fs/gfs2/glock.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   5  */
   6
   7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/sched.h>
  10 #include <linux/slab.h>
  11 #include <linux/spinlock.h>
  12 #include <linux/buffer_head.h>
  13 #include <linux/delay.h>
  14 #include <linux/sort.h>
  15 #include <linux/hash.h>
  16 #include <linux/jhash.h>
  17 #include <linux/kallsyms.h>
  18 #include <linux/gfs2_ondisk.h>
  19 #include <linux/list.h>
  20 #include <linux/wait.h>
  21 #include <linux/module.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/debugfs.h>
  25 #include <linux/kthread.h>
  26 #include <linux/freezer.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/jiffies.h>
  29 #include <linux/rcupdate.h>
  30 #include <linux/rculist_bl.h>
  31 #include <linux/bit_spinlock.h>
  32 #include <linux/percpu.h>
  33 #include <linux/list_sort.h>
  34 #include <linux/lockref.h>
  35 #include <linux/rhashtable.h>
  36
  37 #include "gfs2.h"
  38 #include "incore.h"
  39 #include "glock.h"
  40 #include "glops.h"
  41 #include "inode.h"
  42 #include "lops.h"
  43 #include "meta_io.h"
  44 #include "quota.h"
  45 #include "super.h"
  46 #include "util.h"
  47 #include "bmap.h"
  48 #define CREATE_TRACE_POINTS
  49 #include "trace_gfs2.h"
  50
  51 struct gfs2_glock_iter {
  52         struct gfs2_sbd *sdp;           /* incore superblock           */
  53         struct rhashtable_iter hti;     /* rhashtable iterator         */
  54         struct gfs2_glock *gl;          /* current glock struct        */
  55         loff_t last_pos;                /* last position               */
  56 };
  57
  58 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  59
  60 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  61
  62 static struct dentry *gfs2_root;
  63 static struct workqueue_struct *glock_workqueue;
  64 struct workqueue_struct *gfs2_delete_workqueue;
  65 static LIST_HEAD(lru_list);
  66 static atomic_t lru_count = ATOMIC_INIT(0);
  67 static DEFINE_SPINLOCK(lru_lock);
  68
  69 #define GFS2_GL_HASH_SHIFT      15
  70 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
  71
  72 static const struct rhashtable_params ht_parms = {
  73         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
  74         .key_len = offsetofend(struct lm_lockname, ln_type),
  75         .key_offset = offsetof(struct gfs2_glock, gl_name),
  76         .head_offset = offsetof(struct gfs2_glock, gl_node),
  77 };
  78
  79 static struct rhashtable gl_hash_table;
  80
  81 #define GLOCK_WAIT_TABLE_BITS 12
  82 #define GLOCK_WAIT_TABLE_SIZE (1 << GLOCK_WAIT_TABLE_BITS)
  83 static wait_queue_head_t glock_wait_table[GLOCK_WAIT_TABLE_SIZE] __cacheline_aligned;
  84
  85 struct wait_glock_queue {
  86         struct lm_lockname *name;
  87         wait_queue_entry_t wait;
  88 };
  89
  90 static int glock_wake_function(wait_queue_entry_t *wait, unsigned int mode,
  91                                int sync, void *key)
  92 {
  93         struct wait_glock_queue *wait_glock =
  94                 container_of(wait, struct wait_glock_queue, wait);
  95         struct lm_lockname *wait_name = wait_glock->name;
  96         struct lm_lockname *wake_name = key;
  97
  98         if (wake_name->ln_sbd != wait_name->ln_sbd ||
  99             wake_name->ln_number != wait_name->ln_number ||
 100             wake_name->ln_type != wait_name->ln_type)
 101                 return 0;
 102         return autoremove_wake_function(wait, mode, sync, key);
 103 }
 104
 105 static wait_queue_head_t *glock_waitqueue(struct lm_lockname *name)
 106 {
 107         u32 hash = jhash2((u32 *)name, ht_parms.key_len / 4, 0);
 108
 109         return glock_wait_table + hash_32(hash, GLOCK_WAIT_TABLE_BITS);
 110 }
 111
 112 /**
 113  * wake_up_glock  -  Wake up waiters on a glock
 114  * @gl: the glock
 115  */
 116 static void wake_up_glock(struct gfs2_glock *gl)
 117 {
 118         wait_queue_head_t *wq = glock_waitqueue(&gl->gl_name);
 119
 120         if (waitqueue_active(wq))
 121                 __wake_up(wq, TASK_NORMAL, 1, &gl->gl_name);
 122 }
 123
 124 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 125 {
 126         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 127
 128         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
 129                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 130         } else {
 131                 kfree(gl->gl_lksb.sb_lvbptr);
 132                 kmem_cache_free(gfs2_glock_cachep, gl);
 133         }
 134 }
 135
 136 /**
 137  * glock_blocked_by_withdraw - determine if we can still use a glock
 138  * @gl: the glock
 139  *
 140  * We need to allow some glocks to be enqueued, dequeued, promoted, and demoted
 141  * when we're withdrawn. For example, to maintain metadata integrity, we should
 142  * disallow the use of inode and rgrp glocks when withdrawn. Other glocks, like
 143  * iopen or the transaction glocks may be safely used because none of their
 144  * metadata goes through the journal. So in general, we should disallow all
 145  * glocks that are journaled, and allow all the others. One exception is:
 146  * we need to allow our active journal to be promoted and demoted so others
 147  * may recover it and we can reacquire it when they're done.
 148  */
 149 static bool glock_blocked_by_withdraw(struct gfs2_glock *gl)
 150 {
 151         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 152
 153         if (likely(!gfs2_withdrawn(sdp)))
 154                 return false;
 155         if (gl->gl_ops->go_flags & GLOF_NONDISK)
 156                 return false;
 157         if (!sdp->sd_jdesc ||
 158             gl->gl_name.ln_number == sdp->sd_jdesc->jd_no_addr)
 159                 return false;
 160         return true;
 161 }
 162
 163 void gfs2_glock_free(struct gfs2_glock *gl)
 164 {
 165         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 166
 167         BUG_ON(atomic_read(&gl->gl_revokes));
 168         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 169         smp_mb();
 170         wake_up_glock(gl);
 171         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 172         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 173                 wake_up(&sdp->sd_glock_wait);
 174 }
 175
 176 /**
 177  * gfs2_glock_hold() - increment reference count on glock
 178  * @gl: The glock to hold
 179  *
 180  */
 181
 182 void gfs2_glock_hold(struct gfs2_glock *gl)
 183 {
 184         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 185         lockref_get(&gl->gl_lockref);
 186 }
 187
 188 /**
 189  * demote_ok - Check to see if it's ok to unlock a glock
 190  * @gl: the glock
 191  *
 192  * Returns: 1 if it's ok
 193  */
 194
 195 static int demote_ok(const struct gfs2_glock *gl)
 196 {
 197         const struct gfs2_glock_operations *glops = gl->gl_ops;
 198
 199         if (gl->gl_state == LM_ST_UNLOCKED)
 200                 return 0;
 201         if (!list_empty(&gl->gl_holders))
 202                 return 0;
 203         if (glops->go_demote_ok)
 204                 return glops->go_demote_ok(gl);
 205         return 1;
 206 }
 207
 208
 209 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 210 {
 211         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 212                 return;
 213
 214         spin_lock(&lru_lock);
 215
 216         list_del(&gl->gl_lru);
 217         list_add_tail(&gl->gl_lru, &lru_list);
 218
 219         if (!test_bit(GLF_LRU, &gl->gl_flags)) {
 220                 set_bit(GLF_LRU, &gl->gl_flags);
 221                 atomic_inc(&lru_count);
 222         }
 223
 224         spin_unlock(&lru_lock);
 225 }
 226
 227 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 228 {
 229         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 230                 return;
 231
 232         spin_lock(&lru_lock);
 233         if (test_bit(GLF_LRU, &gl->gl_flags)) {
 234                 list_del_init(&gl->gl_lru);
 235                 atomic_dec(&lru_count);
 236                 clear_bit(GLF_LRU, &gl->gl_flags);
 237         }
 238         spin_unlock(&lru_lock);
 239 }
 240
 241 /*
 242  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 243  * work queue.
 244  */
 245 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 246         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 247                 /*
 248                  * We are holding the lockref spinlock, and the work was still
 249                  * queued above.  The queued work (glock_work_func) takes that
 250                  * spinlock before dropping its glock reference(s), so it
 251                  * cannot have dropped them in the meantime.
 252                  */
 253                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 254                 gl->gl_lockref.count--;
 255         }
 256 }
 257
 258 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 259         spin_lock(&gl->gl_lockref.lock);
 260         __gfs2_glock_queue_work(gl, delay);
 261         spin_unlock(&gl->gl_lockref.lock);
 262 }
 263
 264 static void __gfs2_glock_put(struct gfs2_glock *gl)
 265 {
 266         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 267         struct address_space *mapping = gfs2_glock2aspace(gl);
 268
 269         lockref_mark_dead(&gl->gl_lockref);
 270
 271         gfs2_glock_remove_from_lru(gl);
 272         spin_unlock(&gl->gl_lockref.lock);
 273         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 274         GLOCK_BUG_ON(gl, mapping && mapping->nrpages && !gfs2_withdrawn(sdp));
 275         trace_gfs2_glock_put(gl);
 276         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 277 }
 278
 279 /*
 280  * Cause the glock to be put in work queue context.
