Linux Audio

Check our new training course

Loading...
v4.10.11
   1/* -*- mode: c; c-basic-offset: 8; -*-
   2 * vim: noexpandtab sw=8 ts=8 sts=0:
   3 *
   4 * dlmglue.c
   5 *
   6 * Code which implements an OCFS2 specific interface to our DLM.
   7 *
   8 * Copyright (C) 2003, 2004 Oracle.  All rights reserved.
   9 *
  10 * This program is free software; you can redistribute it and/or
  11 * modify it under the terms of the GNU General Public
  12 * License as published by the Free Software Foundation; either
  13 * version 2 of the License, or (at your option) any later version.
  14 *
  15 * This program is distributed in the hope that it will be useful,
  16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18 * General Public License for more details.
  19 *
  20 * You should have received a copy of the GNU General Public
  21 * License along with this program; if not, write to the
  22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23 * Boston, MA 021110-1307, USA.
  24 */
  25
  26#include <linux/types.h>
  27#include <linux/slab.h>
  28#include <linux/highmem.h>
  29#include <linux/mm.h>
  30#include <linux/kthread.h>
  31#include <linux/pagemap.h>
  32#include <linux/debugfs.h>
  33#include <linux/seq_file.h>
  34#include <linux/time.h>
 
  35#include <linux/quotaops.h>
 
  36
  37#define MLOG_MASK_PREFIX ML_DLM_GLUE
  38#include <cluster/masklog.h>
  39
  40#include "ocfs2.h"
  41#include "ocfs2_lockingver.h"
  42
  43#include "alloc.h"
  44#include "dcache.h"
  45#include "dlmglue.h"
  46#include "extent_map.h"
  47#include "file.h"
  48#include "heartbeat.h"
  49#include "inode.h"
  50#include "journal.h"
  51#include "stackglue.h"
  52#include "slot_map.h"
  53#include "super.h"
  54#include "uptodate.h"
  55#include "quota.h"
  56#include "refcounttree.h"
  57#include "acl.h"
  58
  59#include "buffer_head_io.h"
  60
  61struct ocfs2_mask_waiter {
  62	struct list_head	mw_item;
  63	int			mw_status;
  64	struct completion	mw_complete;
  65	unsigned long		mw_mask;
  66	unsigned long		mw_goal;
  67#ifdef CONFIG_OCFS2_FS_STATS
  68	ktime_t			mw_lock_start;
  69#endif
  70};
  71
  72static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres);
  73static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres);
  74static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres);
  75static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres);
  76
  77/*
  78 * Return value from ->downconvert_worker functions.
  79 *
  80 * These control the precise actions of ocfs2_unblock_lock()
  81 * and ocfs2_process_blocked_lock()
  82 *
  83 */
  84enum ocfs2_unblock_action {
  85	UNBLOCK_CONTINUE	= 0, /* Continue downconvert */
  86	UNBLOCK_CONTINUE_POST	= 1, /* Continue downconvert, fire
  87				      * ->post_unlock callback */
  88	UNBLOCK_STOP_POST	= 2, /* Do not downconvert, fire
  89				      * ->post_unlock() callback. */
  90};
  91
  92struct ocfs2_unblock_ctl {
  93	int requeue;
  94	enum ocfs2_unblock_action unblock_action;
  95};
  96
  97/* Lockdep class keys */
  98struct lock_class_key lockdep_keys[OCFS2_NUM_LOCK_TYPES];
 
 
  99
 100static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
 101					int new_level);
 102static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres);
 103
 104static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
 105				     int blocking);
 106
 107static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
 108				       int blocking);
 109
 110static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
 111				     struct ocfs2_lock_res *lockres);
 112
 113static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres);
 114
 115static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
 116					    int new_level);
 117static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
 118					 int blocking);
 119
 120#define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
 121
 122/* This aids in debugging situations where a bad LVB might be involved. */
 123static void ocfs2_dump_meta_lvb_info(u64 level,
 124				     const char *function,
 125				     unsigned int line,
 126				     struct ocfs2_lock_res *lockres)
 127{
 128	struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
 129
 130	mlog(level, "LVB information for %s (called from %s:%u):\n",
 131	     lockres->l_name, function, line);
 132	mlog(level, "version: %u, clusters: %u, generation: 0x%x\n",
 133	     lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters),
 134	     be32_to_cpu(lvb->lvb_igeneration));
 135	mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n",
 136	     (unsigned long long)be64_to_cpu(lvb->lvb_isize),
 137	     be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid),
 138	     be16_to_cpu(lvb->lvb_imode));
 139	mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
 140	     "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink),
 141	     (long long)be64_to_cpu(lvb->lvb_iatime_packed),
 142	     (long long)be64_to_cpu(lvb->lvb_ictime_packed),
 143	     (long long)be64_to_cpu(lvb->lvb_imtime_packed),
 144	     be32_to_cpu(lvb->lvb_iattr));
 145}
 146
 147
 148/*
 149 * OCFS2 Lock Resource Operations
 150 *
 151 * These fine tune the behavior of the generic dlmglue locking infrastructure.
 152 *
 153 * The most basic of lock types can point ->l_priv to their respective
 154 * struct ocfs2_super and allow the default actions to manage things.
 155 *
 156 * Right now, each lock type also needs to implement an init function,
 157 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
 158 * should be called when the lock is no longer needed (i.e., object
 159 * destruction time).
 160 */
 161struct ocfs2_lock_res_ops {
 162	/*
 163	 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
 164	 * this callback if ->l_priv is not an ocfs2_super pointer
 165	 */
 166	struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *);
 167
 168	/*
 169	 * Optionally called in the downconvert thread after a
 170	 * successful downconvert. The lockres will not be referenced
 171	 * after this callback is called, so it is safe to free
 172	 * memory, etc.
 173	 *
 174	 * The exact semantics of when this is called are controlled
 175	 * by ->downconvert_worker()
 176	 */
 177	void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *);
 178
 179	/*
 180	 * Allow a lock type to add checks to determine whether it is
 181	 * safe to downconvert a lock. Return 0 to re-queue the
 182	 * downconvert at a later time, nonzero to continue.
 183	 *
 184	 * For most locks, the default checks that there are no
 185	 * incompatible holders are sufficient.
 186	 *
 187	 * Called with the lockres spinlock held.
 188	 */
 189	int (*check_downconvert)(struct ocfs2_lock_res *, int);
 190
 191	/*
 192	 * Allows a lock type to populate the lock value block. This
 193	 * is called on downconvert, and when we drop a lock.
 194	 *
 195	 * Locks that want to use this should set LOCK_TYPE_USES_LVB
 196	 * in the flags field.
 197	 *
 198	 * Called with the lockres spinlock held.
 199	 */
 200	void (*set_lvb)(struct ocfs2_lock_res *);
 201
 202	/*
 203	 * Called from the downconvert thread when it is determined
 204	 * that a lock will be downconverted. This is called without
 205	 * any locks held so the function can do work that might
 206	 * schedule (syncing out data, etc).
 207	 *
 208	 * This should return any one of the ocfs2_unblock_action
 209	 * values, depending on what it wants the thread to do.
 210	 */
 211	int (*downconvert_worker)(struct ocfs2_lock_res *, int);
 212
 213	/*
 214	 * LOCK_TYPE_* flags which describe the specific requirements
 215	 * of a lock type. Descriptions of each individual flag follow.
 216	 */
 217	int flags;
 218};
 219
 220/*
 221 * Some locks want to "refresh" potentially stale data when a
 222 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
 223 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
 224 * individual lockres l_flags member from the ast function. It is
 225 * expected that the locking wrapper will clear the
 226 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
 227 */
 228#define LOCK_TYPE_REQUIRES_REFRESH 0x1
 229
 230/*
 231 * Indicate that a lock type makes use of the lock value block. The
 232 * ->set_lvb lock type callback must be defined.
 233 */
 234#define LOCK_TYPE_USES_LVB		0x2
 235
 236static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = {
 237	.get_osb	= ocfs2_get_inode_osb,
 238	.flags		= 0,
 239};
 240
 241static struct ocfs2_lock_res_ops ocfs2_inode_inode_lops = {
 242	.get_osb	= ocfs2_get_inode_osb,
 243	.check_downconvert = ocfs2_check_meta_downconvert,
 244	.set_lvb	= ocfs2_set_meta_lvb,
 245	.downconvert_worker = ocfs2_data_convert_worker,
 246	.flags		= LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
 247};
 248
 249static struct ocfs2_lock_res_ops ocfs2_super_lops = {
 250	.flags		= LOCK_TYPE_REQUIRES_REFRESH,
 251};
 252
 253static struct ocfs2_lock_res_ops ocfs2_rename_lops = {
 254	.flags		= 0,
 255};
 256
 257static struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops = {
 258	.flags		= 0,
 259};
 260
 261static struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops = {
 
 
 
 
 262	.flags		= LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
 263};
 264
 265static struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
 266	.get_osb	= ocfs2_get_dentry_osb,
 267	.post_unlock	= ocfs2_dentry_post_unlock,
 268	.downconvert_worker = ocfs2_dentry_convert_worker,
 269	.flags		= 0,
 270};
 271
 272static struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
 273	.get_osb	= ocfs2_get_inode_osb,
 274	.flags		= 0,
 275};
 276
 277static struct ocfs2_lock_res_ops ocfs2_flock_lops = {
 278	.get_osb	= ocfs2_get_file_osb,
 279	.flags		= 0,
 280};
 281
 282static struct ocfs2_lock_res_ops ocfs2_qinfo_lops = {
 283	.set_lvb	= ocfs2_set_qinfo_lvb,
 284	.get_osb	= ocfs2_get_qinfo_osb,
 285	.flags		= LOCK_TYPE_REQUIRES_REFRESH | LOCK_TYPE_USES_LVB,
 286};
 287
 288static struct ocfs2_lock_res_ops ocfs2_refcount_block_lops = {
 289	.check_downconvert = ocfs2_check_refcount_downconvert,
 290	.downconvert_worker = ocfs2_refcount_convert_worker,
 291	.flags		= 0,
 292};
 293
 294static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
 295{
 296	return lockres->l_type == OCFS2_LOCK_TYPE_META ||
 297		lockres->l_type == OCFS2_LOCK_TYPE_RW ||
 298		lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
 299}
 300
 301static inline struct ocfs2_lock_res *ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb)
 302{
 303	return container_of(lksb, struct ocfs2_lock_res, l_lksb);
 304}
 305
 306static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
 307{
 308	BUG_ON(!ocfs2_is_inode_lock(lockres));
 309
 310	return (struct inode *) lockres->l_priv;
 311}
 312
 313static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres)
 314{
 315	BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY);
 316
 317	return (struct ocfs2_dentry_lock *)lockres->l_priv;
 318}
 319
 320static inline struct ocfs2_mem_dqinfo *ocfs2_lock_res_qinfo(struct ocfs2_lock_res *lockres)
 321{
 322	BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_QINFO);
 323
 324	return (struct ocfs2_mem_dqinfo *)lockres->l_priv;
 325}
 326
 327static inline struct ocfs2_refcount_tree *
 328ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res *res)
 329{
 330	return container_of(res, struct ocfs2_refcount_tree, rf_lockres);
 331}
 332
 333static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres)
 334{
 335	if (lockres->l_ops->get_osb)
 336		return lockres->l_ops->get_osb(lockres);
 337
 338	return (struct ocfs2_super *)lockres->l_priv;
 339}
 340
 341static int ocfs2_lock_create(struct ocfs2_super *osb,
 342			     struct ocfs2_lock_res *lockres,
 343			     int level,
 344			     u32 dlm_flags);
 345static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
 346						     int wanted);
 347static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
 348				   struct ocfs2_lock_res *lockres,
 349				   int level, unsigned long caller_ip);
 350static inline void ocfs2_cluster_unlock(struct ocfs2_super *osb,
 351					struct ocfs2_lock_res *lockres,
 352					int level)
 353{
 354	__ocfs2_cluster_unlock(osb, lockres, level, _RET_IP_);
 355}
 356
 357static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres);
 358static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres);
 359static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres);
 360static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level);
 361static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
 362					struct ocfs2_lock_res *lockres);
 363static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
 364						int convert);
 365#define ocfs2_log_dlm_error(_func, _err, _lockres) do {					\
 366	if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY)				\
 367		mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n",	\
 368		     _err, _func, _lockres->l_name);					\
 369	else										\
 370		mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n",	\
 371		     _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name,	\
 372		     (unsigned int)ocfs2_get_dentry_lock_ino(_lockres));		\
 373} while (0)
 374static int ocfs2_downconvert_thread(void *arg);
 375static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
 376					struct ocfs2_lock_res *lockres);
 377static int ocfs2_inode_lock_update(struct inode *inode,
 378				  struct buffer_head **bh);
 379static void ocfs2_drop_osb_locks(struct ocfs2_super *osb);
 380static inline int ocfs2_highest_compat_lock_level(int level);
 381static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
 382					      int new_level);
 383static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
 384				  struct ocfs2_lock_res *lockres,
 385				  int new_level,
 386				  int lvb,
 387				  unsigned int generation);
 388static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
 389				        struct ocfs2_lock_res *lockres);
 390static int ocfs2_cancel_convert(struct ocfs2_super *osb,
 391				struct ocfs2_lock_res *lockres);
 392
 393
 394static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
 395				  u64 blkno,
 396				  u32 generation,
 397				  char *name)
 398{
 399	int len;
 400
 401	BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
 402
 403	len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
 404		       ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
 405		       (long long)blkno, generation);
 406
 407	BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
 408
 409	mlog(0, "built lock resource with name: %s\n", name);
 410}
 411
 412static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
 413
 414static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
 415				       struct ocfs2_dlm_debug *dlm_debug)
 416{
 417	mlog(0, "Add tracking for lockres %s\n", res->l_name);
 418
 419	spin_lock(&ocfs2_dlm_tracking_lock);
 420	list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
 421	spin_unlock(&ocfs2_dlm_tracking_lock);
 422}
 423
 424static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
 425{
 426	spin_lock(&ocfs2_dlm_tracking_lock);
 427	if (!list_empty(&res->l_debug_list))
 428		list_del_init(&res->l_debug_list);
 429	spin_unlock(&ocfs2_dlm_tracking_lock);
 430}
 431
 432#ifdef CONFIG_OCFS2_FS_STATS
 433static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
 434{
 435	res->l_lock_refresh = 0;
 
