if (error != GLR_TRYFAILED && error != GLR_CANCELED)
1602                                 serious = error;
1603                 }
1604         }
1605
1606         if (!borked) {
1607                 kfree(e);
1608                 return 0;
1609         }
1610
1611         for (x = 0; x < num_gh; x++)
1612                 if (!e[x])
1613                         gfs2_glock_dq(&ghs[x]);
1614
1615         if (serious)
1616                 error = serious;
1617         else {
1618                 for (x = 0; x < num_gh; x++)
1619                         gfs2_holder_reinit(ghs[x].gh_state, ghs[x].gh_flags,
1620                                           &ghs[x]);
1621                 error = nq_m_sync(num_gh, ghs, (struct gfs2_holder **)e);
1622         }
1623
1624         kfree(e);
1625
1626         return error;
1627 }
1628
1629 /**
1630  * gfs2_glock_dq_m - release multiple glocks
1631  * @num_gh: the number of structures
1632  * @ghs: an array of struct gfs2_holder structures
1633  *
1634  */
1635
1636 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1637 {
1638         unsigned int x;
1639
1640         for (x = 0; x < num_gh; x++)
1641                 gfs2_glock_dq(&ghs[x]);
1642 }
1643
1644 /**
1645  * gfs2_glock_dq_uninit_m - release multiple glocks
1646  * @num_gh: the number of structures
1647  * @ghs: an array of struct gfs2_holder structures
1648  *
1649  */
1650
1651 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1652 {
1653         unsigned int x;
1654
1655         for (x = 0; x < num_gh; x++)
1656                 gfs2_glock_dq_uninit(&ghs[x]);
1657 }
1658
1659 /**
1660  * gfs2_glock_prefetch_num - prefetch a glock based on lock number
1661  * @sdp: the filesystem
1662  * @number: the lock number
1663  * @glops: the glock operations for the type of glock
1664  * @state: the state to acquire the glock in
1665  * @flags: modifier flags for the aquisition
1666  *
1667  * Returns: errno
1668  */
1669
1670 void gfs2_glock_prefetch_num(struct gfs2_sbd *sdp, u64 number,
1671                              const struct gfs2_glock_operations *glops,
1672                              unsigned int state, int flags)
1673 {
1674         struct gfs2_glock *gl;
1675         int error;
1676
1677         if (atomic_read(&sdp->sd_reclaim_count) <
1678             gfs2_tune_get(sdp, gt_reclaim_limit)) {
1679                 error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1680                 if (!error) {
1681                         gfs2_glock_prefetch(gl, state, flags);
1682                         gfs2_glock_put(gl);
1683                 }
1684         }
1685 }
1686
1687 /**
1688  * gfs2_lvb_hold - attach a LVB from a glock
1689  * @gl: The glock in question
1690  *
1691  */
1692
1693 int gfs2_lvb_hold(struct gfs2_glock *gl)
1694 {
1695         int error;
1696
1697         gfs2_glmutex_lock(gl);
1698
1699         if (!atomic_read(&gl->gl_lvb_count)) {
1700                 error = gfs2_lm_hold_lvb(gl->gl_sbd, gl->gl_lock, &gl->gl_lvb);
1701                 if (error) {
1702                         gfs2_glmutex_unlock(gl);
1703                         return error;
1704                 }
1705                 gfs2_glock_hold(gl);
1706         }
1707         atomic_inc(&gl->gl_lvb_count);
1708
1709         gfs2_glmutex_unlock(gl);
1710
1711         return 0;
1712 }
1713
1714 /**
1715  * gfs2_lvb_unhold - detach a LVB from a glock
1716  * @gl: The glock in question
1717  *
1718  */
1719
1720 void gfs2_lvb_unhold(struct gfs2_glock *gl)
1721 {
1722         gfs2_glock_hold(gl);
1723         gfs2_glmutex_lock(gl);
1724
1725         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count) > 0);
1726         if (atomic_dec_and_test(&gl->gl_lvb_count)) {
1727                 gfs2_lm_unhold_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1728                 gl->gl_lvb = NULL;
1729                 gfs2_glock_put(gl);
1730         }
1731
1732         gfs2_glmutex_unlock(gl);
1733         gfs2_glock_put(gl);
1734 }
1735
1736 static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
1737                         unsigned int state)
1738 {
1739         struct gfs2_glock *gl;
1740
1741         gl = gfs2_glock_find(sdp, name);
1742         if (!gl)
1743                 return;
1744
1745         if (gl->gl_ops->go_callback)
1746                 gl->gl_ops->go_callback(gl, state);
1747         handle_callback(gl, state);
1748
1749         spin_lock(&gl->gl_spin);
1750         run_queue(gl);
1751         spin_unlock(&gl->gl_spin);
1752
1753         gfs2_glock_put(gl);
1754 }
1755
1756 /**
1757  * gfs2_glock_cb - Callback used by locking module
1758  * @sdp: Pointer to the superblock
1759  * @type: Type of callback
1760  * @data: Type dependent data pointer
1761  *
1762  * Called by the locking module when it wants to tell us something.