 281  */
 282 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 283 {
 284         gfs2_glock_queue_work(gl, 0);
 285 }
 286
 287 /**
 288  * gfs2_glock_put() - Decrement reference count on glock
 289  * @gl: The glock to put
 290  *
 291  */
 292
 293 void gfs2_glock_put(struct gfs2_glock *gl)
 294 {
 295         if (lockref_put_or_lock(&gl->gl_lockref))
 296                 return;
 297
 298         __gfs2_glock_put(gl);
 299 }
 300
 301 /**
 302  * may_grant - check if its ok to grant a new lock
 303  * @gl: The glock
 304  * @gh: The lock request which we wish to grant
 305  *
 306  * Returns: true if its ok to grant the lock
 307  */
 308
 309 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 310 {
 311         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 312         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 313              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 314                 return 0;
 315         if (gl->gl_state == gh->gh_state)
 316                 return 1;
 317         if (gh->gh_flags & GL_EXACT)
 318                 return 0;
 319         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 320                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 321                         return 1;
 322                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 323                         return 1;
 324         }
 325         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 326                 return 1;
 327         return 0;
 328 }
 329
 330 static void gfs2_holder_wake(struct gfs2_holder *gh)
 331 {
 332         clear_bit(HIF_WAIT, &gh->gh_iflags);
 333         smp_mb__after_atomic();
 334         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 335         if (gh->gh_flags & GL_ASYNC) {
 336                 struct gfs2_sbd *sdp = gh->gh_gl->gl_name.ln_sbd;
 337
 338                 wake_up(&sdp->sd_async_glock_wait);
 339         }
 340 }
 341
 342 /**
 343  * do_error - Something unexpected has happened during a lock request
 344  *
 345  */
 346
 347 static void do_error(struct gfs2_glock *gl, const int ret)
 348 {
 349         struct gfs2_holder *gh, *tmp;
 350
 351         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 352                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 353                         continue;
 354                 if (ret & LM_OUT_ERROR)
 355                         gh->gh_error = -EIO;
 356                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 357                         gh->gh_error = GLR_TRYFAILED;
 358                 else
 359                         continue;
 360                 list_del_init(&gh->gh_list);
 361                 trace_gfs2_glock_queue(gh, 0);
 362                 gfs2_holder_wake(gh);
 363         }
 364 }
 365
 366 /**
 367  * do_promote - promote as many requests as possible on the current queue
 368  * @gl: The glock
 369  *
 370  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 371  *          if a type specific operation is underway.
 372  */
 373
 374 static int do_promote(struct gfs2_glock *gl)
 375 __releases(&gl->gl_lockref.lock)
 376 __acquires(&gl->gl_lockref.lock)
 377 {
 378         const struct gfs2_glock_operations *glops = gl->gl_ops;
 379         struct gfs2_holder *gh, *tmp;
 380         int ret;
 381
 382 restart:
 383         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 384                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 385                         continue;
 386                 if (may_grant(gl, gh)) {
 387                         if (gh->gh_list.prev == &gl->gl_holders &&
 388                             glops->go_lock) {
 389                                 spin_unlock(&gl->gl_lockref.lock);
 390                                 /* FIXME: eliminate this eventually */
 391                                 ret = glops->go_lock(gh);
 392                                 spin_lock(&gl->gl_lockref.lock);
 393                                 if (ret) {
 394                                         if (ret == 1)
 395                                                 return 2;
 396                                         gh->gh_error = ret;
 397                                         list_del_init(&gh->gh_list);
 398                                         trace_gfs2_glock_queue(gh, 0);
 399                                         gfs2_holder_wake(gh);
 400                                         goto restart;
 401                                 }
 402                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 403                                 trace_gfs2_promote(gh, 1);
 404                                 gfs2_holder_wake(gh);
 405                                 goto restart;
 406                         }
 407                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 408                         trace_gfs2_promote(gh, 0);
 409                         gfs2_holder_wake(gh);
 410                         continue;
 411                 }
 412                 if (gh->gh_list.prev == &gl->gl_holders)
 413                         return 1;
 414                 do_error(gl, 0);
 415                 break;
 416         }
 417         return 0;
 418 }
 419
 420 /**
 421  * find_first_waiter - find the first gh that's waiting for the glock
 422  * @gl: the glock
 423  */
 424
 425 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 426 {
 427         struct gfs2_holder *gh;
 428
 429         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 430                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 431                         return gh;
 432         }
 433         return NULL;
 434 }
 435
 436 /**
 437  * state_change - record that the glock is now in a different state
 438  * @gl: the glock
 439  * @new_state the new state
 440  *
 441  */
 442
 443 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 444 {
 445         int held1, held2;
 446
 447         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 448         held2 = (new_state != LM_ST_UNLOCKED);
 449
 450         if (held1 != held2) {
 451                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 452                 if (held2)
 453                         gl->gl_lockref.count++;
 454                 else
 455                         gl->gl_lockref.count--;
 456         }
 457         if (held1 && held2 && list_empty(&gl->gl_holders))
 458                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 459
 460         if (new_state != gl->gl_target)
 461                 /* shorten our minimum hold time */
 462                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 463                                        GL_GLOCK_MIN_HOLD);
 464         gl->gl_state = new_state;
 465         gl->gl_tchange = jiffies;
 466 }
 467
 468 static void gfs2_demote_wake(struct gfs2_glock *gl)
 469 {
 470         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 471         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 472         smp_mb__after_atomic();
 473         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 474 }
 475
 476 /**
 477  * finish_xmote - The DLM has replied to one of our lock requests
 478  * @gl: The glock
 479  * @ret: The status from the DLM
 480  *
 481  */
 482
 483 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 484 {
 485         const struct gfs2_glock_operations *glops = gl->gl_ops;
 486         struct gfs2_holder *gh;
 487         unsigned state = ret & LM_OUT_ST_MASK;
 488         int rv;
 489
 490         spin_lock(&gl->gl_lockref.lock);
 491         trace_gfs2_glock_state_change(gl, state);
 492         state_change(gl, state);
 493         gh = find_first_waiter(gl);
 494
 495         /* Demote to UN request arrived during demote to SH or DF */
 496         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 497             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 498                 gl->gl_target = LM_ST_UNLOCKED;
 499
 500         /* Check for state != intended state */
 501         if (unlikely(state != gl->gl_target)) {
 502                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 503                         /* move to back of queue and try next entry */
 504                         if (ret & LM_OUT_CANCELED) {
 505                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 506                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 507                                 gh = find_first_waiter(gl);
 508                                 gl->gl_target = gh->gh_state;
 509                                 goto retry;
 510                         }
 511                         /* Some error or failed "try lock" - report it */
 512                         if ((ret & LM_OUT_ERROR) ||
 513                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 514                                 gl->gl_target = gl->gl_state;
 515                                 do_error(gl, ret);
 516                                 goto out;
 517                         }
 518                 }
 519                 switch(state) {
 520                 /* Unlocked due to conversion deadlock, try again */
 521                 case LM_ST_UNLOCKED:
 522 retry:
 523                         do_xmote(gl, gh, gl->gl_target);
 524                         break;
 525                 /* Conversion fails, unlock and try again */
 526                 case LM_ST_SHARED:
 527                 case LM_ST_DEFERRED:
 528                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 529                         break;
 530                 default: /* Everything else */
 531                         fs_err(gl->gl_name.ln_sbd, "wanted %u got %u\n",
 532                                gl->gl_target, state);
 533                         GLOCK_BUG_ON(gl, 1);
 534                 }
 535                 spin_unlock(&gl->gl_lockref.lock);
 536                 return;
 537         }
 538
 539         /* Fast path - we got what we asked for */
 540         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 541                 gfs2_demote_wake(gl);
 542         if (state != LM_ST_UNLOCKED) {
 543                 if (glops->go_xmote_bh) {
 544                         spin_unlock(&gl->gl_lockref.lock);
 545                         rv = glops->go_xmote_bh(gl, gh);
 546                         spin_lock(&gl->gl_lockref.lock);
 547                         if (rv) {
 548                                 do_error(gl, rv);
 549                                 goto out;
 550                         }
 551                 }
 552                 rv = do_promote(gl);
 553                 if (rv == 2)
 554                         goto out_locked;
 555         }
 556 out:
 557         clear_bit(GLF_LOCK, &gl->gl_flags);
 558 out_locked:
 559         spin_unlock(&gl->gl_lockref.lock);
 560 }
 561
 562 /**
 563  * do_xmote - Calls the DLM to change the state of a lock
 564  * @gl: The lock state
 565  * @gh: The holder (only for promotes)
 566  * @target: The target lock state
 567  *
 568  */
 569
 570 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 571 __releases(&gl->gl_lockref.lock)
 572 __acquires(&gl->gl_lockref.lock)
 573 {
 574         const struct gfs2_glock_operations *glops = gl->gl_ops;
 575         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 576         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 577         int ret;
 578
 579         if (target != LM_ST_UNLOCKED && glock_blocked_by_withdraw(gl) &&
 580             gh && !(gh->gh_flags & LM_FLAG_NOEXP))
 581                 return;
 582         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 583                       LM_FLAG_PRIORITY);
 584         GLOCK_BUG_ON(gl, gl->gl_state == target);
 585         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 586         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 587             glops->go_inval) {
 588                 /*
 589                  * If another process is already doing the invalidate, let that
 590                  * finish first.  The glock state machine will get back to this
 591                  * holder again later.