 436	memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats));
 437	memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats));
 438}
 439
 440static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
 441				    struct ocfs2_mask_waiter *mw, int ret)
 442{
 443	u32 usec;
 444	ktime_t kt;
 445	struct ocfs2_lock_stats *stats;
 446
 447	if (level == LKM_PRMODE)
 448		stats = &res->l_lock_prmode;
 449	else if (level == LKM_EXMODE)
 450		stats = &res->l_lock_exmode;
 451	else
 452		return;
 453
 454	kt = ktime_sub(ktime_get(), mw->mw_lock_start);
 455	usec = ktime_to_us(kt);
 456
 457	stats->ls_gets++;
 458	stats->ls_total += ktime_to_ns(kt);
 459	/* overflow */
 460	if (unlikely(stats->ls_gets == 0)) {
 461		stats->ls_gets++;
 462		stats->ls_total = ktime_to_ns(kt);
 463	}
 464
 465	if (stats->ls_max < usec)
 466		stats->ls_max = usec;
 467
 468	if (ret)
 469		stats->ls_fail++;
 
 
 470}
 471
 472static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
 473{
 474	lockres->l_lock_refresh++;
 475}
 476
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 477static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
 478{
 479	mw->mw_lock_start = ktime_get();
 480}
 481#else
 482static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
 483{
 484}
 485static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res,
 486			   int level, struct ocfs2_mask_waiter *mw, int ret)
 487{
 488}
 489static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
 490{
 491}
 
 
 
 492static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
 493{
 494}
 495#endif
 496
 497static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
 498				       struct ocfs2_lock_res *res,
 499				       enum ocfs2_lock_type type,
 500				       struct ocfs2_lock_res_ops *ops,
 501				       void *priv)
 502{
 503	res->l_type          = type;
 504	res->l_ops           = ops;
 505	res->l_priv          = priv;
 506
 507	res->l_level         = DLM_LOCK_IV;
 508	res->l_requested     = DLM_LOCK_IV;
 509	res->l_blocking      = DLM_LOCK_IV;
 510	res->l_action        = OCFS2_AST_INVALID;
 511	res->l_unlock_action = OCFS2_UNLOCK_INVALID;
 512
 513	res->l_flags         = OCFS2_LOCK_INITIALIZED;
 514
 515	ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
 516
 517	ocfs2_init_lock_stats(res);
 518#ifdef CONFIG_DEBUG_LOCK_ALLOC
 519	if (type != OCFS2_LOCK_TYPE_OPEN)
 520		lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type],
 521				 &lockdep_keys[type], 0);
 522	else
 523		res->l_lockdep_map.key = NULL;
 524#endif
 525}
 526
 527void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
 528{
 529	/* This also clears out the lock status block */
 530	memset(res, 0, sizeof(struct ocfs2_lock_res));
 531	spin_lock_init(&res->l_lock);
 532	init_waitqueue_head(&res->l_event);
 533	INIT_LIST_HEAD(&res->l_blocked_list);
 534	INIT_LIST_HEAD(&res->l_mask_waiters);
 
 535}
 536
 537void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
 538			       enum ocfs2_lock_type type,
 539			       unsigned int generation,
 540			       struct inode *inode)
 541{
 542	struct ocfs2_lock_res_ops *ops;
 543
 544	switch(type) {
 545		case OCFS2_LOCK_TYPE_RW:
 546			ops = &ocfs2_inode_rw_lops;
 547			break;
 548		case OCFS2_LOCK_TYPE_META:
 549			ops = &ocfs2_inode_inode_lops;
 550			break;
 551		case OCFS2_LOCK_TYPE_OPEN:
 552			ops = &ocfs2_inode_open_lops;
 553			break;
 554		default:
 555			mlog_bug_on_msg(1, "type: %d\n", type);
 556			ops = NULL; /* thanks, gcc */
 557			break;
 558	};
 559
 560	ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
 561			      generation, res->l_name);
 562	ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
 563}
 564
 565static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
 566{
 567	struct inode *inode = ocfs2_lock_res_inode(lockres);
 568
 569	return OCFS2_SB(inode->i_sb);
 570}
 571
 572static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres)
 573{
 574	struct ocfs2_mem_dqinfo *info = lockres->l_priv;
 575
 576	return OCFS2_SB(info->dqi_gi.dqi_sb);
 577}
 578
 579static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres)
 580{
 581	struct ocfs2_file_private *fp = lockres->l_priv;
 582
 583	return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb);
 584}
 585
 586static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
 587{
 588	__be64 inode_blkno_be;
 589
 590	memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
 591	       sizeof(__be64));
 592
 593	return be64_to_cpu(inode_blkno_be);
 594}
 595
 596static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
 597{
 598	struct ocfs2_dentry_lock *dl = lockres->l_priv;
 599
 600	return OCFS2_SB(dl->dl_inode->i_sb);
 601}
 602
 603void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
 604				u64 parent, struct inode *inode)
 605{
 606	int len;
 607	u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
 608	__be64 inode_blkno_be = cpu_to_be64(inode_blkno);
 609	struct ocfs2_lock_res *lockres = &dl->dl_lockres;
 610
 611	ocfs2_lock_res_init_once(lockres);
 612
 613	/*
 614	 * Unfortunately, the standard lock naming scheme won't work
 615	 * here because we have two 16 byte values to use. Instead,
 616	 * we'll stuff the inode number as a binary value. We still
 617	 * want error prints to show something without garbling the
 618	 * display, so drop a null byte in there before the inode
 619	 * number. A future version of OCFS2 will likely use all
 620	 * binary lock names. The stringified names have been a
 621	 * tremendous aid in debugging, but now that the debugfs
 622	 * interface exists, we can mangle things there if need be.
 623	 *
 624	 * NOTE: We also drop the standard "pad" value (the total lock
 625	 * name size stays the same though - the last part is all
 626	 * zeros due to the memset in ocfs2_lock_res_init_once()
 627	 */
 628	len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
 629		       "%c%016llx",
 630		       ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
 631		       (long long)parent);
 632
 633	BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
 634
 635	memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
 636	       sizeof(__be64));
 637
 638	ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
 639				   OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
 640				   dl);
 641}
 642
 643static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
 644				      struct ocfs2_super *osb)
 645{
 646	/* Superblock lockres doesn't come from a slab so we call init
 647	 * once on it manually.  */
 648	ocfs2_lock_res_init_once(res);
 649	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
 650			      0, res->l_name);
 651	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
 652				   &ocfs2_super_lops, osb);
 653}
 654
 655static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
 656				       struct ocfs2_super *osb)
 657{
 658	/* Rename lockres doesn't come from a slab so we call init
 659	 * once on it manually.  */
 660	ocfs2_lock_res_init_once(res);
 661	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
 662	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
 663				   &ocfs2_rename_lops, osb);
 664}
 665
 666static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res,
 667					 struct ocfs2_super *osb)
 668{
 669	/* nfs_sync lockres doesn't come from a slab so we call init
 670	 * once on it manually.  */
 671	ocfs2_lock_res_init_once(res);
 672	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name);
 673	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC,
 674				   &ocfs2_nfs_sync_lops, osb);
 675}
 676
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 677static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res,
 678					    struct ocfs2_super *osb)
 679{
 680	ocfs2_lock_res_init_once(res);
 681	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name);
 682	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN,
 683				   &ocfs2_orphan_scan_lops, osb);
 684}
 685
 686void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres,
 687			      struct ocfs2_file_private *fp)
 688{
 689	struct inode *inode = fp->fp_file->f_mapping->host;
 690	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 691
 692	ocfs2_lock_res_init_once(lockres);
 693	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno,
 694			      inode->i_generation, lockres->l_name);
 695	ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
 696				   OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops,
 697				   fp);
 698	lockres->l_flags |= OCFS2_LOCK_NOCACHE;
 699}
 700
 701void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres,
 702			       struct ocfs2_mem_dqinfo *info)
 703{
 704	ocfs2_lock_res_init_once(lockres);
 705	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type,
 706			      0, lockres->l_name);
 707	ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres,
 708				   OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops,
 709				   info);
 710}
 711
 712void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres,
 713				  struct ocfs2_super *osb, u64 ref_blkno,
 714				  unsigned int generation)
 715{
 716	ocfs2_lock_res_init_once(lockres);
 717	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno,
 718			      generation, lockres->l_name);
 719	ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT,
 720				   &ocfs2_refcount_block_lops, osb);
 721}
 722
 723void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
 724{
 725	if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
 726		return;
 727
 728	ocfs2_remove_lockres_tracking(res);
 729
 730	mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
 731			"Lockres %s is on the blocked list\n",
 732			res->l_name);
 733	mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
 734			"Lockres %s has mask waiters pending\n",
 735			res->l_name);
 736	mlog_bug_on_msg(spin_is_locked(&res->l_lock),
 737			"Lockres %s is locked\n",
 738			res->l_name);
 739	mlog_bug_on_msg(res->l_ro_holders,
 740			"Lockres %s has %u ro holders\n",
 741			res->l_name, res->l_ro_holders);
 742	mlog_bug_on_msg(res->l_ex_holders,
 743			"Lockres %s has %u ex holders\n",
 744			res->l_name, res->l_ex_holders);
 745
 746	/* Need to clear out the lock status block for the dlm */
 747	memset(&res->l_lksb, 0, sizeof(res->l_lksb));
 748
 749	res->l_flags = 0UL;
 750}
 751
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 752static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
 753				     int level)
 754{
 755	BUG_ON(!lockres);
 756
 757	switch(level) {
 758	case DLM_LOCK_EX:
 759		lockres->l_ex_holders++;
 760		break;
 761	case DLM_LOCK_PR:
 762		lockres->l_ro_holders++;
 763		break;
 764	default:
 765		BUG();
 766	}
 767}
 768
 769static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
 770				     int level)
 771{
 772	BUG_ON(!lockres);
 773
 774	switch(level) {
 775	case DLM_LOCK_EX:
 776		BUG_ON(!lockres->l_ex_holders);
 777		lockres->l_ex_holders--;
 778		break;
 779	case DLM_LOCK_PR:
 780		BUG_ON(!lockres->l_ro_holders);
 781		lockres->l_ro_holders--;
 782		break;
 783	default:
 784		BUG();
 785	}
 786}
 787
 788/* WARNING: This function lives in a world where the only three lock
 789 * levels are EX, PR, and NL. It *will* have to be adjusted when more
 790 * lock types are added. */
 791static inline int ocfs2_highest_compat_lock_level(int level)
 792{
 793	int new_level = DLM_LOCK_EX;
 794
 795	if (level == DLM_LOCK_EX)
 796		new_level = DLM_LOCK_NL;
 797	else if (level == DLM_LOCK_PR)
 798		new_level = DLM_LOCK_PR;
 799	return new_level;
 800}
 801
 802static void lockres_set_flags(struct ocfs2_lock_res *lockres,
 803			      unsigned long newflags)
 804{
 805	struct ocfs2_mask_waiter *mw, *tmp;
 806
 807 	assert_spin_locked(&lockres->l_lock);
 808
 809	lockres->l_flags = newflags;
 810
 811	list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) {
 812		if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
 813			continue;
 814
 815		list_del_init(&mw->mw_item);
 816		mw->mw_status = 0;
 817		complete(&mw->mw_complete);
 