1763  * Either we need to drop a lock, one of our ASYNC requests completed, or
1764  * a journal from another client needs to be recovered.
1765  */
1766
1767 void gfs2_glock_cb(void *cb_data, unsigned int type, void *data)
1768 {
1769         struct gfs2_sbd *sdp = cb_data;
1770
1771         switch (type) {
1772         case LM_CB_NEED_E:
1773                 blocking_cb(sdp, data, LM_ST_UNLOCKED);
1774                 return;
1775
1776         case LM_CB_NEED_D:
1777                 blocking_cb(sdp, data, LM_ST_DEFERRED);
1778                 return;
1779
1780         case LM_CB_NEED_S:
1781                 blocking_cb(sdp, data, LM_ST_SHARED);
1782                 return;
1783
1784         case LM_CB_ASYNC: {
1785                 struct lm_async_cb *async = data;
1786                 struct gfs2_glock *gl;
1787
1788                 gl = gfs2_glock_find(sdp, &async->lc_name);
1789                 if (gfs2_assert_warn(sdp, gl))
1790                         return;
1791                 if (!gfs2_assert_warn(sdp, gl->gl_req_bh))
1792                         gl->gl_req_bh(gl, async->lc_ret);
1793                 gfs2_glock_put(gl);
1794                 return;
1795         }
1796
1797         case LM_CB_NEED_RECOVERY:
1798                 gfs2_jdesc_make_dirty(sdp, *(unsigned int *)data);
1799                 if (sdp->sd_recoverd_process)
1800                         wake_up_process(sdp->sd_recoverd_process);
1801                 return;
1802
1803         case LM_CB_DROPLOCKS:
1804                 gfs2_gl_hash_clear(sdp, NO_WAIT);
1805                 gfs2_quota_scan(sdp);
1806                 return;
1807
1808         default:
1809                 gfs2_assert_warn(sdp, 0);
1810                 return;
1811         }
1812 }
1813
1814 /**
1815  * gfs2_iopen_go_callback - Try to kick the inode/vnode associated with an
1816  *                          iopen glock from memory
1817  * @io_gl: the iopen glock
1818  * @state: the state into which the glock should be put
1819  *
1820  */
1821
1822 void gfs2_iopen_go_callback(struct gfs2_glock *io_gl, unsigned int state)
1823 {
1824
1825         if (state != LM_ST_UNLOCKED)
1826                 return;
1827         /* FIXME: remove this? */
1828 }
1829
1830 /**
1831  * demote_ok - Check to see if it's ok to unlock a glock
1832  * @gl: the glock
1833  *
1834  * Returns: 1 if it's ok
1835  */
1836
1837 static int demote_ok(struct gfs2_glock *gl)
1838 {
1839         struct gfs2_sbd *sdp = gl->gl_sbd;
1840         const struct gfs2_glock_operations *glops = gl->gl_ops;
1841         int demote = 1;
1842
1843         if (test_bit(GLF_STICKY, &gl->gl_flags))
1844                 demote = 0;
1845         else if (test_bit(GLF_PREFETCH, &gl->gl_flags))
1846                 demote = time_after_eq(jiffies, gl->gl_stamp +
1847                                     gfs2_tune_get(sdp, gt_prefetch_secs) * HZ);
1848         else if (glops->go_demote_ok)
1849                 demote = glops->go_demote_ok(gl);
1850
1851         return demote;
1852 }
1853
1854 /**
1855  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
1856  * @gl: the glock
1857  *
1858  */
1859
1860 void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
1861 {
1862         struct gfs2_sbd *sdp = gl->gl_sbd;