 592                  */
 593                 if (test_and_set_bit(GLF_INVALIDATE_IN_PROGRESS,
 594                                      &gl->gl_flags))
 595                         return;
 596                 do_error(gl, 0); /* Fail queued try locks */
 597         }
 598         gl->gl_req = target;
 599         set_bit(GLF_BLOCKING, &gl->gl_flags);
 600         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 601             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 602             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 603                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 604         spin_unlock(&gl->gl_lockref.lock);
 605         if (glops->go_sync) {
 606                 ret = glops->go_sync(gl);
 607                 /* If we had a problem syncing (due to io errors or whatever,
 608                  * we should not invalidate the metadata or tell dlm to
 609                  * release the glock to other nodes.
 610                  */
 611                 if (ret) {
 612                         if (cmpxchg(&sdp->sd_log_error, 0, ret)) {
 613                                 fs_err(sdp, "Error %d syncing glock \n", ret);
 614                                 gfs2_dump_glock(NULL, gl, true);
 615                         }
 616                         return;
 617                 }
 618         }
 619         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags)) {
 620                 /*
 621                  * The call to go_sync should have cleared out the ail list.
 622                  * If there are still items, we have a problem. We ought to
 623                  * withdraw, but we can't because the withdraw code also uses
 624                  * glocks. Warn about the error, dump the glock, then fall
 625                  * through and wait for logd to do the withdraw for us.
 626                  */
 627                 if ((atomic_read(&gl->gl_ail_count) != 0) &&
 628                     (!cmpxchg(&sdp->sd_log_error, 0, -EIO))) {
 629                         gfs2_assert_warn(sdp, !atomic_read(&gl->gl_ail_count));
 630                         gfs2_dump_glock(NULL, gl, true);
 631                 }
 632                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 633                 clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 634         }
 635
 636         gfs2_glock_hold(gl);
 637         /*
 638          * Check for an error encountered since we called go_sync and go_inval.
 639          * If so, we can't withdraw from the glock code because the withdraw
 640          * code itself uses glocks (see function signal_our_withdraw) to
 641          * change the mount to read-only. Most importantly, we must not call
 642          * dlm to unlock the glock until the journal is in a known good state
 643          * (after journal replay) otherwise other nodes may use the object
 644          * (rgrp or dinode) and then later, journal replay will corrupt the
 645          * file system. The best we can do here is wait for the logd daemon
 646          * to see sd_log_error and withdraw, and in the meantime, requeue the
 647          * work for later.
 648          *
 649          * However, if we're just unlocking the lock (say, for unmount, when
 650          * gfs2_gl_hash_clear calls clear_glock) and recovery is complete
 651          * then it's okay to tell dlm to unlock it.
 652          */
 653         if (unlikely(sdp->sd_log_error && !gfs2_withdrawn(sdp)))
 654                 gfs2_withdraw_delayed(sdp);
 655         if (glock_blocked_by_withdraw(gl)) {
 656                 if (target != LM_ST_UNLOCKED ||
 657                     test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags)) {
 658                         gfs2_glock_queue_work(gl, GL_GLOCK_DFT_HOLD);
 659                         goto out;
 660                 }
 661         }
 662
 663         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 664                 /* lock_dlm */
 665                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 666                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 667                     target == LM_ST_UNLOCKED &&
 668                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 669                         finish_xmote(gl, target);
 670                         gfs2_glock_queue_work(gl, 0);
 671                 } else if (ret) {
 672                         fs_err(sdp, "lm_lock ret %d\n", ret);
 673                         GLOCK_BUG_ON(gl, !gfs2_withdrawn(sdp));
 674                 }
 675         } else { /* lock_nolock */
 676                 finish_xmote(gl, target);
 677                 gfs2_glock_queue_work(gl, 0);
 678         }
 679 out:
 680         spin_lock(&gl->gl_lockref.lock);
 681 }
 682
 683 /**
 684  * find_first_holder - find the first "holder" gh
 685  * @gl: the glock
 686  */
 687
 688 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 689 {
 690         struct gfs2_holder *gh;
 691
 692         if (!list_empty(&gl->gl_holders)) {
 693                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 694                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 695                         return gh;
 696         }
 697         return NULL;
 698 }
 699
 700 /**
 701  * run_queue - do all outstanding tasks related to a glock
 702  * @gl: The glock in question
 703  * @nonblock: True if we must not block in run_queue
 704  *
 705  */
 706
 707 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 708 __releases(&gl->gl_lockref.lock)
 709 __acquires(&gl->gl_lockref.lock)
 710 {
 711         struct gfs2_holder *gh = NULL;
 712         int ret;
 713
 714         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 715                 return;
 716
 717         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 718
 719         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 720             gl->gl_demote_state != gl->gl_state) {
 721                 if (find_first_holder(gl))
 722                         goto out_unlock;
 723                 if (nonblock)
 724                         goto out_sched;
 725                 smp_mb();
 726                 if (atomic_read(&gl->gl_revokes) != 0)
 727                         goto out_sched;
 728                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 729                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 730                 gl->gl_target = gl->gl_demote_state;
 731         } else {
 732                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 733                         gfs2_demote_wake(gl);
 734                 ret = do_promote(gl);
 735                 if (ret == 0)
 736                         goto out_unlock;
 737                 if (ret == 2)
 738                         goto out;
 739                 gh = find_first_waiter(gl);
 740                 gl->gl_target = gh->gh_state;
 741                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 742                         do_error(gl, 0); /* Fail queued try locks */
 743         }
 744         do_xmote(gl, gh, gl->gl_target);
 745 out:
 746         return;
 747
 748 out_sched:
 749         clear_bit(GLF_LOCK, &gl->gl_flags);
 750         smp_mb__after_atomic();
 751         gl->gl_lockref.count++;
 752         __gfs2_glock_queue_work(gl, 0);
 753         return;
 754
 755 out_unlock:
 756         clear_bit(GLF_LOCK, &gl->gl_flags);
 757         smp_mb__after_atomic();
 758         return;
 759 }
 760
 761 static void delete_work_func(struct work_struct *work)
 762 {
 763         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 764         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 765         struct inode *inode;
 766         u64 no_addr = gl->gl_name.ln_number;
 767
 768         /* If someone's using this glock to create a new dinode, the block must
 769            have been freed by another node, then re-used, in which case our
 770            iopen callback is too late after the fact. Ignore it. */
 771         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 772                 goto out;
 773
 774         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 775         if (!IS_ERR_OR_NULL(inode)) {
 776                 d_prune_aliases(inode);
 777                 iput(inode);
 778         }
 779 out:
 780         gfs2_glock_put(gl);
 781 }
 782
 783 static void glock_work_func(struct work_struct *work)
 784 {
 785         unsigned long delay = 0;
 786         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 787         unsigned int drop_refs = 1;
 788
 789         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 790                 finish_xmote(gl, gl->gl_reply);
 791                 drop_refs++;
 792         }
 793         spin_lock(&gl->gl_lockref.lock);
 794         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 795             gl->gl_state != LM_ST_UNLOCKED &&
 796             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 797                 unsigned long holdtime, now = jiffies;
 798
 799                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 800                 if (time_before(now, holdtime))
 801                         delay = holdtime - now;
 802
 803                 if (!delay) {
 804                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 805                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 806                 }
 807         }
 808         run_queue(gl, 0);
 809         if (delay) {
 810                 /* Keep one glock reference for the work we requeue. */
 811                 drop_refs--;
 812                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 813                         delay = 0;
 814                 __gfs2_glock_queue_work(gl, delay);
 815         }
 816
 817         /*
 818          * Drop the remaining glock references manually here. (Mind that
 819          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 820          * here as well.)
 821          */
 822         gl->gl_lockref.count -= drop_refs;
 823         if (!gl->gl_lockref.count) {
 824                 __gfs2_glock_put(gl);
 825                 return;
 826         }
 827         spin_unlock(&gl->gl_lockref.lock);
 828 }
 829
 830 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 831                                             struct gfs2_glock *new)
 832 {
 833         struct wait_glock_queue wait;
 834         wait_queue_head_t *wq = glock_waitqueue(name);
 835         struct gfs2_glock *gl;
 836
 837         wait.name = name;
 838         init_wait(&wait.wait);
 839         wait.wait.func = glock_wake_function;
 840
 841 again:
 842         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 843         rcu_read_lock();
 844         if (new) {
 845                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 846                         &new->gl_node, ht_parms);
 847                 if (IS_ERR(gl))
 848                         goto out;
 849         } else {
 850                 gl = rhashtable_lookup_fast(&gl_hash_table,
 851                         name, ht_parms);
 852         }
 853         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 854                 rcu_read_unlock();
 855                 schedule();
 856                 goto again;
 857         }
 858 out:
 859         rcu_read_unlock();
 860         finish_wait(wq, &wait.wait);
 861         return gl;
 862 }
 863
 864 /**
 865  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 866  * @sdp: The GFS2 superblock
 867  * @number: the lock number
 868  * @glops: The glock_operations to use
 869  * @create: If 0, don't create the glock if it doesn't exist
 870  * @glp: the glock is returned here
 871  *
 872  * This does not lock a glock, just finds/creates structures for one.