 818	}
 819}
 820static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
 821{
 822	lockres_set_flags(lockres, lockres->l_flags | or);
 823}
 824static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
 825				unsigned long clear)
 826{
 827	lockres_set_flags(lockres, lockres->l_flags & ~clear);
 828}
 829
 830static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
 831{
 832	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
 833	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
 834	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
 835	BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
 836
 837	lockres->l_level = lockres->l_requested;
 838	if (lockres->l_level <=
 839	    ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
 840		lockres->l_blocking = DLM_LOCK_NL;
 841		lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
 842	}
 843	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
 844}
 845
 846static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
 847{
 848	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
 849	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
 850
 851	/* Convert from RO to EX doesn't really need anything as our
 852	 * information is already up to data. Convert from NL to
 853	 * *anything* however should mark ourselves as needing an
 854	 * update */
 855	if (lockres->l_level == DLM_LOCK_NL &&
 856	    lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
 857		lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
 858
 859	lockres->l_level = lockres->l_requested;
 860
 861	/*
 862	 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
 863	 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
 864	 * downconverting the lock before the upconvert has fully completed.
 865	 * Do not prevent the dc thread from downconverting if NONBLOCK lock
 866	 * had already returned.
 867	 */
 868	if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED))
 869		lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
 870	else
 871		lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED);
 872
 873	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
 874}
 875
 876static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
 877{
 878	BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY)));
 879	BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
 880
 881	if (lockres->l_requested > DLM_LOCK_NL &&
 882	    !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
 883	    lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
 884		lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
 885
 886	lockres->l_level = lockres->l_requested;
 887	lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
 888	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
 889}
 890
 891static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
 892				     int level)
 893{
 894	int needs_downconvert = 0;
 895
 896	assert_spin_locked(&lockres->l_lock);
 897
 898	if (level > lockres->l_blocking) {
 899		/* only schedule a downconvert if we haven't already scheduled
 900		 * one that goes low enough to satisfy the level we're
 901		 * blocking.  this also catches the case where we get
 902		 * duplicate BASTs */
 903		if (ocfs2_highest_compat_lock_level(level) <
 904		    ocfs2_highest_compat_lock_level(lockres->l_blocking))
 905			needs_downconvert = 1;
 906
 907		lockres->l_blocking = level;
 908	}
 909
 910	mlog(ML_BASTS, "lockres %s, block %d, level %d, l_block %d, dwn %d\n",
 911	     lockres->l_name, level, lockres->l_level, lockres->l_blocking,
 912	     needs_downconvert);
 913
 914	if (needs_downconvert)
 915		lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
 916	mlog(0, "needs_downconvert = %d\n", needs_downconvert);
 917	return needs_downconvert;
 918}
 919
 920/*
 921 * OCFS2_LOCK_PENDING and l_pending_gen.
 922 *
 923 * Why does OCFS2_LOCK_PENDING exist?  To close a race between setting
 924 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock().  See ocfs2_unblock_lock()
 925 * for more details on the race.
 926 *
 927 * OCFS2_LOCK_PENDING closes the race quite nicely.  However, it introduces
 928 * a race on itself.  In o2dlm, we can get the ast before ocfs2_dlm_lock()
 929 * returns.  The ast clears OCFS2_LOCK_BUSY, and must therefore clear
 930 * OCFS2_LOCK_PENDING at the same time.  When ocfs2_dlm_lock() returns,
 931 * the caller is going to try to clear PENDING again.  If nothing else is
 932 * happening, __lockres_clear_pending() sees PENDING is unset and does
 933 * nothing.
 934 *
 935 * But what if another path (eg downconvert thread) has just started a
 936 * new locking action?  The other path has re-set PENDING.  Our path
 937 * cannot clear PENDING, because that will re-open the original race
 938 * window.
 939 *
 940 * [Example]
 941 *
 942 * ocfs2_meta_lock()
 943 *  ocfs2_cluster_lock()
 944 *   set BUSY
 945 *   set PENDING
 946 *   drop l_lock
 947 *   ocfs2_dlm_lock()
 948 *    ocfs2_locking_ast()		ocfs2_downconvert_thread()
 949 *     clear PENDING			 ocfs2_unblock_lock()
 950 *					  take_l_lock
 951 *					  !BUSY
 952 *					  ocfs2_prepare_downconvert()
 953 *					   set BUSY
 954 *					   set PENDING
 955 *					  drop l_lock
 956 *   take l_lock
 957 *   clear PENDING
 958 *   drop l_lock
 959 *			<window>
 960 *					  ocfs2_dlm_lock()
 961 *
 962 * So as you can see, we now have a window where l_lock is not held,
 963 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
 964 *
 965 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
 966 * set by ocfs2_prepare_downconvert().  That wasn't nice.
 967 *
 968 * To solve this we introduce l_pending_gen.  A call to
 969 * lockres_clear_pending() will only do so when it is passed a generation
 970 * number that matches the lockres.  lockres_set_pending() will return the
 971 * current generation number.  When ocfs2_cluster_lock() goes to clear
 972 * PENDING, it passes the generation it got from set_pending().  In our
 973 * example above, the generation numbers will *not* match.  Thus,
 974 * ocfs2_cluster_lock() will not clear the PENDING set by
 975 * ocfs2_prepare_downconvert().
 976 */
 977
 978/* Unlocked version for ocfs2_locking_ast() */
 979static void __lockres_clear_pending(struct ocfs2_lock_res *lockres,
 980				    unsigned int generation,
 981				    struct ocfs2_super *osb)
 982{
 983	assert_spin_locked(&lockres->l_lock);
 984
 985	/*
 986	 * The ast and locking functions can race us here.  The winner
 987	 * will clear pending, the loser will not.
 988	 */
 989	if (!(lockres->l_flags & OCFS2_LOCK_PENDING) ||
 990	    (lockres->l_pending_gen != generation))
 991		return;
 992
 993	lockres_clear_flags(lockres, OCFS2_LOCK_PENDING);
 994	lockres->l_pending_gen++;
 995
 996	/*
 997	 * The downconvert thread may have skipped us because we
 998	 * were PENDING.  Wake it up.
 999	 */
1000	if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1001		ocfs2_wake_downconvert_thread(osb);
1002}
1003
1004/* Locked version for callers of ocfs2_dlm_lock() */
1005static void lockres_clear_pending(struct ocfs2_lock_res *lockres,
1006				  unsigned int generation,
1007				  struct ocfs2_super *osb)
1008{
1009	unsigned long flags;
1010
1011	spin_lock_irqsave(&lockres->l_lock, flags);
1012	__lockres_clear_pending(lockres, generation, osb);
1013	spin_unlock_irqrestore(&lockres->l_lock, flags);
1014}
1015
1016static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres)
1017{
1018	assert_spin_locked(&lockres->l_lock);
1019	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
1020
1021	lockres_or_flags(lockres, OCFS2_LOCK_PENDING);
1022
1023	return lockres->l_pending_gen;
1024}
1025
1026static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level)
1027{
1028	struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1029	struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1030	int needs_downconvert;
1031	unsigned long flags;
1032
1033	BUG_ON(level <= DLM_LOCK_NL);
1034
1035	mlog(ML_BASTS, "BAST fired for lockres %s, blocking %d, level %d, "
1036	     "type %s\n", lockres->l_name, level, lockres->l_level,
1037	     ocfs2_lock_type_string(lockres->l_type));
1038
1039	/*
1040	 * We can skip the bast for locks which don't enable caching -
1041	 * they'll be dropped at the earliest possible time anyway.
1042	 */
1043	if (lockres->l_flags & OCFS2_LOCK_NOCACHE)
1044		return;
1045
1046	spin_lock_irqsave(&lockres->l_lock, flags);
1047	needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
1048	if (needs_downconvert)
1049		ocfs2_schedule_blocked_lock(osb, lockres);
1050	spin_unlock_irqrestore(&lockres->l_lock, flags);
1051
1052	wake_up(&lockres->l_event);
1053
1054	ocfs2_wake_downconvert_thread(osb);
1055}
1056
1057static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb)
1058{
1059	struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1060	struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1061	unsigned long flags;
1062	int status;
1063
1064	spin_lock_irqsave(&lockres->l_lock, flags);
1065
1066	status = ocfs2_dlm_lock_status(&lockres->l_lksb);
1067
1068	if (status == -EAGAIN) {
1069		lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1070		goto out;
1071	}
1072
1073	if (status) {
1074		mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n",
1075		     lockres->l_name, status);
1076		spin_unlock_irqrestore(&lockres->l_lock, flags);
1077		return;
1078	}
1079
1080	mlog(ML_BASTS, "AST fired for lockres %s, action %d, unlock %d, "
1081	     "level %d => %d\n", lockres->l_name, lockres->l_action,
1082	     lockres->l_unlock_action, lockres->l_level, lockres->l_requested);
1083
1084	switch(lockres->l_action) {
1085	case OCFS2_AST_ATTACH:
1086		ocfs2_generic_handle_attach_action(lockres);
1087		lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
1088		break;
1089	case OCFS2_AST_CONVERT:
1090		ocfs2_generic_handle_convert_action(lockres);
1091		break;
1092	case OCFS2_AST_DOWNCONVERT:
1093		ocfs2_generic_handle_downconvert_action(lockres);
1094		break;
1095	default:
1096		mlog(ML_ERROR, "lockres %s: AST fired with invalid action: %u, "
1097		     "flags 0x%lx, unlock: %u\n",
1098		     lockres->l_name, lockres->l_action, lockres->l_flags,
1099		     lockres->l_unlock_action);
1100		BUG();
1101	}
1102out:
1103	/* set it to something invalid so if we get called again we
1104	 * can catch it. */
1105	lockres->l_action = OCFS2_AST_INVALID;
1106
1107	/* Did we try to cancel this lock?  Clear that state */
1108	if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT)
1109		lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1110
1111	/*
1112	 * We may have beaten the locking functions here.  We certainly
1113	 * know that dlm_lock() has been called :-)
1114	 * Because we can't have two lock calls in flight at once, we
1115	 * can use lockres->l_pending_gen.
1116	 */
1117	__lockres_clear_pending(lockres, lockres->l_pending_gen,  osb);
1118
1119	wake_up(&lockres->l_event);
1120	spin_unlock_irqrestore(&lockres->l_lock, flags);
1121}
1122
1123static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error)
1124{
1125	struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1126	unsigned long flags;
1127
1128	mlog(ML_BASTS, "UNLOCK AST fired for lockres %s, action = %d\n",
1129	     lockres->l_name, lockres->l_unlock_action);
1130
1131	spin_lock_irqsave(&lockres->l_lock, flags);
1132	if (error) {
1133		mlog(ML_ERROR, "Dlm passes error %d for lock %s, "
1134		     "unlock_action %d\n", error, lockres->l_name,
1135		     lockres->l_unlock_action);
1136		spin_unlock_irqrestore(&lockres->l_lock, flags);
1137		return;
1138	}
1139
1140	switch(lockres->l_unlock_action) {
1141	case OCFS2_UNLOCK_CANCEL_CONVERT:
1142		mlog(0, "Cancel convert success for %s\n", lockres->l_name);
1143		lockres->l_action = OCFS2_AST_INVALID;
1144		/* Downconvert thread may have requeued this lock, we
1145		 * need to wake it. */
1146		if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1147			ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres));
1148		break;
1149	case OCFS2_UNLOCK_DROP_LOCK:
1150		lockres->l_level = DLM_LOCK_IV;
1151		break;
1152	default:
1153		BUG();
1154	}
1155
1156	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1157	lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1158	wake_up(&lockres->l_event);
1159	spin_unlock_irqrestore(&lockres->l_lock, flags);
1160}
1161
1162/*
1163 * This is the filesystem locking protocol.  It provides the lock handling
1164 * hooks for the underlying DLM.  It has a maximum version number.
1165 * The version number allows interoperability with systems running at
1166 * the same major number and an equal or smaller minor number.
1167 *
1168 * Whenever the filesystem does new things with locks (adds or removes a
1169 * lock, orders them differently, does different things underneath a lock),
1170 * the version must be changed.  The protocol is negotiated when joining
1171 * the dlm domain.  A node may join the domain if its major version is
1172 * identical to all other nodes and its minor version is greater than
1173 * or equal to all other nodes.  When its minor version is greater than
1174 * the other nodes, it will run at the minor version specified by the
1175 * other nodes.
1176 *
1177 * If a locking change is made that will not be compatible with older
1178 * versions, the major number must be increased and the minor version set
1179 * to zero.  If a change merely adds a behavior that can be disabled when
1180 * speaking to older versions, the minor version must be increased.  If a
1181 * change adds a fully backwards compatible change (eg, LVB changes that
1182 * are just ignored by older versions), the version does not need to be
1183 * updated.
1184 */
1185static struct ocfs2_locking_protocol lproto = {
1186	.lp_max_version = {
1187		.pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
1188		.pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
1189	},
1190	.lp_lock_ast		= ocfs2_locking_ast,
1191	.lp_blocking_ast	= ocfs2_blocking_ast,
1192	.lp_unlock_ast		= ocfs2_unlock_ast,
1193};
1194
1195void ocfs2_set_locking_protocol(void)
1196{
1197	ocfs2_stack_glue_set_max_proto_version(&lproto.lp_max_version);
1198}
1199
1200static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
1201						int convert)
1202{
1203	unsigned long flags;
1204
1205	spin_lock_irqsave(&lockres->l_lock, flags);
1206	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1207	lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1208	if (convert)
1209		lockres->l_action = OCFS2_AST_INVALID;
1210	else
1211		lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1212	spin_unlock_irqrestore(&lockres->l_lock, flags);
1213
1214	wake_up(&lockres->l_event);
1215}
1216
1217/* Note: If we detect another process working on the lock (i.e.,
1218 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1219 * to do the right thing in that case.
1220 */
1221static int ocfs2_lock_create(struct ocfs2_super *osb,
1222			     struct ocfs2_lock_res *lockres,
1223			     int level,
1224			     u32 dlm_flags)
1225{
1226	int ret = 0;
1227	unsigned long flags;
1228	unsigned int gen;
1229
1230	mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level,
1231	     dlm_flags);
1232
1233	spin_lock_irqsave(&lockres->l_lock, flags);
1234	if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
1235	    (lockres->l_flags & OCFS2_LOCK_BUSY)) {
1236		spin_unlock_irqrestore(&lockres->l_lock, flags);
1237		goto bail;
1238	}
1239
1240	lockres->l_action = OCFS2_AST_ATTACH;
1241	lockres->l_requested = level;
1242	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1243	gen = lockres_set_pending(lockres);
1244	spin_unlock_irqrestore(&lockres->l_lock, flags);
1245
1246	ret = ocfs2_dlm_lock(osb->cconn,
1247			     level,
1248			     &lockres->l_lksb,
1249			     dlm_flags,
1250			     lockres->l_name,
1251			     OCFS2_LOCK_ID_MAX_LEN - 1);
1252	lockres_clear_pending(lockres, gen, osb);
1253	if (ret) {
1254		ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1255		ocfs2_recover_from_dlm_error(lockres, 1);
1256	}
1257
1258	mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name);
1259
1260bail:
1261	return ret;
1262}
1263
1264static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
1265					int flag)
1266{
1267	unsigned long flags;
1268	int ret;
1269
1270	spin_lock_irqsave(&lockres->l_lock, flags);
1271	ret = lockres->l_flags & flag;
1272	spin_unlock_irqrestore(&lockres->l_lock, flags);
1273
1274	return ret;
1275}
1276
1277static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
1278
1279{
1280	wait_event(lockres->l_event,
1281		   !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
1282}
1283
1284static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
1285
1286{
1287	wait_event(lockres->l_event,
1288		   !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
1289}
1290
1291/* predict what lock level we'll be dropping down to on behalf
1292 * of another node, and return true if the currently wanted
1293 * level will be compatible with it. */
1294static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
1295						     int wanted)
1296{
1297	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
1298
1299	return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
1300}
1301
1302static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
1303{
1304	INIT_LIST_HEAD(&mw->mw_item);
1305	init_completion(&mw->mw_complete);
1306	ocfs2_init_start_time(mw);
1307}
1308
1309static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
1310{
1311	wait_for_completion(&mw->mw_complete);
1312	/* Re-arm the completion in case we want to wait on it again */
1313	reinit_completion(&mw->mw_complete);
1314	return mw->mw_status;
1315}
1316
1317static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
1318				    struct ocfs2_mask_waiter *mw,
1319				    unsigned long mask,
1320				    unsigned long goal)
1321{
1322	BUG_ON(!list_empty(&mw->mw_item));
1323
1324	assert_spin_locked(&lockres->l_lock);
1325
1326	list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
1327	mw->mw_mask = mask;
1328	mw->mw_goal = goal;
 