1863
1864         spin_lock(&sdp->sd_reclaim_lock);
1865         if (list_empty(&gl->gl_reclaim)) {
1866                 gfs2_glock_hold(gl);
1867                 list_add(&gl->gl_reclaim, &sdp->sd_reclaim_list);
1868                 atomic_inc(&sdp->sd_reclaim_count);
1869         }
1870         spin_unlock(&sdp->sd_reclaim_lock);
1871
1872         wake_up(&sdp->sd_reclaim_wq);
1873 }
1874
1875 /**
1876  * gfs2_reclaim_glock - process the next glock on the filesystem's reclaim list
1877  * @sdp: the filesystem
1878  *
1879  * Called from gfs2_glockd() glock reclaim daemon, or when promoting a
1880  * different glock and we notice that there are a lot of glocks in the
1881  * reclaim list.
1882  *
1883  */
1884
1885 void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
1886 {
1887         struct gfs2_glock *gl;
1888
1889         spin_lock(&sdp->sd_reclaim_lock);
1890         if (list_empty(&sdp->sd_reclaim_list)) {
1891                 spin_unlock(&sdp->sd_reclaim_lock);
1892                 return;
1893         }
1894         gl = list_entry(sdp->sd_reclaim_list.next,
1895                         struct gfs2_glock, gl_reclaim);
1896         list_del_init(&gl->gl_reclaim);
1897         spin_unlock(&sdp->sd_reclaim_lock);
1898
1899         atomic_dec(&sdp->sd_reclaim_count);
1900         atomic_inc(&sdp->sd_reclaimed);
1901
1902         if (gfs2_glmutex_trylock(gl)) {
1903                 if (queue_empty(gl, &gl->gl_holders) &&
1904                     gl->gl_state != LM_ST_UNLOCKED && demote_ok(gl))
1905                         handle_callback(gl, LM_ST_UNLOCKED);
1906                 gfs2_glmutex_unlock(gl);
1907         }
1908
1909         gfs2_glock_put(gl);
1910 }
1911
1912 /**
1913  * examine_bucket - Call a function for glock in a hash bucket
1914  * @examiner: the function
1915  * @sdp: the filesystem
1916  * @bucket: the bucket
1917  *
1918  * Returns: 1 if the bucket has entries
1919  */
1920
1921 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
1922                           unsigned int hash)
1923 {
1924         struct glock_plug plug;
1925         struct list_head *tmp;
1926         struct gfs2_glock *gl;
1927         int entries;
1928
1929         /* Add "plug" to end of bucket list, work back up list from there */
1930         memset(&plug.gl_flags, 0, sizeof(unsigned long));
1931         set_bit(GLF_PLUG, &plug.gl_flags);
1932
1933         write_lock(gl_lock_addr(hash));
1934         list_add(&plug.gl_list, &gl_hash_table[hash].hb_list);
1935         write_unlock(gl_lock_addr(hash));
1936
1937         for (;;) {
1938                 write_lock(gl_lock_addr(hash));
1939
1940                 for (;;) {
1941                         tmp = plug.gl_list.next;
1942
1943                         if (tmp == &gl_hash_table[hash].hb_list) {
1944                                 list_del(&plug.gl_list);
1945                                 entries = !list_empty(&gl_hash_table[hash].hb_list);
1946                                 write_unlock(gl_lock_addr(hash));
1947                                 return entries;
1948                         }
1949                         gl = list_entry(tmp, struct gfs2_glock, gl_list);
1950