 873  *
 874  * Returns: errno
 875  */
 876
 877 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 878                    const struct gfs2_glock_operations *glops, int create,
 879                    struct gfs2_glock **glp)
 880 {
 881         struct super_block *s = sdp->sd_vfs;
 882         struct lm_lockname name = { .ln_number = number,
 883                                     .ln_type = glops->go_type,
 884                                     .ln_sbd = sdp };
 885         struct gfs2_glock *gl, *tmp;
 886         struct address_space *mapping;
 887         struct kmem_cache *cachep;
 888         int ret = 0;
 889
 890         gl = find_insert_glock(&name, NULL);
 891         if (gl) {
 892                 *glp = gl;
 893                 return 0;
 894         }
 895         if (!create)
 896                 return -ENOENT;
 897
 898         if (glops->go_flags & GLOF_ASPACE)
 899                 cachep = gfs2_glock_aspace_cachep;
 900         else
 901                 cachep = gfs2_glock_cachep;
 902         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 903         if (!gl)
 904                 return -ENOMEM;
 905
 906         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 907
 908         if (glops->go_flags & GLOF_LVB) {
 909                 gl->gl_lksb.sb_lvbptr = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
 910                 if (!gl->gl_lksb.sb_lvbptr) {
 911                         kmem_cache_free(cachep, gl);
 912                         return -ENOMEM;
 913                 }
 914         }
 915
 916         atomic_inc(&sdp->sd_glock_disposal);
 917         gl->gl_node.next = NULL;
 918         gl->gl_flags = 0;
 919         gl->gl_name = name;
 920         gl->gl_lockref.count = 1;
 921         gl->gl_state = LM_ST_UNLOCKED;
 922         gl->gl_target = LM_ST_UNLOCKED;
 923         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 924         gl->gl_ops = glops;
 925         gl->gl_dstamp = 0;
 926         preempt_disable();
 927         /* We use the global stats to estimate the initial per-glock stats */
 928         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 929         preempt_enable();
 930         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 931         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 932         gl->gl_tchange = jiffies;
 933         gl->gl_object = NULL;
 934         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 935         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 936         INIT_WORK(&gl->gl_delete, delete_work_func);
 937
 938         mapping = gfs2_glock2aspace(gl);
 939         if (mapping) {
 940                 mapping->a_ops = &gfs2_meta_aops;
 941                 mapping->host = s->s_bdev->bd_inode;
 942                 mapping->flags = 0;
 943                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 944                 mapping->private_data = NULL;
 945                 mapping->writeback_index = 0;
 946         }
 947
 948         tmp = find_insert_glock(&name, gl);
 949         if (!tmp) {
 950                 *glp = gl;
 951                 goto out;
 952         }
 953         if (IS_ERR(tmp)) {
 954                 ret = PTR_ERR(tmp);
 955                 goto out_free;
 956         }
 957         *glp = tmp;
 958
 959 out_free:
 960         kfree(gl->gl_lksb.sb_lvbptr);
 961         kmem_cache_free(cachep, gl);
 962         atomic_dec(&sdp->sd_glock_disposal);
 963
 964 out:
 965         return ret;
 966 }
 967
 968 /**
 969  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 970  * @gl: the glock
 971  * @state: the state we're requesting
 972  * @flags: the modifier flags
 973  * @gh: the holder structure
 974  *
 975  */
 976
 977 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 978                       struct gfs2_holder *gh)
 979 {
 980         INIT_LIST_HEAD(&gh->gh_list);
 981         gh->gh_gl = gl;
 982         gh->gh_ip = _RET_IP_;
 983         gh->gh_owner_pid = get_pid(task_pid(current));
 984         gh->gh_state = state;
 985         gh->gh_flags = flags;
 986         gh->gh_error = 0;
 987         gh->gh_iflags = 0;
 988         gfs2_glock_hold(gl);
 989 }
 990
 991 /**
 992  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 993  * @state: the state we're requesting
 994  * @flags: the modifier flags
 995  * @gh: the holder structure
 996  *
 997  * Don't mess with the glock.
 998  *
 999  */
1000
1001 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
1002 {
1003         gh->gh_state = state;
1004         gh->gh_flags = flags;
1005         gh->gh_iflags = 0;
1006         gh->gh_ip = _RET_IP_;
1007         put_pid(gh->gh_owner_pid);
1008         gh->gh_owner_pid = get_pid(task_pid(current));
1009 }
1010
1011 /**
1012  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
1013  * @gh: the holder structure
1014  *
1015  */
1016
1017 void gfs2_holder_uninit(struct gfs2_holder *gh)
1018 {
1019         put_pid(gh->gh_owner_pid);
1020         gfs2_glock_put(gh->gh_gl);
1021         gfs2_holder_mark_uninitialized(gh);
1022         gh->gh_ip = 0;
1023 }
1024
1025 static void gfs2_glock_update_hold_time(struct gfs2_glock *gl,
1026                                         unsigned long start_time)
1027 {
1028         /* Have we waited longer that a second? */
1029         if (time_after(jiffies, start_time + HZ)) {
1030                 /* Lengthen the minimum hold time. */
1031                 gl->gl_hold_time = min(gl->gl_hold_time + GL_GLOCK_HOLD_INCR,
1032                                        GL_GLOCK_MAX_HOLD);
1033         }
1034 }
1035
1036 /**
1037  * gfs2_glock_wait - wait on a glock acquisition
1038  * @gh: the glock holder
1039  *
1040  * Returns: 0 on success
1041  */
1042
1043 int gfs2_glock_wait(struct gfs2_holder *gh)
1044 {
1045         unsigned long start_time = jiffies;
1046
1047         might_sleep();
1048         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
1049         gfs2_glock_update_hold_time(gh->gh_gl, start_time);
1050         return gh->gh_error;
1051 }
1052
1053 static int glocks_pending(unsigned int num_gh, struct gfs2_holder *ghs)
1054 {
1055         int i;
1056
1057         for (i = 0; i < num_gh; i++)
1058                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags))
1059                         return 1;
1060         return 0;
1061 }
1062
1063 /**
1064  * gfs2_glock_async_wait - wait on multiple asynchronous glock acquisitions
1065  * @num_gh: the number of holders in the array
1066  * @ghs: the glock holder array
1067  *
1068  * Returns: 0 on success, meaning all glocks have been granted and are held.
1069  *          -ESTALE if the request timed out, meaning all glocks were released,
1070  *          and the caller should retry the operation.
1071  */
1072
1073 int gfs2_glock_async_wait(unsigned int num_gh, struct gfs2_holder *ghs)
1074 {
1075         struct gfs2_sbd *sdp = ghs[0].gh_gl->gl_name.ln_sbd;
1076         int i, ret = 0, timeout = 0;
1077         unsigned long start_time = jiffies;
1078         bool keep_waiting;
1079
1080         might_sleep();
1081         /*
1082          * Total up the (minimum hold time * 2) of all glocks and use that to
1083          * determine the max amount of time we should wait.
1084          */
1085         for (i = 0; i < num_gh; i++)
1086                 timeout += ghs[i].gh_gl->gl_hold_time << 1;
1087
1088 wait_for_dlm:
1089         if (!wait_event_timeout(sdp->sd_async_glock_wait,
1090                                 !glocks_pending(num_gh, ghs), timeout))
1091                 ret = -ESTALE; /* request timed out. */
1092
1093         /*
1094          * If dlm granted all our requests, we need to adjust the glock
1095          * minimum hold time values according to how long we waited.
1096          *
1097          * If our request timed out, we need to repeatedly release any held
1098          * glocks we acquired thus far to allow dlm to acquire the remaining
1099          * glocks without deadlocking.  We cannot currently cancel outstanding
1100          * glock acquisitions.
1101          *
1102          * The HIF_WAIT bit tells us which requests still need a response from
1103          * dlm.
1104          *
1105          * If dlm sent us any errors, we return the first error we find.
1106          */
1107         keep_waiting = false;
1108         for (i = 0; i < num_gh; i++) {
1109                 /* Skip holders we have already dequeued below. */
1110                 if (!gfs2_holder_queued(&ghs[i]))
1111                         continue;
1112                 /* Skip holders with a pending DLM response. */
1113                 if (test_bit(HIF_WAIT, &ghs[i].gh_iflags)) {
1114                         keep_waiting = true;
1115                         continue;
1116                 }
1117
1118                 if (test_bit(HIF_HOLDER, &ghs[i].gh_iflags)) {
1119                         if (ret == -ESTALE)
1120                                 gfs2_glock_dq(&ghs[i]);
1121                         else
1122                                 gfs2_glock_update_hold_time(ghs[i].gh_gl,
1123                                                             start_time);
1124                 }
1125                 if (!ret)
1126                         ret = ghs[i].gh_error;
1127         }
1128
1129         if (keep_waiting)
1130                 goto wait_for_dlm;
1131
1132         /*
1133          * At this point, we've either acquired all locks or released them all.