1329}
1330
1331/* returns 0 if the mw that was removed was already satisfied, -EBUSY
1332 * if the mask still hadn't reached its goal */
1333static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1334				      struct ocfs2_mask_waiter *mw)
1335{
1336	int ret = 0;
1337
1338	assert_spin_locked(&lockres->l_lock);
1339	if (!list_empty(&mw->mw_item)) {
1340		if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
1341			ret = -EBUSY;
1342
1343		list_del_init(&mw->mw_item);
1344		init_completion(&mw->mw_complete);
 
1345	}
1346
1347	return ret;
1348}
1349
1350static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1351				      struct ocfs2_mask_waiter *mw)
1352{
1353	unsigned long flags;
1354	int ret = 0;
1355
1356	spin_lock_irqsave(&lockres->l_lock, flags);
1357	ret = __lockres_remove_mask_waiter(lockres, mw);
1358	spin_unlock_irqrestore(&lockres->l_lock, flags);
1359
1360	return ret;
1361
1362}
1363
1364static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw,
1365					     struct ocfs2_lock_res *lockres)
1366{
1367	int ret;
1368
1369	ret = wait_for_completion_interruptible(&mw->mw_complete);
1370	if (ret)
1371		lockres_remove_mask_waiter(lockres, mw);
1372	else
1373		ret = mw->mw_status;
1374	/* Re-arm the completion in case we want to wait on it again */
1375	reinit_completion(&mw->mw_complete);
1376	return ret;
1377}
1378
1379static int __ocfs2_cluster_lock(struct ocfs2_super *osb,
1380				struct ocfs2_lock_res *lockres,
1381				int level,
1382				u32 lkm_flags,
1383				int arg_flags,
1384				int l_subclass,
1385				unsigned long caller_ip)
1386{
1387	struct ocfs2_mask_waiter mw;
1388	int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
1389	int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1390	unsigned long flags;
1391	unsigned int gen;
1392	int noqueue_attempted = 0;
1393	int dlm_locked = 0;
1394	int kick_dc = 0;
1395
1396	if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED)) {
1397		mlog_errno(-EINVAL);
1398		return -EINVAL;
1399	}
1400
1401	ocfs2_init_mask_waiter(&mw);
1402
1403	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
1404		lkm_flags |= DLM_LKF_VALBLK;
1405
1406again:
1407	wait = 0;
1408
1409	spin_lock_irqsave(&lockres->l_lock, flags);
1410
1411	if (catch_signals && signal_pending(current)) {
1412		ret = -ERESTARTSYS;
1413		goto unlock;
1414	}
1415
1416	mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
1417			"Cluster lock called on freeing lockres %s! flags "
1418			"0x%lx\n", lockres->l_name, lockres->l_flags);
1419
1420	/* We only compare against the currently granted level
1421	 * here. If the lock is blocked waiting on a downconvert,
1422	 * we'll get caught below. */
1423	if (lockres->l_flags & OCFS2_LOCK_BUSY &&
1424	    level > lockres->l_level) {
1425		/* is someone sitting in dlm_lock? If so, wait on
1426		 * them. */
1427		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1428		wait = 1;
1429		goto unlock;
1430	}
1431
1432	if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
1433		/*
1434		 * We've upconverted. If the lock now has a level we can
1435		 * work with, we take it. If, however, the lock is not at the
1436		 * required level, we go thru the full cycle. One way this could
1437		 * happen is if a process requesting an upconvert to PR is
1438		 * closely followed by another requesting upconvert to an EX.
1439		 * If the process requesting EX lands here, we want it to
1440		 * continue attempting to upconvert and let the process
1441		 * requesting PR take the lock.
1442		 * If multiple processes request upconvert to PR, the first one
1443		 * here will take the lock. The others will have to go thru the
1444		 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1445		 * downconvert request.
1446		 */
1447		if (level <= lockres->l_level)
1448			goto update_holders;
1449	}
1450
1451	if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
1452	    !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
1453		/* is the lock is currently blocked on behalf of
1454		 * another node */
1455		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
1456		wait = 1;
1457		goto unlock;
1458	}
1459
1460	if (level > lockres->l_level) {
1461		if (noqueue_attempted > 0) {
1462			ret = -EAGAIN;
1463			goto unlock;
1464		}
1465		if (lkm_flags & DLM_LKF_NOQUEUE)
1466			noqueue_attempted = 1;
1467
1468		if (lockres->l_action != OCFS2_AST_INVALID)
1469			mlog(ML_ERROR, "lockres %s has action %u pending\n",
1470			     lockres->l_name, lockres->l_action);
1471
1472		if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1473			lockres->l_action = OCFS2_AST_ATTACH;
1474			lkm_flags &= ~DLM_LKF_CONVERT;
1475		} else {
1476			lockres->l_action = OCFS2_AST_CONVERT;
1477			lkm_flags |= DLM_LKF_CONVERT;
1478		}
1479
1480		lockres->l_requested = level;
1481		lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1482		gen = lockres_set_pending(lockres);
1483		spin_unlock_irqrestore(&lockres->l_lock, flags);
1484
1485		BUG_ON(level == DLM_LOCK_IV);
1486		BUG_ON(level == DLM_LOCK_NL);
1487
1488		mlog(ML_BASTS, "lockres %s, convert from %d to %d\n",
1489		     lockres->l_name, lockres->l_level, level);
1490
1491		/* call dlm_lock to upgrade lock now */
1492		ret = ocfs2_dlm_lock(osb->cconn,
1493				     level,
1494				     &lockres->l_lksb,
1495				     lkm_flags,
1496				     lockres->l_name,
1497				     OCFS2_LOCK_ID_MAX_LEN - 1);
1498		lockres_clear_pending(lockres, gen, osb);
1499		if (ret) {
1500			if (!(lkm_flags & DLM_LKF_NOQUEUE) ||
1501			    (ret != -EAGAIN)) {
1502				ocfs2_log_dlm_error("ocfs2_dlm_lock",
1503						    ret, lockres);
1504			}
1505			ocfs2_recover_from_dlm_error(lockres, 1);
1506			goto out;
1507		}
1508		dlm_locked = 1;
1509
1510		mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1511		     lockres->l_name);
1512
1513		/* At this point we've gone inside the dlm and need to
1514		 * complete our work regardless. */
1515		catch_signals = 0;
1516
1517		/* wait for busy to clear and carry on */
1518		goto again;
1519	}
1520
1521update_holders:
1522	/* Ok, if we get here then we're good to go. */
1523	ocfs2_inc_holders(lockres, level);
1524
1525	ret = 0;
1526unlock:
1527	lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1528
1529	/* ocfs2_unblock_lock reques on seeing OCFS2_LOCK_UPCONVERT_FINISHING */
1530	kick_dc = (lockres->l_flags & OCFS2_LOCK_BLOCKED);
1531
1532	spin_unlock_irqrestore(&lockres->l_lock, flags);
1533	if (kick_dc)
1534		ocfs2_wake_downconvert_thread(osb);
1535out:
1536	/*
1537	 * This is helping work around a lock inversion between the page lock
1538	 * and dlm locks.  One path holds the page lock while calling aops
1539	 * which block acquiring dlm locks.  The voting thread holds dlm
1540	 * locks while acquiring page locks while down converting data locks.
1541	 * This block is helping an aop path notice the inversion and back
1542	 * off to unlock its page lock before trying the dlm lock again.
1543	 */
1544	if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1545	    mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1546		wait = 0;
1547		spin_lock_irqsave(&lockres->l_lock, flags);
1548		if (__lockres_remove_mask_waiter(lockres, &mw)) {
1549			if (dlm_locked)
1550				lockres_or_flags(lockres,
1551					OCFS2_LOCK_NONBLOCK_FINISHED);
1552			spin_unlock_irqrestore(&lockres->l_lock, flags);
1553			ret = -EAGAIN;
1554		} else {
1555			spin_unlock_irqrestore(&lockres->l_lock, flags);
1556			goto again;
1557		}
1558	}
1559	if (wait) {
1560		ret = ocfs2_wait_for_mask(&mw);
1561		if (ret == 0)
1562			goto again;
1563		mlog_errno(ret);
1564	}
1565	ocfs2_update_lock_stats(lockres, level, &mw, ret);
1566
1567#ifdef CONFIG_DEBUG_LOCK_ALLOC
1568	if (!ret && lockres->l_lockdep_map.key != NULL) {
1569		if (level == DLM_LOCK_PR)
1570			rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass,
1571				!!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1572				caller_ip);
1573		else
1574			rwsem_acquire(&lockres->l_lockdep_map, l_subclass,
1575				!!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1576				caller_ip);
1577	}
1578#endif
1579	return ret;
1580}
1581
1582static inline int ocfs2_cluster_lock(struct ocfs2_super *osb,
1583				     struct ocfs2_lock_res *lockres,
1584				     int level,
1585				     u32 lkm_flags,
1586				     int arg_flags)
1587{
1588	return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags,
1589				    0, _RET_IP_);
1590}
1591
1592
1593static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
1594				   struct ocfs2_lock_res *lockres,
1595				   int level,
1596				   unsigned long caller_ip)
1597{
1598	unsigned long flags;
1599
1600	spin_lock_irqsave(&lockres->l_lock, flags);
1601	ocfs2_dec_holders(lockres, level);
1602	ocfs2_downconvert_on_unlock(osb, lockres);
1603	spin_unlock_irqrestore(&lockres->l_lock, flags);
1604#ifdef CONFIG_DEBUG_LOCK_ALLOC
1605	if (lockres->l_lockdep_map.key != NULL)
1606		rwsem_release(&lockres->l_lockdep_map, 1, caller_ip);
1607#endif
1608}
1609
1610static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1611				 struct ocfs2_lock_res *lockres,
1612				 int ex,
1613				 int local)
1614{
1615	int level =  ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1616	unsigned long flags;
1617	u32 lkm_flags = local ? DLM_LKF_LOCAL : 0;
1618
1619	spin_lock_irqsave(&lockres->l_lock, flags);
1620	BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1621	lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1622	spin_unlock_irqrestore(&lockres->l_lock, flags);
1623
1624	return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1625}
1626
1627/* Grants us an EX lock on the data and metadata resources, skipping
1628 * the normal cluster directory lookup. Use this ONLY on newly created
1629 * inodes which other nodes can't possibly see, and which haven't been
1630 * hashed in the inode hash yet. This can give us a good performance
1631 * increase as it'll skip the network broadcast normally associated
1632 * with creating a new lock resource. */
1633int ocfs2_create_new_inode_locks(struct inode *inode)
1634{
1635	int ret;
1636	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1637
1638	BUG_ON(!ocfs2_inode_is_new(inode));
1639
1640	mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1641
1642	/* NOTE: That we don't increment any of the holder counts, nor
1643	 * do we add anything to a journal handle. Since this is
1644	 * supposed to be a new inode which the cluster doesn't know
1645	 * about yet, there is no need to.  As far as the LVB handling
1646	 * is concerned, this is basically like acquiring an EX lock
1647	 * on a resource which has an invalid one -- we'll set it
1648	 * valid when we release the EX. */
1649
1650	ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1651	if (ret) {
1652		mlog_errno(ret);
1653		goto bail;
1654	}
1655
1656	/*
1657	 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1658	 * don't use a generation in their lock names.
1659	 */
1660	ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0);
1661	if (ret) {
1662		mlog_errno(ret);
1663		goto bail;
1664	}
1665
1666	ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1667	if (ret)
1668		mlog_errno(ret);
1669
1670bail:
1671	return ret;
1672}
1673
1674int ocfs2_rw_lock(struct inode *inode, int write)
1675{
1676	int status, level;
1677	struct ocfs2_lock_res *lockres;
1678	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1679
1680	mlog(0, "inode %llu take %s RW lock\n",
1681	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1682	     write ? "EXMODE" : "PRMODE");
1683
1684	if (ocfs2_mount_local(osb))
1685		return 0;
1686
1687	lockres = &OCFS2_I(inode)->ip_rw_lockres;
1688
1689	level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1690
1691	status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres, level, 0,
1692				    0);
1693	if (status < 0)
1694		mlog_errno(status);
1695
1696	return status;
1697}
1698
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1699void ocfs2_rw_unlock(struct inode *inode, int write)
1700{
1701	int level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1702	struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1703	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1704
1705	mlog(0, "inode %llu drop %s RW lock\n",
1706	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1707	     write ? "EXMODE" : "PRMODE");
1708
1709	if (!ocfs2_mount_local(osb))
1710		ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1711}
1712
1713/*
1714 * ocfs2_open_lock always get PR mode lock.
1715 */
1716int ocfs2_open_lock(struct inode *inode)
1717{
1718	int status = 0;
1719	struct ocfs2_lock_res *lockres;
1720	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1721
1722	mlog(0, "inode %llu take PRMODE open lock\n",
1723	     (unsigned long long)OCFS2_I(inode)->ip_blkno);
1724
1725	if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1726		goto out;
1727
1728	lockres = &OCFS2_I(inode)->ip_open_lockres;
1729
1730	status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1731				    DLM_LOCK_PR, 0, 0);
1732	if (status < 0)
1733		mlog_errno(status);
1734
1735out:
1736	return status;
1737}
1738
1739int ocfs2_try_open_lock(struct inode *inode, int write)
1740{
1741	int status = 0, level;
1742	struct ocfs2_lock_res *lockres;
1743	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1744
1745	mlog(0, "inode %llu try to take %s open lock\n",
1746	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1747	     write ? "EXMODE" : "PRMODE");
1748
1749	if (ocfs2_is_hard_readonly(osb)) {
1750		if (write)
1751			status = -EROFS;
1752		goto out;
1753	}
1754
1755	if (ocfs2_mount_local(osb))
1756		goto out;
1757
1758	lockres = &OCFS2_I(inode)->ip_open_lockres;
1759
1760	level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1761
1762	/*
1763	 * The file system may already holding a PRMODE/EXMODE open lock.
1764	 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1765	 * other nodes and the -EAGAIN will indicate to the caller that
1766	 * this inode is still in use.
1767	 */
1768	status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1769				    level, DLM_LKF_NOQUEUE, 0);
1770
1771out:
1772	return status;
1773}
1774
1775/*
1776 * ocfs2_open_unlock unlock PR and EX mode open locks.
1777 */
1778void ocfs2_open_unlock(struct inode *inode)
1779{
1780	struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1781	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1782
1783	mlog(0, "inode %llu drop open lock\n",
1784	     (unsigned long long)OCFS2_I(inode)->ip_blkno);
1785
1786	if (ocfs2_mount_local(osb))
1787		goto out;
1788
1789	if(lockres->l_ro_holders)
1790		ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1791				     DLM_LOCK_PR);
1792	if(lockres->l_ex_holders)
1793		ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1794				     DLM_LOCK_EX);
1795
1796out:
1797	return;
1798}
1799
1800static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres,
1801				     int level)
1802{
1803	int ret;
1804	struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1805	unsigned long flags;
1806	struct ocfs2_mask_waiter mw;
1807
1808	ocfs2_init_mask_waiter(&mw);
1809
1810retry_cancel:
1811	spin_lock_irqsave(&lockres->l_lock, flags);
1812	if (lockres->l_flags & OCFS2_LOCK_BUSY) {
1813		ret = ocfs2_prepare_cancel_convert(osb, lockres);
1814		if (ret) {
1815			spin_unlock_irqrestore(&lockres->l_lock, flags);
1816			ret = ocfs2_cancel_convert(osb, lockres);
1817			if (ret < 0) {
1818				mlog_errno(ret);
1819				goto out;
1820			}
1821			goto retry_cancel;
1822		}
1823		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1824		spin_unlock_irqrestore(&lockres->l_lock, flags);
1825
1826		ocfs2_wait_for_mask(&mw);
1827		goto retry_cancel;
1828	}
1829
1830	ret = -ERESTARTSYS;
1831	/*
1832	 * We may still have gotten the lock, in which case there's no
1833	 * point to restarting the syscall.
1834	 */
1835	if (lockres->l_level == level)
1836		ret = 0;
1837
1838	mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret,
1839	     lockres->l_flags, lockres->l_level, lockres->l_action);
1840
1841	spin_unlock_irqrestore(&lockres->l_lock, flags);
1842
1843out:
1844	return ret;
1845}
1846
1847/*
1848 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1849 * flock() calls. The locking approach this requires is sufficiently
1850 * different from all other cluster lock types that we implement a
1851 * separate path to the "low-level" dlm calls. In particular:
1852 *
1853 * - No optimization of lock levels is done - we take at exactly
1854 *   what's been requested.
1855 *
1856 * - No lock caching is employed. We immediately downconvert to
1857 *   no-lock at unlock time. This also means flock locks never go on
1858 *   the blocking list).
1859 *
1860 * - Since userspace can trivially deadlock itself with flock, we make
1861 *   sure to allow cancellation of a misbehaving applications flock()
1862 *   request.
1863 *
1864 * - Access to any flock lockres doesn't require concurrency, so we
1865 *   can simplify the code by requiring the caller to guarantee
1866 *   serialization of dlmglue flock calls.
1867 */
1868int ocfs2_file_lock(struct file *file, int ex, int trylock)
1869{
1870	int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1871	unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
1872	unsigned long flags;
1873	struct ocfs2_file_private *fp = file->private_data;
1874	struct ocfs2_lock_res *lockres = &fp->fp_flock;
1875	struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1876	struct ocfs2_mask_waiter mw;
1877
1878	ocfs2_init_mask_waiter(&mw);
1879
1880	if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
1881	    (lockres->l_level > DLM_LOCK_NL)) {
1882		mlog(ML_ERROR,
1883		     "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1884		     "level: %u\n", lockres->l_name, lockres->l_flags,
1885		     lockres->l_level);
1886		return -EINVAL;
1887	}
1888
1889	spin_lock_irqsave(&lockres->l_lock, flags);
1890	if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1891		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1892		spin_unlock_irqrestore(&lockres->l_lock, flags);
1893
1894		/*
1895		 * Get the lock at NLMODE to start - that way we
1896		 * can cancel the upconvert request if need be.
1897		 */
1898		ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
1899		if (ret < 0) {
1900			mlog_errno(ret);
1901			goto out;
1902		}
1903
1904		ret = ocfs2_wait_for_mask(&mw);
1905		if (ret) {
1906			mlog_errno(ret);
1907			goto out;
1908		}
1909		spin_lock_irqsave(&lockres->l_lock, flags);
1910	}
1911
1912	lockres->l_action = OCFS2_AST_CONVERT;
1913	lkm_flags |= DLM_LKF_CONVERT;
1914	lockres->l_requested = level;
1915	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1916
1917	lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1918	spin_unlock_irqrestore(&lockres->l_lock, flags);
1919
1920	ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
1921			     lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1);
1922	if (ret) {
1923		if (!trylock || (ret != -EAGAIN)) {
1924			ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1925			ret = -EINVAL;
1926		}
1927
1928		ocfs2_recover_from_dlm_error(lockres, 1);
1929		lockres_remove_mask_waiter(lockres, &mw);
1930		goto out;
1931	}
1932
1933	ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
1934	if (ret == -ERESTARTSYS) {
1935		/*
1936		 * Userspace can cause deadlock itself with
1937		 * flock(). Current behavior locally is to allow the
1938		 * deadlock, but abort the system call if a signal is
1939		 * received. We follow this example, otherwise a
1940		 * poorly written program could sit in kernel until
1941		 * reboot.
1942		 *
1943		 * Handling this is a bit more complicated for Ocfs2
1944		 * though. We can't exit this function with an
1945		 * outstanding lock request, so a cancel convert is
1946		 * required. We intentionally overwrite 'ret' - if the
1947		 * cancel fails and the lock was granted, it's easier
1948		 * to just bubble success back up to the user.
1949		 */
1950		ret = ocfs2_flock_handle_signal(lockres, level);
1951	} else if (!ret && (level > lockres->l_level)) {
1952		/* Trylock failed asynchronously */
1953		BUG_ON(!trylock);
1954		ret = -EAGAIN;
1955	}
1956
1957out:
1958
1959	mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
1960	     lockres->l_name, ex, trylock, ret);
1961	return ret;
1962}
1963
1964void ocfs2_file_unlock(struct file *file)
1965{
1966	int ret;
1967	unsigned int gen;
1968	unsigned long flags;
1969	struct ocfs2_file_private *fp = file->private_data;
1970	struct ocfs2_lock_res *lockres = &fp->fp_flock;
1971	struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1972	struct ocfs2_mask_waiter mw;
1973
1974	ocfs2_init_mask_waiter(&mw);
1975
1976	if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
1977		return;
1978
1979	if (lockres->l_level == DLM_LOCK_NL)
1980		return;
1981
1982	mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
1983	     lockres->l_name, lockres->l_flags, lockres->l_level,
1984	     lockres->l_action);
1985
1986	spin_lock_irqsave(&lockres->l_lock, flags);
1987	/*
1988	 * Fake a blocking ast for the downconvert code.
1989	 */
1990	lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
1991	lockres->l_blocking = DLM_LOCK_EX;
1992
1993	gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
1994	lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1995	spin_unlock_irqrestore(&lockres->l_lock, flags);
1996
1997	ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
1998	if (ret) {
1999		mlog_errno(ret);
2000		return;
2001	}
2002
2003	ret = ocfs2_wait_for_mask(&mw);
2004	if (ret)
2005		mlog_errno(ret);
2006}
2007
2008static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
2009					struct ocfs2_lock_res *lockres)
2010{
2011	int kick = 0;
2012
2013	/* If we know that another node is waiting on our lock, kick
2014	 * the downconvert thread * pre-emptively when we reach a release
2015	 * condition. */
2016	if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
2017		switch(lockres->l_blocking) {
2018		case DLM_LOCK_EX:
2019			if (!lockres->l_ex_holders && !lockres->l_ro_holders)
2020				kick = 1;
2021			break;
2022		case DLM_LOCK_PR:
2023			if (!lockres->l_ex_holders)
2024				kick = 1;
2025			break;
2026		default:
2027			BUG();
2028		}
2029	}
2030
2031	if (kick)
2032		ocfs2_wake_downconvert_thread(osb);
2033}
2034
2035#define OCFS2_SEC_BITS   34
2036#define OCFS2_SEC_SHIFT  (64 - 34)
2037#define OCFS2_NSEC_MASK  ((1ULL << OCFS2_SEC_SHIFT) - 1)
2038
2039/* LVB only has room for 64 bits of time here so we pack it for
2040 * now. */
2041static u64 ocfs2_pack_timespec(struct timespec *spec)
2042{
2043	u64 res;
2044	u64 sec = spec->tv_sec;
2045	u32 nsec = spec->tv_nsec;
2046
2047	res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
2048
2049	return res;
2050}
2051
2052/* Call this with the lockres locked. I am reasonably sure we don't
2053 * need ip_lock in this function as anyone who would be changing those
2054 * values is supposed to be blocked in ocfs2_inode_lock right now. */
2055static void __ocfs2_stuff_meta_lvb(struct inode *inode)
2056{
2057	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2058	struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2059	struct ocfs2_meta_lvb *lvb;
 