1951                         /* Move plug up list */
1952                         list_move(&plug.gl_list, &gl->gl_list);
1953
1954                         if (test_bit(GLF_PLUG, &gl->gl_flags))
1955                                 continue;
1956                         if (gl->gl_sbd != sdp)
1957                                 continue;
1958
1959                         /* examiner() must glock_put() */
1960                         gfs2_glock_hold(gl);
1961
1962                         break;
1963                 }
1964
1965                 write_unlock(gl_lock_addr(hash));
1966
1967                 examiner(gl);
1968         }
1969 }
1970
1971 /**
1972  * scan_glock - look at a glock and see if we can reclaim it
1973  * @gl: the glock to look at
1974  *
1975  */
1976
1977 static void scan_glock(struct gfs2_glock *gl)
1978 {
1979         if (gl->gl_ops == &gfs2_inode_glops)
1980                 goto out;
1981
1982         if (gfs2_glmutex_trylock(gl)) {
1983                 if (queue_empty(gl, &gl->gl_holders) &&
1984                     gl->gl_state != LM_ST_UNLOCKED &&
1985                     demote_ok(gl))
1986                         goto out_schedule;
1987                 gfs2_glmutex_unlock(gl);
1988         }
1989 out:
1990         gfs2_glock_put(gl);
1991         return;
1992
1993 out_schedule:
1994         gfs2_glmutex_unlock(gl);
1995         gfs2_glock_schedule_for_reclaim(gl);
1996         gfs2_glock_put(gl);
1997 }
1998
1999 /**
2000  * gfs2_scand_internal - Look for glocks and inodes to toss from memory
2001  * @sdp: the filesystem
2002  *
2003  */
2004
2005 void gfs2_scand_internal(struct gfs2_sbd *sdp)
2006 {
2007         unsigned int x;
2008
2009         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2010                 examine_bucket(scan_glock, sdp, x);
2011                 cond_resched();
2012         }
2013 }
2014
2015 /**
2016  * clear_glock - look at a glock and see if we can free it from glock cache
2017  * @gl: the glock to look at
2018  *
2019  */
2020
2021 static void clear_glock(struct gfs2_glock *gl)
2022 {
2023         struct gfs2_sbd *sdp = gl->gl_sbd;
2024         int released;
2025
2026         spin_lock(&sdp->sd_reclaim_lock);
2027         if (!list_empty(&gl->gl_reclaim)) {
2028                 list_del_init(&gl->gl_reclaim);
2029                 atomic_dec(&sdp->sd_reclaim_count);
2030                 spin_unlock(&sdp->sd_reclaim_lock);
2031                 released = gfs2_glock_put(gl);
2032                 gfs2_assert(sdp, !released);
2033         } else {
2034                 spin_unlock(&sdp->sd_reclaim_lock);
2035         }
2036
2037         if (gfs2_glmutex_trylock(gl)) {
2038                 if (queue_empty(gl, &gl->gl_holders) &&
2039                     gl->gl_state != LM_ST_UNLOCKED)
2040                         handle_callback(gl, LM_ST_UNLOCKED);
2041
2042                 gfs2_glmutex_unlock(gl);
2043         }
2044
2045         gfs2_glock_put(gl);
2046 }
2047
2048 /**
2049  * gfs2_gl_hash_clear - Empty out the glock hash table
2050  * @sdp: the filesystem
2051  * @wait: wait until it's all gone
2052  *
2053  * Called when unmounting the filesystem, or when inter-node lock manager
2054  * requests DROPLOCKS because it is running out of capacity.