1134          */
1135         return ret;
1136 }
1137
1138 /**
1139  * handle_callback - process a demote request
1140  * @gl: the glock
1141  * @state: the state the caller wants us to change to
1142  *
1143  * There are only two requests that we are going to see in actual
1144  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
1145  */
1146
1147 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
1148                             unsigned long delay, bool remote)
1149 {
1150         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
1151
1152         set_bit(bit, &gl->gl_flags);
1153         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
1154                 gl->gl_demote_state = state;
1155                 gl->gl_demote_time = jiffies;
1156         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
1157                         gl->gl_demote_state != state) {
1158                 gl->gl_demote_state = LM_ST_UNLOCKED;
1159         }
1160         if (gl->gl_ops->go_callback)
1161                 gl->gl_ops->go_callback(gl, remote);
1162         trace_gfs2_demote_rq(gl, remote);
1163 }
1164
1165 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
1166 {
1167         struct va_format vaf;
1168         va_list args;
1169
1170         va_start(args, fmt);
1171
1172         if (seq) {
1173                 seq_vprintf(seq, fmt, args);
1174         } else {
1175                 vaf.fmt = fmt;
1176                 vaf.va = &args;
1177
1178                 pr_err("%pV", &vaf);
1179         }
1180
1181         va_end(args);
1182 }
1183
1184 /**
1185  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1186  * @gh: the holder structure to add
1187  *
1188  * Eventually we should move the recursive locking trap to a
1189  * debugging option or something like that. This is the fast
1190  * path and needs to have the minimum number of distractions.
1191  *
1192  */
1193
1194 static inline void add_to_queue(struct gfs2_holder *gh)
1195 __releases(&gl->gl_lockref.lock)
1196 __acquires(&gl->gl_lockref.lock)
1197 {
1198         struct gfs2_glock *gl = gh->gh_gl;
1199         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1200         struct list_head *insert_pt = NULL;
1201         struct gfs2_holder *gh2;
1202         int try_futile = 0;
1203
1204         GLOCK_BUG_ON(gl, gh->gh_owner_pid == NULL);
1205         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1206                 GLOCK_BUG_ON(gl, true);
1207
1208         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1209                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1210                         try_futile = !may_grant(gl, gh);
1211                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1212                         goto fail;
1213         }
1214
1215         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1216                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1217                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1218                         goto trap_recursive;
1219                 if (try_futile &&
1220                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1221 fail:
1222                         gh->gh_error = GLR_TRYFAILED;
1223                         gfs2_holder_wake(gh);
1224                         return;
1225                 }
1226                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1227                         continue;
1228                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1229                         insert_pt = &gh2->gh_list;
1230         }
1231         set_bit(GLF_QUEUED, &gl->gl_flags);
1232         trace_gfs2_glock_queue(gh, 1);
1233         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1234         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1235         if (likely(insert_pt == NULL)) {
1236                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1237                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1238                         goto do_cancel;
1239                 return;
1240         }
1241         list_add_tail(&gh->gh_list, insert_pt);
1242 do_cancel:
1243         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1244         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1245                 spin_unlock(&gl->gl_lockref.lock);
1246                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1247                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1248                 spin_lock(&gl->gl_lockref.lock);
1249         }
1250         return;
1251
1252 trap_recursive:
1253         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1254         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1255         fs_err(sdp, "lock type: %d req lock state : %d\n",
1256                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1257         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1258         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1259         fs_err(sdp, "lock type: %d req lock state : %d\n",
1260                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1261         gfs2_dump_glock(NULL, gl, true);
1262         BUG();
1263 }
1264
1265 /**
1266  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1267  * @gh: the holder structure
1268  *
1269  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1270  *
1271  * Returns: 0, GLR_TRYFAILED, or errno on failure
1272  */
1273
1274 int gfs2_glock_nq(struct gfs2_holder *gh)
1275 {
1276         struct gfs2_glock *gl = gh->gh_gl;
1277         int error = 0;
1278
1279         if (glock_blocked_by_withdraw(gl) && !(gh->gh_flags & LM_FLAG_NOEXP))
1280                 return -EIO;
1281
1282         if (test_bit(GLF_LRU, &gl->gl_flags))
1283                 gfs2_glock_remove_from_lru(gl);
1284
1285         spin_lock(&gl->gl_lockref.lock);
1286         add_to_queue(gh);
1287         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1288                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1289                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1290                 gl->gl_lockref.count++;
1291                 __gfs2_glock_queue_work(gl, 0);
1292         }
1293         run_queue(gl, 1);
1294         spin_unlock(&gl->gl_lockref.lock);
1295
1296         if (!(gh->gh_flags & GL_ASYNC))
1297                 error = gfs2_glock_wait(gh);
1298
1299         return error;
1300 }
1301
1302 /**
1303  * gfs2_glock_poll - poll to see if an async request has been completed
1304  * @gh: the holder
1305  *
1306  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1307  */
1308
1309 int gfs2_glock_poll(struct gfs2_holder *gh)
1310 {
1311         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1312 }
1313
1314 /**
1315  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1316  * @gh: the glock holder
1317  *
1318  */
1319
1320 void gfs2_glock_dq(struct gfs2_holder *gh)
1321 {
1322         struct gfs2_glock *gl = gh->gh_gl;
1323         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1324         unsigned delay = 0;
1325         int fast_path = 0;
1326
1327         spin_lock(&gl->gl_lockref.lock);
1328         /*
1329          * If we're in the process of file system withdraw, we cannot just
1330          * dequeue any glocks until our journal is recovered, lest we
1331          * introduce file system corruption. We need two exceptions to this
1332          * rule: We need to allow unlocking of nondisk glocks and the glock
1333          * for our own journal that needs recovery.
1334          */
1335         if (test_bit(SDF_WITHDRAW_RECOVERY, &sdp->sd_flags) &&
1336             glock_blocked_by_withdraw(gl) &&
1337             gh->gh_gl != sdp->sd_jinode_gl) {
1338                 sdp->sd_glock_dqs_held++;
1339                 might_sleep();
1340                 wait_on_bit(&sdp->sd_flags, SDF_WITHDRAW_RECOVERY,
1341                             TASK_UNINTERRUPTIBLE);
1342         }
1343         if (gh->gh_flags & GL_NOCACHE)
1344                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1345
1346         list_del_init(&gh->gh_list);
1347         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1348         if (find_first_holder(gl) == NULL) {
1349                 if (list_empty(&gl->gl_holders) &&
1350                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1351                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1352                         fast_path = 1;
1353         }
1354         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1355                 gfs2_glock_add_to_lru(gl);
1356
1357         trace_gfs2_glock_queue(gh, 0);
1358         if (unlikely(!fast_path)) {
1359                 gl->gl_lockref.count++;
1360                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1361                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1362                     gl->gl_name.ln_type == LM_TYPE_INODE)
1363                         delay = gl->gl_hold_time;
1364                 __gfs2_glock_queue_work(gl, delay);
1365         }
1366         spin_unlock(&gl->gl_lockref.lock);
1367 }
1368
1369 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1370 {
1371         struct gfs2_glock *gl = gh->gh_gl;
1372         gfs2_glock_dq(gh);
1373         might_sleep();
1374         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1375 }
1376
1377 /**
1378  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1379  * @gh: the holder structure
1380  *
1381  */
1382
1383 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1384 {
1385         gfs2_glock_dq(gh);
1386         gfs2_holder_uninit(gh);
1387 }
1388
1389 /**
1390  * gfs2_glock_nq_num - acquire a glock based on lock number
1391  * @sdp: the filesystem
1392  * @number: the lock number
1393  * @glops: the glock operations for the type of glock
1394  * @state: the state to acquire the glock in
1395  * @flags: modifier flags for the acquisition
1396  * @gh: the struct gfs2_holder
1397  *
1398  * Returns: errno
1399  */
1400
1401 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1402                       const struct gfs2_glock_operations *glops,
1403                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1404 {
1405         struct gfs2_glock *gl;
1406         int error;
1407
1408         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1409         if (!error) {
1410                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1411                 gfs2_glock_put(gl);
1412         }
1413
1414         return error;
1415 }
1416
1417 /**
1418  * glock_compare - Compare two struct gfs2_glock structures for sorting
1419  * @arg_a: the first structure
1420  * @arg_b: the second structure
1421  *
1422  */
1423
1424 static int glock_compare(const void *arg_a, const void *arg_b)