2060
2061	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2062
2063	/*
2064	 * Invalidate the LVB of a deleted inode - this way other
2065	 * nodes are forced to go to disk and discover the new inode
2066	 * status.
2067	 */
2068	if (oi->ip_flags & OCFS2_INODE_DELETED) {
2069		lvb->lvb_version = 0;
2070		goto out;
2071	}
2072
2073	lvb->lvb_version   = OCFS2_LVB_VERSION;
2074	lvb->lvb_isize	   = cpu_to_be64(i_size_read(inode));
2075	lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
2076	lvb->lvb_iuid      = cpu_to_be32(i_uid_read(inode));
2077	lvb->lvb_igid      = cpu_to_be32(i_gid_read(inode));
2078	lvb->lvb_imode     = cpu_to_be16(inode->i_mode);
2079	lvb->lvb_inlink    = cpu_to_be16(inode->i_nlink);
2080	lvb->lvb_iatime_packed  =
2081		cpu_to_be64(ocfs2_pack_timespec(&inode->i_atime));
2082	lvb->lvb_ictime_packed =
2083		cpu_to_be64(ocfs2_pack_timespec(&inode->i_ctime));
2084	lvb->lvb_imtime_packed =
2085		cpu_to_be64(ocfs2_pack_timespec(&inode->i_mtime));
2086	lvb->lvb_iattr    = cpu_to_be32(oi->ip_attr);
2087	lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features);
2088	lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
2089
2090out:
2091	mlog_meta_lvb(0, lockres);
2092}
2093
2094static void ocfs2_unpack_timespec(struct timespec *spec,
2095				  u64 packed_time)
2096{
2097	spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
2098	spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
2099}
2100
2101static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
2102{
2103	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2104	struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2105	struct ocfs2_meta_lvb *lvb;
 
2106
2107	mlog_meta_lvb(0, lockres);
2108
2109	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
 
 
2110
2111	/* We're safe here without the lockres lock... */
2112	spin_lock(&oi->ip_lock);
2113	oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
2114	i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
2115
2116	oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
2117	oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures);
2118	ocfs2_set_inode_flags(inode);
2119
2120	/* fast-symlinks are a special case */
2121	if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
2122		inode->i_blocks = 0;
2123	else
2124		inode->i_blocks = ocfs2_inode_sector_count(inode);
2125
2126	i_uid_write(inode, be32_to_cpu(lvb->lvb_iuid));
2127	i_gid_write(inode, be32_to_cpu(lvb->lvb_igid));
2128	inode->i_mode    = be16_to_cpu(lvb->lvb_imode);
2129	set_nlink(inode, be16_to_cpu(lvb->lvb_inlink));
2130	ocfs2_unpack_timespec(&inode->i_atime,
2131			      be64_to_cpu(lvb->lvb_iatime_packed));
2132	ocfs2_unpack_timespec(&inode->i_mtime,
2133			      be64_to_cpu(lvb->lvb_imtime_packed));
2134	ocfs2_unpack_timespec(&inode->i_ctime,
2135			      be64_to_cpu(lvb->lvb_ictime_packed));
2136	spin_unlock(&oi->ip_lock);
 