2055  */
2056
2057 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
2058 {
2059         unsigned long t;
2060         unsigned int x;
2061         int cont;
2062
2063         t = jiffies;
2064
2065         for (;;) {
2066                 cont = 0;
2067
2068                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
2069                         if (examine_bucket(clear_glock, sdp, x))
2070                                 cont = 1;
2071
2072                 if (!wait || !cont)
2073                         break;
2074
2075                 if (time_after_eq(jiffies,
2076                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
2077                         fs_warn(sdp, "Unmount seems to be stalled. "
2078                                      "Dumping lock state...\n");
2079                         gfs2_dump_lockstate(sdp);
2080                         t = jiffies;
2081                 }
2082
2083                 invalidate_inodes(sdp->sd_vfs);
2084                 msleep(10);
2085         }
2086 }
2087
2088 /*
2089  *  Diagnostic routines to help debug distributed deadlock
2090  */
2091
2092 /**
2093  * dump_holder - print information about a glock holder
2094  * @str: a string naming the type of holder
2095  * @gh: the glock holder
2096  *
2097  * Returns: 0 on success, -ENOBUFS when we run out of space
2098  */
2099
2100 static int dump_holder(char *str, struct gfs2_holder *gh)
2101 {
2102         unsigned int x;
2103         int error = -ENOBUFS;
2104
2105         printk(KERN_INFO "  %s\n", str);
2106         printk(KERN_INFO "    owner = %ld\n",
2107                    (gh->gh_owner) ? (long)gh->gh_owner->pid : -1);
2108         printk(KERN_INFO "    gh_state = %u\n", gh->gh_state);
2109         printk(KERN_INFO "    gh_flags =");
2110         for (x = 0; x < 32; x++)
2111                 if (gh->gh_flags & (1 << x))
2112                         printk(" %u", x);
2113         printk(" \n");
2114         printk(KERN_INFO "    error = %d\n", gh->gh_error);
2115         printk(KERN_INFO "    gh_iflags =");
2116         for (x = 0; x < 32; x++)
2117                 if (test_bit(x, &gh->gh_iflags))
2118                         printk(" %u", x);
2119         printk(" \n");
2120         print_symbol(KERN_INFO "    initialized at: %s\n", gh->gh_ip);
2121
2122         error = 0;
2123
2124         return error;
2125 }
2126
2127 /**
2128  * dump_inode - print information about an inode
2129  * @ip: the inode
2130  *
2131  * Returns: 0 on success, -ENOBUFS when we run out of space
2132  */
2133
2134 static int dump_inode(struct gfs2_inode *ip)
2135 {
2136         unsigned int x;
2137         int error = -ENOBUFS;
2138
2139         printk(KERN_INFO "  Inode:\n");
2140         printk(KERN_INFO "    num = %llu %llu\n",
2141                     (unsigned long long)ip->i_num.no_formal_ino,
2142                     (unsigned long long)ip->i_num.no_addr);
2143         printk(KERN_INFO "    type = %u\n", IF2DT(ip->i_di.di_mode));
2144         printk(KERN_INFO "    i_flags =");
2145         for (x = 0; x < 32; x++)
2146                 if (test_bit(x, &ip->i_flags))
2147                         printk(" %u", x);
2148         printk(" \n");
2149
2150         error = 0;
2151
2152         return error;
2153 }
2154
2155 /**
2156  * dump_glock - print information about a glock
2157  * @gl: the glock
2158  * @count: where we are in the buffer
2159  *
2160  * Returns: 0 on success, -ENOBUFS when we run out of space
2161  */
2162
2163 static int dump_glock(struct gfs2_glock *gl)
2164 {
2165         struct gfs2_holder *gh;
2166         unsigned int x;
2167         int error = -ENOBUFS;
2168
2169         spin_lock(&gl->gl_spin);
2170
2171         printk(KERN_INFO "Glock 0x%p (%u, %llu)\n", gl, gl->gl_name.ln_type,
2172                (unsigned long long)gl->gl_name.ln_number);
2173         printk(KERN_INFO "  gl_flags =");
2174         for (x = 0; x < 32; x++) {
2175                 if (test_bit(x, &gl->gl_flags))
2176                         printk(" %u", x);
2177         }
2178         printk(" \n");
2179         printk(KERN_INFO "  gl_ref = %d\n", atomic_read(&gl->gl_ref.refcount));
2180         printk(KERN_INFO "  gl_state = %u\n", gl->gl_state);
2181         printk(KERN_INFO "  gl_owner = %s\n", gl->gl_owner->comm);