1425 {
1426         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1427         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1428         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1429         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1430
1431         if (a->ln_number > b->ln_number)
1432                 return 1;
1433         if (a->ln_number < b->ln_number)
1434                 return -1;
1435         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1436         return 0;
1437 }
1438
1439 /**
1440  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1441  * @num_gh: the number of structures
1442  * @ghs: an array of struct gfs2_holder structures
1443  *
1444  * Returns: 0 on success (all glocks acquired),
1445  *          errno on failure (no glocks acquired)
1446  */
1447
1448 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1449                      struct gfs2_holder **p)
1450 {
1451         unsigned int x;
1452         int error = 0;
1453
1454         for (x = 0; x < num_gh; x++)
1455                 p[x] = &ghs[x];
1456
1457         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1458
1459         for (x = 0; x < num_gh; x++) {
1460                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1461
1462                 error = gfs2_glock_nq(p[x]);
1463                 if (error) {
1464                         while (x--)
1465                                 gfs2_glock_dq(p[x]);
1466                         break;
1467                 }
1468         }
1469
1470         return error;
1471 }
1472
1473 /**
1474  * gfs2_glock_nq_m - acquire multiple glocks
1475  * @num_gh: the number of structures
1476  * @ghs: an array of struct gfs2_holder structures
1477  *
1478  *
1479  * Returns: 0 on success (all glocks acquired),
1480  *          errno on failure (no glocks acquired)
1481  */
1482
1483 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1484 {
1485         struct gfs2_holder *tmp[4];
1486         struct gfs2_holder **pph = tmp;
1487         int error = 0;
1488
1489         switch(num_gh) {
1490         case 0:
1491                 return 0;
1492         case 1:
1493                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1494                 return gfs2_glock_nq(ghs);
1495         default:
1496                 if (num_gh <= 4)
1497                         break;
1498                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1499                                     GFP_NOFS);
1500                 if (!pph)
1501                         return -ENOMEM;
1502         }
1503
1504         error = nq_m_sync(num_gh, ghs, pph);
1505
1506         if (pph != tmp)
1507                 kfree(pph);
1508
1509         return error;
1510 }
1511
1512 /**
1513  * gfs2_glock_dq_m - release multiple glocks
1514  * @num_gh: the number of structures
1515  * @ghs: an array of struct gfs2_holder structures
1516  *
1517  */
1518
1519 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1520 {
1521         while (num_gh--)
1522                 gfs2_glock_dq(&ghs[num_gh]);
1523 }
1524
1525 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1526 {
1527         unsigned long delay = 0;
1528         unsigned long holdtime;
1529         unsigned long now = jiffies;
1530
1531         gfs2_glock_hold(gl);
1532         holdtime = gl->gl_tchange + gl->gl_hold_time;
1533         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1534             gl->gl_name.ln_type == LM_TYPE_INODE) {
1535                 if (time_before(now, holdtime))
1536                         delay = holdtime - now;
1537                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1538                         delay = gl->gl_hold_time;
1539         }
1540
1541         spin_lock(&gl->gl_lockref.lock);
1542         handle_callback(gl, state, delay, true);
1543         __gfs2_glock_queue_work(gl, delay);
1544         spin_unlock(&gl->gl_lockref.lock);
1545 }
1546
1547 /**
1548  * gfs2_should_freeze - Figure out if glock should be frozen
1549  * @gl: The glock in question
1550  *
1551  * Glocks are not frozen if (a) the result of the dlm operation is
1552  * an error, (b) the locking operation was an unlock operation or
1553  * (c) if there is a "noexp" flagged request anywhere in the queue
1554  *
1555  * Returns: 1 if freezing should occur, 0 otherwise
1556  */
1557
1558 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1559 {
1560         const struct gfs2_holder *gh;
1561
1562         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1563                 return 0;
1564         if (gl->gl_target == LM_ST_UNLOCKED)
1565                 return 0;
1566
1567         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1568                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1569                         continue;
1570                 if (LM_FLAG_NOEXP & gh->gh_flags)
1571                         return 0;
1572         }
1573
1574         return 1;
1575 }
1576
1577 /**
1578  * gfs2_glock_complete - Callback used by locking
1579  * @gl: Pointer to the glock
1580  * @ret: The return value from the dlm
1581  *
1582  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1583  * to use a bitfield shared with other glock state fields.
1584  */
1585
1586 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1587 {
1588         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1589
1590         spin_lock(&gl->gl_lockref.lock);
1591         gl->gl_reply = ret;
1592
1593         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1594                 if (gfs2_should_freeze(gl)) {
1595                         set_bit(GLF_FROZEN, &gl->gl_flags);
1596                         spin_unlock(&gl->gl_lockref.lock);
1597                         return;
1598                 }
1599         }
1600
1601         gl->gl_lockref.count++;
1602         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1603         __gfs2_glock_queue_work(gl, 0);
1604         spin_unlock(&gl->gl_lockref.lock);
1605 }
1606
1607 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1608 {
1609         struct gfs2_glock *gla, *glb;
1610
1611         gla = list_entry(a, struct gfs2_glock, gl_lru);
1612         glb = list_entry(b, struct gfs2_glock, gl_lru);
1613
1614         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1615                 return 1;
1616         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1617                 return -1;
1618
1619         return 0;
1620 }
1621
1622 /**
1623  * gfs2_dispose_glock_lru - Demote a list of glocks
1624  * @list: The list to dispose of
1625  *
1626  * Disposing of glocks may involve disk accesses, so that here we sort
1627  * the glocks by number (i.e. disk location of the inodes) so that if
1628  * there are any such accesses, they'll be sent in order (mostly).
1629  *
1630  * Must be called under the lru_lock, but may drop and retake this
1631  * lock. While the lru_lock is dropped, entries may vanish from the
1632  * list, but no new entries will appear on the list (since it is
1633  * private)
1634  */
1635
1636 static void gfs2_dispose_glock_lru(struct list_head *list)
1637 __releases(&lru_lock)
1638 __acquires(&lru_lock)
1639 {
1640         struct gfs2_glock *gl;
1641
1642         list_sort(NULL, list, glock_cmp);
1643
1644         while(!list_empty(list)) {
1645                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1646                 list_del_init(&gl->gl_lru);
1647                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1648 add_back_to_lru:
1649                         list_add(&gl->gl_lru, &lru_list);
1650                         set_bit(GLF_LRU, &gl->gl_flags);
1651                         atomic_inc(&lru_count);
1652                         continue;
1653                 }
1654                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1655                         spin_unlock(&gl->gl_lockref.lock);
1656                         goto add_back_to_lru;
1657                 }
1658                 gl->gl_lockref.count++;
1659                 if (demote_ok(gl))
1660                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1661                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1662                 __gfs2_glock_queue_work(gl, 0);
1663                 spin_unlock(&gl->gl_lockref.lock);
1664                 cond_resched_lock(&lru_lock);
1665         }
1666 }
1667
1668 /**
1669  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1670  * @nr: The number of entries to scan
1671  *
1672  * This function selects the entries on the LRU which are able to
1673  * be demoted, and then kicks off the process by calling
1674  * gfs2_dispose_glock_lru() above.
1675  */
1676
1677 static long gfs2_scan_glock_lru(int nr)
1678 {
1679         struct gfs2_glock *gl;
1680         LIST_HEAD(skipped);
1681         LIST_HEAD(dispose);
1682         long freed = 0;
1683
1684         spin_lock(&lru_lock);
1685         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1686                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1687
1688                 /* Test for being demotable */
1689                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1690                         list_move(&gl->gl_lru, &dispose);
1691                         atomic_dec(&lru_count);
1692                         clear_bit(GLF_LRU, &gl->gl_flags);
1693                         freed++;
1694                         continue;
1695                 }
1696
1697                 list_move(&gl->gl_lru, &skipped);
1698         }
1699         list_splice(&skipped, &lru_list);
1700         if (!list_empty(&dispose))
1701                 gfs2_dispose_glock_lru(&dispose);
1702         spin_unlock(&lru_lock);
1703
1704         return freed;
1705 }
1706
1707 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1708                                             struct shrink_control *sc)
1709 {
1710         if (!(sc->gfp_mask & __GFP_FS))
1711                 return SHRINK_STOP;
1712         return gfs2_scan_glock_lru(sc->nr_to_scan);
1713 }
1714
1715 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1716                                              struct shrink_control *sc)
1717 {
1718         return vfs_pressure_ratio(atomic_read(&lru_count));
1719 }
1720
1721 static struct shrinker glock_shrinker = {
1722         .seeks = DEFAULT_SEEKS,
1723         .count_objects = gfs2_glock_shrink_count,
1724         .scan_objects = gfs2_glock_shrink_scan,
1725 };
1726
1727 /**
1728  * examine_bucket - Call a function for glock in a hash bucket
1729  * @examiner: the function
1730  * @sdp: the filesystem
1731  * @bucket: the bucket
1732  *
1733  * Note that the function can be called multiple times on the same
1734  * object.  So the user must ensure that the function can cope with
1735  * that.