2137}
2138
2139static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
2140					      struct ocfs2_lock_res *lockres)
2141{
2142	struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2143
2144	if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)
2145	    && lvb->lvb_version == OCFS2_LVB_VERSION
2146	    && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
2147		return 1;
2148	return 0;
2149}
2150
2151/* Determine whether a lock resource needs to be refreshed, and
2152 * arbitrate who gets to refresh it.
2153 *
2154 *   0 means no refresh needed.
2155 *
2156 *   > 0 means you need to refresh this and you MUST call
2157 *   ocfs2_complete_lock_res_refresh afterwards. */
2158static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
2159{
2160	unsigned long flags;
2161	int status = 0;
2162
2163refresh_check:
2164	spin_lock_irqsave(&lockres->l_lock, flags);
2165	if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
2166		spin_unlock_irqrestore(&lockres->l_lock, flags);
2167		goto bail;
2168	}
2169
2170	if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
2171		spin_unlock_irqrestore(&lockres->l_lock, flags);
2172
2173		ocfs2_wait_on_refreshing_lock(lockres);
2174		goto refresh_check;
2175	}
2176
2177	/* Ok, I'll be the one to refresh this lock. */
2178	lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
2179	spin_unlock_irqrestore(&lockres->l_lock, flags);
2180
2181	status = 1;
2182bail:
2183	mlog(0, "status %d\n", status);
2184	return status;
2185}
2186
2187/* If status is non zero, I'll mark it as not being in refresh
2188 * anymroe, but i won't clear the needs refresh flag. */
2189static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
2190						   int status)
2191{
2192	unsigned long flags;
2193
2194	spin_lock_irqsave(&lockres->l_lock, flags);
2195	lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
2196	if (!status)
2197		lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
2198	spin_unlock_irqrestore(&lockres->l_lock, flags);
2199
2200	wake_up(&lockres->l_event);
2201}
2202
2203/* may or may not return a bh if it went to disk. */
2204static int ocfs2_inode_lock_update(struct inode *inode,
2205				  struct buffer_head **bh)
2206{
2207	int status = 0;
2208	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2209	struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2210	struct ocfs2_dinode *fe;
2211	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2212
2213	if (ocfs2_mount_local(osb))
2214		goto bail;
2215
2216	spin_lock(&oi->ip_lock);
2217	if (oi->ip_flags & OCFS2_INODE_DELETED) {
2218		mlog(0, "Orphaned inode %llu was deleted while we "
2219		     "were waiting on a lock. ip_flags = 0x%x\n",
2220		     (unsigned long long)oi->ip_blkno, oi->ip_flags);
2221		spin_unlock(&oi->ip_lock);
2222		status = -ENOENT;
2223		goto bail;
2224	}
2225	spin_unlock(&oi->ip_lock);
2226
2227	if (!ocfs2_should_refresh_lock_res(lockres))
2228		goto bail;
2229
2230	/* This will discard any caching information we might have had
2231	 * for the inode metadata. */
2232	ocfs2_metadata_cache_purge(INODE_CACHE(inode));
2233
2234	ocfs2_extent_map_trunc(inode, 0);
2235
2236	if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
2237		mlog(0, "Trusting LVB on inode %llu\n",
2238		     (unsigned long long)oi->ip_blkno);
2239		ocfs2_refresh_inode_from_lvb(inode);
 
2240	} else {
2241		/* Boo, we have to go to disk. */
2242		/* read bh, cast, ocfs2_refresh_inode */
2243		status = ocfs2_read_inode_block(inode, bh);
2244		if (status < 0) {
2245			mlog_errno(status);
2246			goto bail_refresh;
2247		}
2248		fe = (struct ocfs2_dinode *) (*bh)->b_data;
 
 
 
 
2249
2250		/* This is a good chance to make sure we're not
2251		 * locking an invalid object.  ocfs2_read_inode_block()
2252		 * already checked that the inode block is sane.
2253		 *
2254		 * We bug on a stale inode here because we checked
2255		 * above whether it was wiped from disk. The wiping
2256		 * node provides a guarantee that we receive that
2257		 * message and can mark the inode before dropping any
2258		 * locks associated with it. */
2259		mlog_bug_on_msg(inode->i_generation !=
2260				le32_to_cpu(fe->i_generation),
2261				"Invalid dinode %llu disk generation: %u "
2262				"inode->i_generation: %u\n",
2263				(unsigned long long)oi->ip_blkno,
2264				le32_to_cpu(fe->i_generation),
2265				inode->i_generation);
2266		mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
2267				!(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
2268				"Stale dinode %llu dtime: %llu flags: 0x%x\n",
2269				(unsigned long long)oi->ip_blkno,
2270				(unsigned long long)le64_to_cpu(fe->i_dtime),
2271				le32_to_cpu(fe->i_flags));
2272
2273		ocfs2_refresh_inode(inode, fe);
2274		ocfs2_track_lock_refresh(lockres);
2275	}
2276
2277	status = 0;
2278bail_refresh:
2279	ocfs2_complete_lock_res_refresh(lockres, status);
2280bail:
2281	return status;
2282}
2283
2284static int ocfs2_assign_bh(struct inode *inode,
2285			   struct buffer_head **ret_bh,
2286			   struct buffer_head *passed_bh)
2287{
2288	int status;
2289
2290	if (passed_bh) {
2291		/* Ok, the update went to disk for us, use the
2292		 * returned bh. */
2293		*ret_bh = passed_bh;
2294		get_bh(*ret_bh);
2295
2296		return 0;
2297	}
2298
2299	status = ocfs2_read_inode_block(inode, ret_bh);
2300	if (status < 0)
2301		mlog_errno(status);
2302
2303	return status;
2304}
2305
2306/*
2307 * returns < 0 error if the callback will never be called, otherwise
2308 * the result of the lock will be communicated via the callback.
2309 */
2310int ocfs2_inode_lock_full_nested(struct inode *inode,
2311				 struct buffer_head **ret_bh,
2312				 int ex,
2313				 int arg_flags,
2314				 int subclass)
2315{
2316	int status, level, acquired;
2317	u32 dlm_flags;
2318	struct ocfs2_lock_res *lockres = NULL;
2319	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2320	struct buffer_head *local_bh = NULL;
2321
2322	mlog(0, "inode %llu, take %s META lock\n",
2323	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
2324	     ex ? "EXMODE" : "PRMODE");
2325
2326	status = 0;
2327	acquired = 0;
2328	/* We'll allow faking a readonly metadata lock for
2329	 * rodevices. */
2330	if (ocfs2_is_hard_readonly(osb)) {
2331		if (ex)
2332			status = -EROFS;
2333		goto getbh;
2334	}
2335
2336	if (ocfs2_mount_local(osb))
2337		goto local;
 
2338
2339	if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2340		ocfs2_wait_for_recovery(osb);
2341
2342	lockres = &OCFS2_I(inode)->ip_inode_lockres;
2343	level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2344	dlm_flags = 0;
2345	if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
2346		dlm_flags |= DLM_LKF_NOQUEUE;
2347
2348	status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
2349				      arg_flags, subclass, _RET_IP_);
2350	if (status < 0) {
2351		if (status != -EAGAIN)
2352			mlog_errno(status);
2353		goto bail;
2354	}
2355
2356	/* Notify the error cleanup path to drop the cluster lock. */
2357	acquired = 1;
2358
2359	/* We wait twice because a node may have died while we were in
2360	 * the lower dlm layers. The second time though, we've
2361	 * committed to owning this lock so we don't allow signals to
2362	 * abort the operation. */
2363	if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2364		ocfs2_wait_for_recovery(osb);
2365
2366local:
2367	/*
2368	 * We only see this flag if we're being called from
2369	 * ocfs2_read_locked_inode(). It means we're locking an inode
2370	 * which hasn't been populated yet, so clear the refresh flag
2371	 * and let the caller handle it.
2372	 */
2373	if (inode->i_state & I_NEW) {
2374		status = 0;
2375		if (lockres)
2376			ocfs2_complete_lock_res_refresh(lockres, 0);
2377		goto bail;
2378	}
2379
2380	/* This is fun. The caller may want a bh back, or it may
2381	 * not. ocfs2_inode_lock_update definitely wants one in, but
2382	 * may or may not read one, depending on what's in the
2383	 * LVB. The result of all of this is that we've *only* gone to
2384	 * disk if we have to, so the complexity is worthwhile. */
2385	status = ocfs2_inode_lock_update(inode, &local_bh);
2386	if (status < 0) {
2387		if (status != -ENOENT)
2388			mlog_errno(status);
2389		goto bail;
2390	}
2391getbh:
2392	if (ret_bh) {
2393		status = ocfs2_assign_bh(inode, ret_bh, local_bh);
2394		if (status < 0) {
2395			mlog_errno(status);
2396			goto bail;
2397		}
2398	}
2399
2400bail:
2401	if (status < 0) {
2402		if (ret_bh && (*ret_bh)) {
2403			brelse(*ret_bh);
2404			*ret_bh = NULL;
2405		}
2406		if (acquired)
2407			ocfs2_inode_unlock(inode, ex);
2408	}
2409
2410	if (local_bh)
2411		brelse(local_bh);
2412
2413	return status;
2414}
2415
2416/*
2417 * This is working around a lock inversion between tasks acquiring DLM
2418 * locks while holding a page lock and the downconvert thread which
2419 * blocks dlm lock acquiry while acquiring page locks.
2420 *
2421 * ** These _with_page variantes are only intended to be called from aop
2422 * methods that hold page locks and return a very specific *positive* error
2423 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2424 *
2425 * The DLM is called such that it returns -EAGAIN if it would have
2426 * blocked waiting for the downconvert thread.  In that case we unlock
2427 * our page so the downconvert thread can make progress.  Once we've
2428 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2429 * that called us can bubble that back up into the VFS who will then
2430 * immediately retry the aop call.
2431 */
2432int ocfs2_inode_lock_with_page(struct inode *inode,
2433			      struct buffer_head **ret_bh,
2434			      int ex,
2435			      struct page *page)
2436{
2437	int ret;
2438
2439	ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
2440	if (ret == -EAGAIN) {
2441		unlock_page(page);
 
 
 
 
 
 
 
 
 
2442		ret = AOP_TRUNCATED_PAGE;
2443	}
2444
2445	return ret;
2446}
2447
2448int ocfs2_inode_lock_atime(struct inode *inode,
2449			  struct vfsmount *vfsmnt,
2450			  int *level)
2451{
2452	int ret;
2453
2454	ret = ocfs2_inode_lock(inode, NULL, 0);
 
 
 
 
2455	if (ret < 0) {
2456		mlog_errno(ret);
 
2457		return ret;
2458	}
2459
2460	/*
2461	 * If we should update atime, we will get EX lock,
2462	 * otherwise we just get PR lock.
2463	 */
2464	if (ocfs2_should_update_atime(inode, vfsmnt)) {
2465		struct buffer_head *bh = NULL;
2466
2467		ocfs2_inode_unlock(inode, 0);
2468		ret = ocfs2_inode_lock(inode, &bh, 1);
 
 
 
 
2469		if (ret < 0) {
2470			mlog_errno(ret);
 
2471			return ret;
2472		}
2473		*level = 1;
2474		if (ocfs2_should_update_atime(inode, vfsmnt))
2475			ocfs2_update_inode_atime(inode, bh);
2476		if (bh)
2477			brelse(bh);
2478	} else
2479		*level = 0;
2480
2481	return ret;
2482}
2483
2484void ocfs2_inode_unlock(struct inode *inode,
2485		       int ex)
2486{
2487	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2488	struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres;
2489	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2490
2491	mlog(0, "inode %llu drop %s META lock\n",
2492	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
2493	     ex ? "EXMODE" : "PRMODE");
2494
2495	if (!ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb)) &&
2496	    !ocfs2_mount_local(osb))
2497		ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2498}
2499
2500int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno)
2501{
2502	struct ocfs2_lock_res *lockres;
2503	struct ocfs2_orphan_scan_lvb *lvb;
2504	int status = 0;
2505
2506	if (ocfs2_is_hard_readonly(osb))
2507		return -EROFS;
2508
2509	if (ocfs2_mount_local(osb))
2510		return 0;
2511
2512	lockres = &osb->osb_orphan_scan.os_lockres;
2513	status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2514	if (status < 0)
2515		return status;
2516
2517	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2518	if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2519	    lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION)
2520		*seqno = be32_to_cpu(lvb->lvb_os_seqno);
2521	else
2522		*seqno = osb->osb_orphan_scan.os_seqno + 1;
2523
2524	return status;
2525}
2526
2527void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno)
2528{
2529	struct ocfs2_lock_res *lockres;
2530	struct ocfs2_orphan_scan_lvb *lvb;
2531
2532	if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) {
2533		lockres = &osb->osb_orphan_scan.os_lockres;
2534		lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2535		lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION;
2536		lvb->lvb_os_seqno = cpu_to_be32(seqno);
2537		ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2538	}
2539}
2540
2541int ocfs2_super_lock(struct ocfs2_super *osb,
2542		     int ex)
2543{
2544	int status = 0;
2545	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2546	struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2547
2548	if (ocfs2_is_hard_readonly(osb))
2549		return -EROFS;
2550
2551	if (ocfs2_mount_local(osb))
2552		goto bail;
2553
2554	status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
2555	if (status < 0) {
2556		mlog_errno(status);
2557		goto bail;
2558	}
2559
2560	/* The super block lock path is really in the best position to
2561	 * know when resources covered by the lock need to be
2562	 * refreshed, so we do it here. Of course, making sense of
2563	 * everything is up to the caller :) */
2564	status = ocfs2_should_refresh_lock_res(lockres);
2565	if (status) {
2566		status = ocfs2_refresh_slot_info(osb);
2567
2568		ocfs2_complete_lock_res_refresh(lockres, status);
2569
2570		if (status < 0) {
2571			ocfs2_cluster_unlock(osb, lockres, level);
2572			mlog_errno(status);
2573		}
2574		ocfs2_track_lock_refresh(lockres);
2575	}
2576bail:
2577	return status;
2578}
2579
2580void ocfs2_super_unlock(struct ocfs2_super *osb,
2581			int ex)
2582{
2583	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2584	struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2585
2586	if (!ocfs2_mount_local(osb))
2587		ocfs2_cluster_unlock(osb, lockres, level);
2588}
2589
2590int ocfs2_rename_lock(struct ocfs2_super *osb)
2591{
2592	int status;
2593	struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2594
2595	if (ocfs2_is_hard_readonly(osb))
2596		return -EROFS;
2597
2598	if (ocfs2_mount_local(osb))
2599		return 0;
2600
2601	status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2602	if (status < 0)
2603		mlog_errno(status);
2604
2605	return status;
2606}
2607
2608void ocfs2_rename_unlock(struct ocfs2_super *osb)
2609{
2610	struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2611
2612	if (!ocfs2_mount_local(osb))
2613		ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2614}
2615
2616int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex)
2617{
2618	int status;
2619	struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2620
2621	if (ocfs2_is_hard_readonly(osb))
2622		return -EROFS;
2623
 