2182         print_symbol(KERN_INFO "  gl_ip = %s\n", gl->gl_ip);
2183         printk(KERN_INFO "  req_gh = %s\n", (gl->gl_req_gh) ? "yes" : "no");
2184         printk(KERN_INFO "  req_bh = %s\n", (gl->gl_req_bh) ? "yes" : "no");
2185         printk(KERN_INFO "  lvb_count = %d\n", atomic_read(&gl->gl_lvb_count));
2186         printk(KERN_INFO "  object = %s\n", (gl->gl_object) ? "yes" : "no");
2187         printk(KERN_INFO "  le = %s\n",
2188                    (list_empty(&gl->gl_le.le_list)) ? "no" : "yes");
2189         printk(KERN_INFO "  reclaim = %s\n",
2190                     (list_empty(&gl->gl_reclaim)) ? "no" : "yes");
2191         if (gl->gl_aspace)
2192                 printk(KERN_INFO "  aspace = 0x%p nrpages = %lu\n", gl->gl_aspace,
2193                        gl->gl_aspace->i_mapping->nrpages);
2194         else
2195                 printk(KERN_INFO "  aspace = no\n");
2196         printk(KERN_INFO "  ail = %d\n", atomic_read(&gl->gl_ail_count));
2197         if (gl->gl_req_gh) {
2198                 error = dump_holder("Request", gl->gl_req_gh);
2199                 if (error)
2200                         goto out;
2201         }
2202         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
2203                 error = dump_holder("Holder", gh);
2204                 if (error)
2205                         goto out;
2206         }
2207         list_for_each_entry(gh, &gl->gl_waiters1, gh_list) {
2208                 error = dump_holder("Waiter1", gh);
2209                 if (error)
2210                         goto out;
2211         }
2212         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
2213                 error = dump_holder("Waiter2", gh);
2214                 if (error)
2215                         goto out;
2216         }
2217         list_for_each_entry(gh, &gl->gl_waiters3, gh_list) {
2218                 error = dump_holder("Waiter3", gh);
2219                 if (error)
2220                         goto out;
2221         }
2222         if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object) {
2223                 if (!test_bit(GLF_LOCK, &gl->gl_flags) &&
2224                     list_empty(&gl->gl_holders)) {
2225                         error = dump_inode(gl->gl_object);
2226                         if (error)
2227                                 goto out;
2228                 } else {
2229                         error = -ENOBUFS;
2230                         printk(KERN_INFO "  Inode: busy\n");
2231                 }
2232         }
2233
2234         error = 0;
2235
2236 out:
2237         spin_unlock(&gl->gl_spin);
2238         return error;
2239 }
2240
2241 /**
2242  * gfs2_dump_lockstate - print out the current lockstate
2243  * @sdp: the filesystem
2244  * @ub: the buffer to copy the information into
2245  *
2246  * If @ub is NULL, dump the lockstate to the console.
2247  *
2248  */
2249
2250 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
2251 {
2252         struct gfs2_glock *gl;
2253         unsigned int x;
2254         int error = 0;
2255
2256         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2257
2258                 read_lock(gl_lock_addr(x));
2259
2260                 list_for_each_entry(gl, &gl_hash_table[x].hb_list, gl_list) {
2261                         if (test_bit(GLF_PLUG, &gl->gl_flags))
2262                                 continue;
2263                         if (gl->gl_sbd != sdp)
2264                                 continue;
2265
2266                         error = dump_glock(gl);
2267                         if (error)
2268                                 break;
2269                 }
2270
2271                 read_unlock(gl_lock_addr(x));
2272
2273                 if (error)
2274                         break;
2275         }
2276
2277
2278         return error;
2279 }
2280
2281 int __init gfs2_glock_init(void)
2282 {
2283         unsigned i;
2284         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
2285                 INIT_LIST_HEAD(&gl_hash_table[i].hb_list);
2286         }
2287 #ifdef GL_HASH_LOCK_SZ
2288         for(i = 0; i < GL_HASH_LOCK_SZ; i++) {
2289                 rwlock_init(&gl_hash_locks[i]);
2290         }
2291 #endif
2292         return 0;
2293 }
2294