1736  */
1737
1738 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1739 {
1740         struct gfs2_glock *gl;
1741         struct rhashtable_iter iter;
1742
1743         rhashtable_walk_enter(&gl_hash_table, &iter);
1744
1745         do {
1746                 rhashtable_walk_start(&iter);
1747
1748                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1749                         if (gl->gl_name.ln_sbd == sdp &&
1750                             lockref_get_not_dead(&gl->gl_lockref))
1751                                 examiner(gl);
1752
1753                 rhashtable_walk_stop(&iter);
1754         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1755
1756         rhashtable_walk_exit(&iter);
1757 }
1758
1759 /**
1760  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1761  * @gl: The glock to thaw
1762  *
1763  */
1764
1765 static void thaw_glock(struct gfs2_glock *gl)
1766 {
1767         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1768                 gfs2_glock_put(gl);
1769                 return;
1770         }
1771         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1772         gfs2_glock_queue_work(gl, 0);
1773 }
1774
1775 /**
1776  * clear_glock - look at a glock and see if we can free it from glock cache
1777  * @gl: the glock to look at
1778  *
1779  */
1780
1781 static void clear_glock(struct gfs2_glock *gl)
1782 {
1783         gfs2_glock_remove_from_lru(gl);
1784
1785         spin_lock(&gl->gl_lockref.lock);
1786         if (gl->gl_state != LM_ST_UNLOCKED)
1787                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1788         __gfs2_glock_queue_work(gl, 0);
1789         spin_unlock(&gl->gl_lockref.lock);
1790 }
1791
1792 /**
1793  * gfs2_glock_thaw - Thaw any frozen glocks
1794  * @sdp: The super block
1795  *
1796  */
1797
1798 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1799 {
1800         glock_hash_walk(thaw_glock, sdp);
1801 }
1802
1803 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1804 {
1805         spin_lock(&gl->gl_lockref.lock);
1806         gfs2_dump_glock(seq, gl, fsid);
1807         spin_unlock(&gl->gl_lockref.lock);
1808 }
1809
1810 static void dump_glock_func(struct gfs2_glock *gl)
1811 {
1812         dump_glock(NULL, gl, true);
1813 }
1814
1815 /**
1816  * gfs2_gl_hash_clear - Empty out the glock hash table
1817  * @sdp: the filesystem
1818  * @wait: wait until it's all gone
1819  *
1820  * Called when unmounting the filesystem.
1821  */
1822
1823 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1824 {
1825         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1826         flush_workqueue(glock_workqueue);
1827         glock_hash_walk(clear_glock, sdp);
1828         flush_workqueue(glock_workqueue);
1829         wait_event_timeout(sdp->sd_glock_wait,
1830                            atomic_read(&sdp->sd_glock_disposal) == 0,
1831                            HZ * 600);
1832         glock_hash_walk(dump_glock_func, sdp);
1833 }
1834
1835 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1836 {
1837         struct gfs2_glock *gl = ip->i_gl;
1838         int ret;
1839
1840         ret = gfs2_truncatei_resume(ip);
1841         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1842
1843         spin_lock(&gl->gl_lockref.lock);
1844         clear_bit(GLF_LOCK, &gl->gl_flags);
1845         run_queue(gl, 1);
1846         spin_unlock(&gl->gl_lockref.lock);
1847 }
1848
1849 static const char *state2str(unsigned state)
1850 {
1851         switch(state) {
1852         case LM_ST_UNLOCKED:
1853                 return "UN";
1854         case LM_ST_SHARED:
1855                 return "SH";
1856         case LM_ST_DEFERRED:
1857                 return "DF";
1858         case LM_ST_EXCLUSIVE:
1859                 return "EX";
1860         }
1861         return "??";
1862 }
1863
1864 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1865 {
1866         char *p = buf;
1867         if (flags & LM_FLAG_TRY)
1868                 *p++ = 't';
1869         if (flags & LM_FLAG_TRY_1CB)
1870                 *p++ = 'T';
1871         if (flags & LM_FLAG_NOEXP)
1872                 *p++ = 'e';
1873         if (flags & LM_FLAG_ANY)
1874                 *p++ = 'A';
1875         if (flags & LM_FLAG_PRIORITY)
1876                 *p++ = 'p';
1877         if (flags & GL_ASYNC)
1878                 *p++ = 'a';
1879         if (flags & GL_EXACT)
1880                 *p++ = 'E';
1881         if (flags & GL_NOCACHE)
1882                 *p++ = 'c';
1883         if (test_bit(HIF_HOLDER, &iflags))
1884                 *p++ = 'H';
1885         if (test_bit(HIF_WAIT, &iflags))
1886                 *p++ = 'W';
1887         if (test_bit(HIF_FIRST, &iflags))
1888                 *p++ = 'F';
1889         *p = 0;
1890         return buf;
1891 }
1892
1893 /**
1894  * dump_holder - print information about a glock holder
1895  * @seq: the seq_file struct
1896  * @gh: the glock holder
1897  * @fs_id_buf: pointer to file system id (if requested)
1898  *
1899  */
1900
1901 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh,
1902                         const char *fs_id_buf)
1903 {
1904         struct task_struct *gh_owner = NULL;
1905         char flags_buf[32];
1906
1907         rcu_read_lock();
1908         if (gh->gh_owner_pid)
1909                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1910         gfs2_print_dbg(seq, "%s H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1911                        fs_id_buf, state2str(gh->gh_state),
1912                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1913                        gh->gh_error,
1914                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1915                        gh_owner ? gh_owner->comm : "(ended)",
1916                        (void *)gh->gh_ip);
1917         rcu_read_unlock();
1918 }
1919
1920 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1921 {
1922         const unsigned long *gflags = &gl->gl_flags;
1923         char *p = buf;
1924
1925         if (test_bit(GLF_LOCK, gflags))
1926                 *p++ = 'l';
1927         if (test_bit(GLF_DEMOTE, gflags))
1928                 *p++ = 'D';
1929         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1930                 *p++ = 'd';
1931         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1932                 *p++ = 'p';
1933         if (test_bit(GLF_DIRTY, gflags))
1934                 *p++ = 'y';
1935         if (test_bit(GLF_LFLUSH, gflags))
1936                 *p++ = 'f';
1937         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1938                 *p++ = 'i';
1939         if (test_bit(GLF_REPLY_PENDING, gflags))
1940                 *p++ = 'r';
1941         if (test_bit(GLF_INITIAL, gflags))
1942                 *p++ = 'I';
1943         if (test_bit(GLF_FROZEN, gflags))
1944                 *p++ = 'F';
1945         if (test_bit(GLF_QUEUED, gflags))
1946                 *p++ = 'q';
1947         if (test_bit(GLF_LRU, gflags))
1948                 *p++ = 'L';
1949         if (gl->gl_object)
1950                 *p++ = 'o';
1951         if (test_bit(GLF_BLOCKING, gflags))
1952                 *p++ = 'b';
1953         *p = 0;
1954         return buf;
1955 }
1956
1957 /**
1958  * gfs2_dump_glock - print information about a glock
1959  * @seq: The seq_file struct
1960  * @gl: the glock
1961  * @fsid: If true, also dump the file system id
1962  *
1963  * The file format is as follows:
1964  * One line per object, capital letters are used to indicate objects
1965  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1966  * other objects are indented by a single space and follow the glock to
1967  * which they are related. Fields are indicated by lower case letters
1968  * followed by a colon and the field value, except for strings which are in
1969  * [] so that its possible to see if they are composed of spaces for
1970  * example. The field's are n = number (id of the object), f = flags,
1971  * t = type, s = state, r = refcount, e = error, p = pid.
1972  *
1973  */
1974
1975 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl, bool fsid)
1976 {
1977         const struct gfs2_glock_operations *glops = gl->gl_ops;
1978         unsigned long long dtime;
1979         const struct gfs2_holder *gh;
1980         char gflags_buf[32];
1981         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1982         char fs_id_buf[sizeof(sdp->sd_fsname) + 7];
1983
1984         memset(fs_id_buf, 0, sizeof(fs_id_buf));
1985         if (fsid && sdp) /* safety precaution */
1986                 sprintf(fs_id_buf, "fsid=%s: ", sdp->sd_fsname);
1987         dtime = jiffies - gl->gl_demote_time;
1988         dtime *= 1000000/HZ; /* demote time in uSec */
1989         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1990                 dtime = 0;
1991         gfs2_print_dbg(seq, "%sG:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d "
1992                        "v:%d r:%d m:%ld\n", fs_id_buf, state2str(gl->gl_state),
1993                   gl->gl_name.ln_type,
1994                   (unsigned long long)gl->gl_name.ln_number,
1995                   gflags2str(gflags_buf, gl),
1996                   state2str(gl->gl_target),
1997                   state2str(gl->gl_demote_state), dtime,
1998                   atomic_read(&gl->gl_ail_count),
1999                   atomic_read(&gl->gl_revokes),
2000                   (int)gl->gl_lockref.count, gl->gl_hold_time);
2001
2002         list_for_each_entry(gh, &gl->gl_holders, gh_list)
2003                 dump_holder(seq, gh, fs_id_buf);
2004
2005         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
2006                 glops->go_dump(seq, gl, fs_id_buf);
2007 }
2008
2009 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
2010 {
2011         struct gfs2_glock *gl = iter_ptr;
2012
2013         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
2014                    gl->gl_name.ln_type,
2015                    (unsigned long long)gl->gl_name.ln_number,
2016                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
2017                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
2018                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
2019                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
2020                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
2021                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
2022                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
2023                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
2024         return 0;
2025 }
2026
2027 static const char *gfs2_gltype[] = {
2028         "type",
2029         "reserved",
2030         "nondisk",
2031         "inode",
2032         "rgrp",
2033         "meta",
2034         "iopen",
2035         "flock",
2036         "plock",
2037         "quota",
2038         "journal",
2039 };
2040
2041 static const char *gfs2_stype[] = {