 
 
 
 
2624	if (ocfs2_mount_local(osb))
2625		return 0;
2626
2627	status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE,
2628				    0, 0);
2629	if (status < 0)
2630		mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status);
2631
 
 
 
 
 
 
2632	return status;
2633}
2634
2635void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex)
2636{
2637	struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2638
2639	if (!ocfs2_mount_local(osb))
2640		ocfs2_cluster_unlock(osb, lockres,
2641				     ex ? LKM_EXMODE : LKM_PRMODE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2642}
2643
2644int ocfs2_dentry_lock(struct dentry *dentry, int ex)
2645{
2646	int ret;
2647	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2648	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2649	struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2650
2651	BUG_ON(!dl);
2652
2653	if (ocfs2_is_hard_readonly(osb)) {
2654		if (ex)
2655			return -EROFS;
2656		return 0;
2657	}
2658
2659	if (ocfs2_mount_local(osb))
2660		return 0;
2661
2662	ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2663	if (ret < 0)
2664		mlog_errno(ret);
2665
2666	return ret;
2667}
2668
2669void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2670{
2671	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2672	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2673	struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2674
2675	if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
2676		ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
2677}
2678
2679/* Reference counting of the dlm debug structure. We want this because
2680 * open references on the debug inodes can live on after a mount, so
2681 * we can't rely on the ocfs2_super to always exist. */
2682static void ocfs2_dlm_debug_free(struct kref *kref)
2683{
2684	struct ocfs2_dlm_debug *dlm_debug;
2685
2686	dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
2687
2688	kfree(dlm_debug);
2689}
2690
2691void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
2692{
2693	if (dlm_debug)
2694		kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
2695}
2696
2697static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
2698{
2699	kref_get(&debug->d_refcnt);
2700}
2701
2702struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
2703{
2704	struct ocfs2_dlm_debug *dlm_debug;
2705
2706	dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
2707	if (!dlm_debug) {
2708		mlog_errno(-ENOMEM);
2709		goto out;
2710	}
2711
2712	kref_init(&dlm_debug->d_refcnt);
2713	INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
2714	dlm_debug->d_locking_state = NULL;
2715out:
2716	return dlm_debug;
2717}
2718
2719/* Access to this is arbitrated for us via seq_file->sem. */
2720struct ocfs2_dlm_seq_priv {
2721	struct ocfs2_dlm_debug *p_dlm_debug;
2722	struct ocfs2_lock_res p_iter_res;
2723	struct ocfs2_lock_res p_tmp_res;
2724};
2725
2726static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
2727						 struct ocfs2_dlm_seq_priv *priv)
2728{
2729	struct ocfs2_lock_res *iter, *ret = NULL;
2730	struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
2731
2732	assert_spin_locked(&ocfs2_dlm_tracking_lock);
2733
2734	list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
2735		/* discover the head of the list */
2736		if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
2737			mlog(0, "End of list found, %p\n", ret);
2738			break;
2739		}
2740
2741		/* We track our "dummy" iteration lockres' by a NULL
2742		 * l_ops field. */
2743		if (iter->l_ops != NULL) {
2744			ret = iter;
2745			break;
2746		}
2747	}
2748
2749	return ret;
2750}
2751
2752static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
2753{
2754	struct ocfs2_dlm_seq_priv *priv = m->private;
2755	struct ocfs2_lock_res *iter;
2756
2757	spin_lock(&ocfs2_dlm_tracking_lock);
2758	iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
2759	if (iter) {
2760		/* Since lockres' have the lifetime of their container
2761		 * (which can be inodes, ocfs2_supers, etc) we want to
2762		 * copy this out to a temporary lockres while still
2763		 * under the spinlock. Obviously after this we can't
2764		 * trust any pointers on the copy returned, but that's
2765		 * ok as the information we want isn't typically held
2766		 * in them. */
2767		priv->p_tmp_res = *iter;
2768		iter = &priv->p_tmp_res;
2769	}
2770	spin_unlock(&ocfs2_dlm_tracking_lock);
2771
2772	return iter;
2773}
2774
2775static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
2776{
2777}
2778
2779static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
2780{
2781	struct ocfs2_dlm_seq_priv *priv = m->private;
2782	struct ocfs2_lock_res *iter = v;
2783	struct ocfs2_lock_res *dummy = &priv->p_iter_res;
2784
 
2785	spin_lock(&ocfs2_dlm_tracking_lock);
2786	iter = ocfs2_dlm_next_res(iter, priv);
2787	list_del_init(&dummy->l_debug_list);
2788	if (iter) {
2789		list_add(&dummy->l_debug_list, &iter->l_debug_list);
2790		priv->p_tmp_res = *iter;
2791		iter = &priv->p_tmp_res;
2792	}
2793	spin_unlock(&ocfs2_dlm_tracking_lock);
2794
2795	return iter;
2796}
2797
2798/*
2799 * Version is used by debugfs.ocfs2 to determine the format being used
2800 *
2801 * New in version 2
2802 *	- Lock stats printed
2803 * New in version 3
2804 *	- Max time in lock stats is in usecs (instead of nsecs)
 
 
2805 */
2806#define OCFS2_DLM_DEBUG_STR_VERSION 3
2807static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
2808{
2809	int i;
2810	char *lvb;
2811	struct ocfs2_lock_res *lockres = v;
 
 
 
 
 
2812
2813	if (!lockres)
2814		return -EINVAL;
2815
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2816	seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
2817
2818	if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
2819		seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
2820			   lockres->l_name,
2821			   (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
2822	else
2823		seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
2824
2825	seq_printf(m, "%d\t"
2826		   "0x%lx\t"
2827		   "0x%x\t"
2828		   "0x%x\t"
2829		   "%u\t"
2830		   "%u\t"
2831		   "%d\t"
2832		   "%d\t",
2833		   lockres->l_level,
2834		   lockres->l_flags,
2835		   lockres->l_action,
2836		   lockres->l_unlock_action,
2837		   lockres->l_ro_holders,
2838		   lockres->l_ex_holders,
2839		   lockres->l_requested,
2840		   lockres->l_blocking);
2841
2842	/* Dump the raw LVB */
2843	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2844	for(i = 0; i < DLM_LVB_LEN; i++)
2845		seq_printf(m, "0x%x\t", lvb[i]);
2846
2847#ifdef CONFIG_OCFS2_FS_STATS
2848# define lock_num_prmode(_l)		((_l)->l_lock_prmode.ls_gets)
2849# define lock_num_exmode(_l)		((_l)->l_lock_exmode.ls_gets)
2850# define lock_num_prmode_failed(_l)	((_l)->l_lock_prmode.ls_fail)
2851# define lock_num_exmode_failed(_l)	((_l)->l_lock_exmode.ls_fail)
2852# define lock_total_prmode(_l)		((_l)->l_lock_prmode.ls_total)
2853# define lock_total_exmode(_l)		((_l)->l_lock_exmode.ls_total)
2854# define lock_max_prmode(_l)		((_l)->l_lock_prmode.ls_max)
2855# define lock_max_exmode(_l)		((_l)->l_lock_exmode.ls_max)
2856# define lock_refresh(_l)		((_l)->l_lock_refresh)
 
 
 
2857#else
2858# define lock_num_prmode(_l)		(0)
2859# define lock_num_exmode(_l)		(0)
2860# define lock_num_prmode_failed(_l)	(0)
2861# define lock_num_exmode_failed(_l)	(0)
2862# define lock_total_prmode(_l)		(0ULL)
2863# define lock_total_exmode(_l)		(0ULL)
2864# define lock_max_prmode(_l)		(0)
2865# define lock_max_exmode(_l)		(0)
2866# define lock_refresh(_l)		(0)
 
 
 
2867#endif
2868	/* The following seq_print was added in version 2 of this output */
2869	seq_printf(m, "%u\t"
2870		   "%u\t"
2871		   "%u\t"
2872		   "%u\t"
2873		   "%llu\t"
2874		   "%llu\t"
2875		   "%u\t"
2876		   "%u\t"
2877		   "%u\t",
 
 
 
2878		   lock_num_prmode(lockres),
2879		   lock_num_exmode(lockres),
2880		   lock_num_prmode_failed(lockres),
2881		   lock_num_exmode_failed(lockres),
2882		   lock_total_prmode(lockres),
2883		   lock_total_exmode(lockres),
2884		   lock_max_prmode(lockres),
2885		   lock_max_exmode(lockres),
2886		   lock_refresh(lockres));
 
 
 
2887
2888	/* End the line */
2889	seq_printf(m, "\n");
2890	return 0;
2891}
2892
2893static const struct seq_operations ocfs2_dlm_seq_ops = {
2894	.start =	ocfs2_dlm_seq_start,
2895	.stop =		ocfs2_dlm_seq_stop,
2896	.next =		ocfs2_dlm_seq_next,
2897	.show =		ocfs2_dlm_seq_show,
2898};
2899
2900static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
2901{
2902	struct seq_file *seq = file->private_data;
2903	struct ocfs2_dlm_seq_priv *priv = seq->private;
2904	struct ocfs2_lock_res *res = &priv->p_iter_res;
2905
2906	ocfs2_remove_lockres_tracking(res);
2907	ocfs2_put_dlm_debug(priv->p_dlm_debug);
2908	return seq_release_private(inode, file);
2909}
2910
2911static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
2912{
2913	struct ocfs2_dlm_seq_priv *priv;
2914	struct ocfs2_super *osb;
2915
2916	priv = __seq_open_private(file, &ocfs2_dlm_seq_ops, sizeof(*priv));
2917	if (!priv) {
2918		mlog_errno(-ENOMEM);
2919		return -ENOMEM;
2920	}
2921
2922	osb = inode->i_private;
2923	ocfs2_get_dlm_debug(osb->osb_dlm_debug);
2924	priv->p_dlm_debug = osb->osb_dlm_debug;
2925	INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
2926
2927	ocfs2_add_lockres_tracking(&priv->p_iter_res,
2928				   priv->p_dlm_debug);
2929
2930	return 0;
2931}
2932
2933static const struct file_operations ocfs2_dlm_debug_fops = {
2934	.open =		ocfs2_dlm_debug_open,
2935	.release =	ocfs2_dlm_debug_release,
2936	.read =		seq_read,
2937	.llseek =	seq_lseek,
2938};
2939
2940static int ocfs2_dlm_init_debug(struct ocfs2_super *osb)
2941{
2942	int ret = 0;
2943	struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2944
2945	dlm_debug->d_locking_state = debugfs_create_file("locking_state",
2946							 S_IFREG|S_IRUSR,
2947							 osb->osb_debug_root,
2948							 osb,
2949							 &ocfs2_dlm_debug_fops);
2950	if (!dlm_debug->d_locking_state) {
2951		ret = -EINVAL;
2952		mlog(ML_ERROR,
2953		     "Unable to create locking state debugfs file.\n");
2954		goto out;
2955	}
2956
 
 
2957	ocfs2_get_dlm_debug(dlm_debug);
2958out:
2959	return ret;
2960}
2961
2962static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
2963{
2964	struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2965
2966	if (dlm_debug) {
2967		debugfs_remove(dlm_debug->d_locking_state);
2968		ocfs2_put_dlm_debug(dlm_debug);
2969	}
2970}
2971
2972int ocfs2_dlm_init(struct ocfs2_super *osb)
2973{
2974	int status = 0;
2975	struct ocfs2_cluster_connection *conn = NULL;
2976
2977	if (ocfs2_mount_local(osb)) {
2978		osb->node_num = 0;
2979		goto local;
2980	}
2981
2982	status = ocfs2_dlm_init_debug(osb);
2983	if (status < 0) {
2984		mlog_errno(status);
2985		goto bail;
2986	}
2987
2988	/* launch downconvert thread */
2989	osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc-%s",
2990			osb->uuid_str);
2991	if (IS_ERR(osb->dc_task)) {
2992		status = PTR_ERR(osb->dc_task);
2993		osb->dc_task = NULL;
2994		mlog_errno(status);
2995		goto bail;
2996	}
2997
2998	/* for now, uuid == domain */
2999	status = ocfs2_cluster_connect(osb->osb_cluster_stack,
3000				       osb->osb_cluster_name,
3001				       strlen(osb->osb_cluster_name),
3002				       osb->uuid_str,
3003				       strlen(osb->uuid_str),
3004				       &lproto, ocfs2_do_node_down, osb,
3005				       &conn);
3006	if (status) {
3007		mlog_errno(status);
3008		goto bail;
3009	}
3010
3011	status = ocfs2_cluster_this_node(conn, &osb->node_num);
3012	if (status < 0) {
3013		mlog_errno(status);
3014		mlog(ML_ERROR,
3015		     "could not find this host's node number\n");
3016		ocfs2_cluster_disconnect(conn, 0);
3017		goto bail;
3018	}
3019
3020local:
3021	ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
3022	ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
3023	ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb);
3024	ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb);
3025
3026	osb->cconn = conn;
3027bail:
3028	if (status < 0) {
3029		ocfs2_dlm_shutdown_debug(osb);
3030		if (osb->dc_task)
3031			kthread_stop(osb->dc_task);
3032	}
3033
3034	return status;
3035}
3036
3037void ocfs2_dlm_shutdown(struct ocfs2_super *osb,
3038			int hangup_pending)
3039{
3040	ocfs2_drop_osb_locks(osb);
3041
3042	/*
3043	 * Now that we have dropped all locks and ocfs2_dismount_volume()
3044	 * has disabled recovery, the DLM won't be talking to us.  It's
3045	 * safe to tear things down before disconnecting the cluster.
3046	 */
3047
3048	if (osb->dc_task) {
3049		kthread_stop(osb->dc_task);
3050		osb->dc_task = NULL;
3051	}
3052
3053	ocfs2_lock_res_free(&osb->osb_super_lockres);
3054	ocfs2_lock_res_free(&osb->osb_rename_lockres);
3055	ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres);
3056	ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres);
3057
3058	ocfs2_cluster_disconnect(osb->cconn, hangup_pending);
3059	osb->cconn = NULL;
 