2042         [GFS2_LKS_SRTT]         = "srtt",
2043         [GFS2_LKS_SRTTVAR]      = "srttvar",
2044         [GFS2_LKS_SRTTB]        = "srttb",
2045         [GFS2_LKS_SRTTVARB]     = "srttvarb",
2046         [GFS2_LKS_SIRT]         = "sirt",
2047         [GFS2_LKS_SIRTVAR]      = "sirtvar",
2048         [GFS2_LKS_DCOUNT]       = "dlm",
2049         [GFS2_LKS_QCOUNT]       = "queue",
2050 };
2051
2052 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
2053
2054 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
2055 {
2056         struct gfs2_sbd *sdp = seq->private;
2057         loff_t pos = *(loff_t *)iter_ptr;
2058         unsigned index = pos >> 3;
2059         unsigned subindex = pos & 0x07;
2060         int i;
2061
2062         if (index == 0 && subindex != 0)
2063                 return 0;
2064
2065         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
2066                    (index == 0) ? "cpu": gfs2_stype[subindex]);
2067
2068         for_each_possible_cpu(i) {
2069                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
2070
2071                 if (index == 0)
2072                         seq_printf(seq, " %15u", i);
2073                 else
2074                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
2075                                    lkstats[index - 1].stats[subindex]);
2076         }
2077         seq_putc(seq, '\n');
2078         return 0;
2079 }
2080
2081 int __init gfs2_glock_init(void)
2082 {
2083         int i, ret;
2084
2085         ret = rhashtable_init(&gl_hash_table, &ht_parms);
2086         if (ret < 0)
2087                 return ret;
2088
2089         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
2090                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
2091         if (!glock_workqueue) {
2092                 rhashtable_destroy(&gl_hash_table);
2093                 return -ENOMEM;
2094         }
2095         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
2096                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
2097                                                 0);
2098         if (!gfs2_delete_workqueue) {
2099                 destroy_workqueue(glock_workqueue);
2100                 rhashtable_destroy(&gl_hash_table);
2101                 return -ENOMEM;
2102         }
2103
2104         ret = register_shrinker(&glock_shrinker);
2105         if (ret) {
2106                 destroy_workqueue(gfs2_delete_workqueue);
2107                 destroy_workqueue(glock_workqueue);
2108                 rhashtable_destroy(&gl_hash_table);
2109                 return ret;
2110         }
2111
2112         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
2113                 init_waitqueue_head(glock_wait_table + i);
2114
2115         return 0;
2116 }
2117
2118 void gfs2_glock_exit(void)
2119 {
2120         unregister_shrinker(&glock_shrinker);
2121         rhashtable_destroy(&gl_hash_table);
2122         destroy_workqueue(glock_workqueue);
2123         destroy_workqueue(gfs2_delete_workqueue);
2124 }
2125
2126 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
2127 {
2128         struct gfs2_glock *gl = gi->gl;
2129
2130         if (gl) {
2131                 if (n == 0)
2132                         return;
2133                 if (!lockref_put_not_zero(&gl->gl_lockref))
2134                         gfs2_glock_queue_put(gl);
2135         }
2136         for (;;) {
2137                 gl = rhashtable_walk_next(&gi->hti);
2138                 if (IS_ERR_OR_NULL(gl)) {
2139                         if (gl == ERR_PTR(-EAGAIN)) {
2140                                 n = 1;
2141                                 continue;
2142                         }
2143                         gl = NULL;
2144                         break;
2145                 }
2146                 if (gl->gl_name.ln_sbd != gi->sdp)
2147                         continue;
2148                 if (n <= 1) {
2149                         if (!lockref_get_not_dead(&gl->gl_lockref))
2150                                 continue;
2151                         break;
2152                 } else {
2153                         if (__lockref_is_dead(&gl->gl_lockref))
2154                                 continue;
2155                         n--;
2156                 }
2157         }
2158         gi->gl = gl;
2159 }
2160
2161 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
2162         __acquires(RCU)
2163 {
2164         struct gfs2_glock_iter *gi = seq->private;
2165         loff_t n;
2166
2167         /*
2168          * We can either stay where we are, skip to the next hash table
2169          * entry, or start from the beginning.
2170          */
2171         if (*pos < gi->last_pos) {
2172                 rhashtable_walk_exit(&gi->hti);
2173                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2174                 n = *pos + 1;
2175         } else {
2176                 n = *pos - gi->last_pos;
2177         }
2178
2179         rhashtable_walk_start(&gi->hti);
2180
2181         gfs2_glock_iter_next(gi, n);
2182         gi->last_pos = *pos;
2183         return gi->gl;
2184 }
2185
2186 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
2187                                  loff_t *pos)
2188 {
2189         struct gfs2_glock_iter *gi = seq->private;
2190
2191         (*pos)++;
2192         gi->last_pos = *pos;
2193         gfs2_glock_iter_next(gi, 1);
2194         return gi->gl;
2195 }
2196
2197 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2198         __releases(RCU)
2199 {
2200         struct gfs2_glock_iter *gi = seq->private;
2201
2202         rhashtable_walk_stop(&gi->hti);
2203 }
2204
2205 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2206 {
2207         dump_glock(seq, iter_ptr, false);
2208         return 0;
2209 }
2210
2211 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2212 {
2213         preempt_disable();
2214         if (*pos >= GFS2_NR_SBSTATS)
2215                 return NULL;
2216         return pos;
2217 }
2218
2219 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2220                                    loff_t *pos)
2221 {
2222         (*pos)++;
2223         if (*pos >= GFS2_NR_SBSTATS)
2224                 return NULL;
2225         return pos;
2226 }
2227
2228 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2229 {
2230         preempt_enable();
2231 }
2232
2233 static const struct seq_operations gfs2_glock_seq_ops = {
2234         .start = gfs2_glock_seq_start,
2235         .next  = gfs2_glock_seq_next,
2236         .stop  = gfs2_glock_seq_stop,
2237         .show  = gfs2_glock_seq_show,
2238 };
2239
2240 static const struct seq_operations gfs2_glstats_seq_ops = {
2241         .start = gfs2_glock_seq_start,
2242         .next  = gfs2_glock_seq_next,
2243         .stop  = gfs2_glock_seq_stop,
2244         .show  = gfs2_glstats_seq_show,
2245 };
2246
2247 static const struct seq_operations gfs2_sbstats_seq_ops = {
2248         .start = gfs2_sbstats_seq_start,
2249         .next  = gfs2_sbstats_seq_next,
2250         .stop  = gfs2_sbstats_seq_stop,
2251         .show  = gfs2_sbstats_seq_show,
2252 };
2253
2254 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2255
2256 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2257                               const struct seq_operations *ops)
2258 {
2259         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2260         if (ret == 0) {
2261                 struct seq_file *seq = file->private_data;
2262                 struct gfs2_glock_iter *gi = seq->private;
2263
2264                 gi->sdp = inode->i_private;
2265                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2266                 if (seq->buf)
2267                         seq->size = GFS2_SEQ_GOODSIZE;
2268                 /*
2269                  * Initially, we are "before" the first hash table entry; the
2270                  * first call to rhashtable_walk_next gets us the first entry.
2271                  */
2272                 gi->last_pos = -1;
2273                 gi->gl = NULL;
2274                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2275         }
2276         return ret;
2277 }
2278
2279 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2280 {
2281         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2282 }
2283
2284 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2285 {
2286         struct seq_file *seq = file->private_data;
2287         struct gfs2_glock_iter *gi = seq->private;
2288
2289         if (gi->gl)
2290                 gfs2_glock_put(gi->gl);
2291         rhashtable_walk_exit(&gi->hti);
2292         return seq_release_private(inode, file);
2293 }
2294
2295 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2296 {
2297         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2298 }
2299
2300 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2301 {
2302         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2303         if (ret == 0) {
2304                 struct seq_file *seq = file->private_data;
2305                 seq->private = inode->i_private;  /* sdp */
2306         }
2307         return ret;
2308 }
2309
2310 static const struct file_operations gfs2_glocks_fops = {
2311         .owner   = THIS_MODULE,
2312         .open    = gfs2_glocks_open,
2313         .read    = seq_read,
2314         .llseek  = seq_lseek,
2315         .release = gfs2_glocks_release,
2316 };
2317
2318 static const struct file_operations gfs2_glstats_fops = {
2319         .owner   = THIS_MODULE,
2320         .open    = gfs2_glstats_open,
2321         .read    = seq_read,
2322         .llseek  = seq_lseek,
2323         .release = gfs2_glocks_release,
2324 };
2325
2326 static const struct file_operations gfs2_sbstats_fops = {
2327         .owner   = THIS_MODULE,
2328         .open    = gfs2_sbstats_open,
2329         .read    = seq_read,
2330         .llseek  = seq_lseek,
2331         .release = seq_release,
2332 };
2333
2334 void gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2335 {
2336         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2337
2338         debugfs_create_file("glocks", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2339                             &gfs2_glocks_fops);
2340
2341         debugfs_create_file("glstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2342                             &gfs2_glstats_fops);
2343
2344         debugfs_create_file("sbstats", S_IFREG | S_IRUGO, sdp->debugfs_dir, sdp,
2345                             &gfs2_sbstats_fops);
2346 }
2347
2348 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2349 {
2350         debugfs_remove_recursive(sdp->debugfs_dir);
2351         sdp->debugfs_dir = NULL;
2352 }
2353
2354 void gfs2_register_debugfs(void)
2355 {
2356         gfs2_root = debugfs_create_dir("gfs2", NULL);
2357 }
2358
2359 void gfs2_unregister_debugfs(void)
2360 {
2361         debugfs_remove(gfs2_root);
2362         gfs2_root = NULL;
2363 }