3060
3061	ocfs2_dlm_shutdown_debug(osb);
 
3062}
3063
3064static int ocfs2_drop_lock(struct ocfs2_super *osb,
3065			   struct ocfs2_lock_res *lockres)
3066{
3067	int ret;
3068	unsigned long flags;
3069	u32 lkm_flags = 0;
3070
3071	/* We didn't get anywhere near actually using this lockres. */
3072	if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
3073		goto out;
3074
3075	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3076		lkm_flags |= DLM_LKF_VALBLK;
3077
3078	spin_lock_irqsave(&lockres->l_lock, flags);
3079
3080	mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
3081			"lockres %s, flags 0x%lx\n",
3082			lockres->l_name, lockres->l_flags);
3083
3084	while (lockres->l_flags & OCFS2_LOCK_BUSY) {
3085		mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3086		     "%u, unlock_action = %u\n",
3087		     lockres->l_name, lockres->l_flags, lockres->l_action,
3088		     lockres->l_unlock_action);
3089
3090		spin_unlock_irqrestore(&lockres->l_lock, flags);
3091
3092		/* XXX: Today we just wait on any busy
3093		 * locks... Perhaps we need to cancel converts in the
3094		 * future? */
3095		ocfs2_wait_on_busy_lock(lockres);
3096
3097		spin_lock_irqsave(&lockres->l_lock, flags);
3098	}
3099
3100	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3101		if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
3102		    lockres->l_level == DLM_LOCK_EX &&
3103		    !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3104			lockres->l_ops->set_lvb(lockres);
3105	}
3106
3107	if (lockres->l_flags & OCFS2_LOCK_BUSY)
3108		mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
3109		     lockres->l_name);
3110	if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3111		mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
3112
3113	if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
3114		spin_unlock_irqrestore(&lockres->l_lock, flags);
3115		goto out;
3116	}
3117
3118	lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
3119
3120	/* make sure we never get here while waiting for an ast to
3121	 * fire. */
3122	BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
3123
3124	/* is this necessary? */
3125	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3126	lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
3127	spin_unlock_irqrestore(&lockres->l_lock, flags);
3128
3129	mlog(0, "lock %s\n", lockres->l_name);
3130
3131	ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags);
3132	if (ret) {
3133		ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3134		mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
3135		ocfs2_dlm_dump_lksb(&lockres->l_lksb);
3136		BUG();
3137	}
3138	mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3139	     lockres->l_name);
3140
3141	ocfs2_wait_on_busy_lock(lockres);
3142out:
3143	return 0;
3144}
3145
3146static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3147				       struct ocfs2_lock_res *lockres);
3148
3149/* Mark the lockres as being dropped. It will no longer be
3150 * queued if blocking, but we still may have to wait on it
3151 * being dequeued from the downconvert thread before we can consider
3152 * it safe to drop.
3153 *
3154 * You can *not* attempt to call cluster_lock on this lockres anymore. */
3155void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb,
3156				struct ocfs2_lock_res *lockres)
3157{
3158	int status;
3159	struct ocfs2_mask_waiter mw;
3160	unsigned long flags, flags2;
3161
3162	ocfs2_init_mask_waiter(&mw);
3163
3164	spin_lock_irqsave(&lockres->l_lock, flags);
3165	lockres->l_flags |= OCFS2_LOCK_FREEING;
3166	if (lockres->l_flags & OCFS2_LOCK_QUEUED && current == osb->dc_task) {
3167		/*
3168		 * We know the downconvert is queued but not in progress
3169		 * because we are the downconvert thread and processing
3170		 * different lock. So we can just remove the lock from the
3171		 * queue. This is not only an optimization but also a way
3172		 * to avoid the following deadlock:
3173		 *   ocfs2_dentry_post_unlock()
3174		 *     ocfs2_dentry_lock_put()
3175		 *       ocfs2_drop_dentry_lock()
3176		 *         iput()
3177		 *           ocfs2_evict_inode()
3178		 *             ocfs2_clear_inode()
3179		 *               ocfs2_mark_lockres_freeing()
3180		 *                 ... blocks waiting for OCFS2_LOCK_QUEUED
3181		 *                 since we are the downconvert thread which
3182		 *                 should clear the flag.
3183		 */
3184		spin_unlock_irqrestore(&lockres->l_lock, flags);
3185		spin_lock_irqsave(&osb->dc_task_lock, flags2);
3186		list_del_init(&lockres->l_blocked_list);
3187		osb->blocked_lock_count--;
3188		spin_unlock_irqrestore(&osb->dc_task_lock, flags2);
3189		/*
3190		 * Warn if we recurse into another post_unlock call.  Strictly
3191		 * speaking it isn't a problem but we need to be careful if
3192		 * that happens (stack overflow, deadlocks, ...) so warn if
3193		 * ocfs2 grows a path for which this can happen.
3194		 */
3195		WARN_ON_ONCE(lockres->l_ops->post_unlock);
3196		/* Since the lock is freeing we don't do much in the fn below */
3197		ocfs2_process_blocked_lock(osb, lockres);
3198		return;
3199	}
3200	while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
3201		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
3202		spin_unlock_irqrestore(&lockres->l_lock, flags);
3203
3204		mlog(0, "Waiting on lockres %s\n", lockres->l_name);
3205
3206		status = ocfs2_wait_for_mask(&mw);
3207		if (status)
3208			mlog_errno(status);
3209
3210		spin_lock_irqsave(&lockres->l_lock, flags);
3211	}
3212	spin_unlock_irqrestore(&lockres->l_lock, flags);
3213}
3214
3215void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
3216			       struct ocfs2_lock_res *lockres)
3217{
3218	int ret;
3219
3220	ocfs2_mark_lockres_freeing(osb, lockres);
3221	ret = ocfs2_drop_lock(osb, lockres);
3222	if (ret)
3223		mlog_errno(ret);
3224}
3225
3226static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
3227{
3228	ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
3229	ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
3230	ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres);
3231	ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres);
3232}
3233
3234int ocfs2_drop_inode_locks(struct inode *inode)
3235{
3236	int status, err;
3237
3238	/* No need to call ocfs2_mark_lockres_freeing here -
3239	 * ocfs2_clear_inode has done it for us. */
3240
3241	err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3242			      &OCFS2_I(inode)->ip_open_lockres);
3243	if (err < 0)
3244		mlog_errno(err);
3245
3246	status = err;
3247
3248	err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3249			      &OCFS2_I(inode)->ip_inode_lockres);
3250	if (err < 0)
3251		mlog_errno(err);
3252	if (err < 0 && !status)
3253		status = err;
3254
3255	err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3256			      &OCFS2_I(inode)->ip_rw_lockres);
3257	if (err < 0)
3258		mlog_errno(err);
3259	if (err < 0 && !status)
3260		status = err;
3261
3262	return status;
3263}
3264
3265static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
3266					      int new_level)
3267{
3268	assert_spin_locked(&lockres->l_lock);
3269
3270	BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
3271
3272	if (lockres->l_level <= new_level) {
3273		mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
3274		     "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
3275		     "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
3276		     new_level, list_empty(&lockres->l_blocked_list),
3277		     list_empty(&lockres->l_mask_waiters), lockres->l_type,
3278		     lockres->l_flags, lockres->l_ro_holders,
3279		     lockres->l_ex_holders, lockres->l_action,
3280		     lockres->l_unlock_action, lockres->l_requested,
3281		     lockres->l_blocking, lockres->l_pending_gen);
3282		BUG();
3283	}
3284
3285	mlog(ML_BASTS, "lockres %s, level %d => %d, blocking %d\n",
3286	     lockres->l_name, lockres->l_level, new_level, lockres->l_blocking);
3287
3288	lockres->l_action = OCFS2_AST_DOWNCONVERT;
3289	lockres->l_requested = new_level;
3290	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3291	return lockres_set_pending(lockres);
3292}
3293
3294static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
3295				  struct ocfs2_lock_res *lockres,
3296				  int new_level,
3297				  int lvb,
3298				  unsigned int generation)
3299{
3300	int ret;
3301	u32 dlm_flags = DLM_LKF_CONVERT;
3302
3303	mlog(ML_BASTS, "lockres %s, level %d => %d\n", lockres->l_name,
3304	     lockres->l_level, new_level);
3305
3306	/*
3307	 * On DLM_LKF_VALBLK, fsdlm behaves differently with o2cb. It always
3308	 * expects DLM_LKF_VALBLK being set if the LKB has LVB, so that
3309	 * we can recover correctly from node failure. Otherwise, we may get
3310	 * invalid LVB in LKB, but without DLM_SBF_VALNOTVALID being set.
3311	 */
3312	if (!ocfs2_is_o2cb_active() &&
3313	    lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3314		lvb = 1;
3315
3316	if (lvb)
3317		dlm_flags |= DLM_LKF_VALBLK;
3318
3319	ret = ocfs2_dlm_lock(osb->cconn,
3320			     new_level,
3321			     &lockres->l_lksb,
3322			     dlm_flags,
3323			     lockres->l_name,
3324			     OCFS2_LOCK_ID_MAX_LEN - 1);
3325	lockres_clear_pending(lockres, generation, osb);
3326	if (ret) {
3327		ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
3328		ocfs2_recover_from_dlm_error(lockres, 1);
3329		goto bail;
3330	}
3331
3332	ret = 0;
3333bail:
3334	return ret;
3335}
3336
3337/* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3338static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
3339				        struct ocfs2_lock_res *lockres)
3340{
3341	assert_spin_locked(&lockres->l_lock);
3342
3343	if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
3344		/* If we're already trying to cancel a lock conversion
3345		 * then just drop the spinlock and allow the caller to
3346		 * requeue this lock. */
3347		mlog(ML_BASTS, "lockres %s, skip convert\n", lockres->l_name);
3348		return 0;
3349	}
3350
3351	/* were we in a convert when we got the bast fire? */
3352	BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
3353	       lockres->l_action != OCFS2_AST_DOWNCONVERT);
3354	/* set things up for the unlockast to know to just
3355	 * clear out the ast_action and unset busy, etc. */
3356	lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
3357
3358	mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
3359			"lock %s, invalid flags: 0x%lx\n",
3360			lockres->l_name, lockres->l_flags);
3361
3362	mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3363
3364	return 1;
3365}
3366
3367static int ocfs2_cancel_convert(struct ocfs2_super *osb,
3368				struct ocfs2_lock_res *lockres)
3369{
3370	int ret;
3371
3372	ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb,
3373			       DLM_LKF_CANCEL);
3374	if (ret) {
3375		ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3376		ocfs2_recover_from_dlm_error(lockres, 0);
3377	}
3378
3379	mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3380
3381	return ret;
3382}
3383
3384static int ocfs2_unblock_lock(struct ocfs2_super *osb,
3385			      struct ocfs2_lock_res *lockres,
3386			      struct ocfs2_unblock_ctl *ctl)
3387{
3388	unsigned long flags;
3389	int blocking;
3390	int new_level;
3391	int level;
3392	int ret = 0;
3393	int set_lvb = 0;
3394	unsigned int gen;
3395
3396	spin_lock_irqsave(&lockres->l_lock, flags);
3397
3398recheck:
3399	/*
3400	 * Is it still blocking? If not, we have no more work to do.
3401	 */
3402	if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) {
3403		BUG_ON(lockres->l_blocking != DLM_LOCK_NL);
3404		spin_unlock_irqrestore(&lockres->l_lock, flags);
3405		ret = 0;
3406		goto leave;
3407	}
3408
3409	if (lockres->l_flags & OCFS2_LOCK_BUSY) {
3410		/* XXX
3411		 * This is a *big* race.  The OCFS2_LOCK_PENDING flag
3412		 * exists entirely for one reason - another thread has set
3413		 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3414		 *
3415		 * If we do ocfs2_cancel_convert() before the other thread
3416		 * calls dlm_lock(), our cancel will do nothing.  We will
3417		 * get no ast, and we will have no way of knowing the
3418		 * cancel failed.  Meanwhile, the other thread will call
3419		 * into dlm_lock() and wait...forever.
3420		 *
3421		 * Why forever?  Because another node has asked for the
3422		 * lock first; that's why we're here in unblock_lock().
3423		 *
3424		 * The solution is OCFS2_LOCK_PENDING.  When PENDING is
3425		 * set, we just requeue the unblock.  Only when the other
3426		 * thread has called dlm_lock() and cleared PENDING will
3427		 * we then cancel their request.
3428		 *
3429		 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3430		 * at the same time they set OCFS2_DLM_BUSY.  They must
3431		 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3432		 */
3433		if (lockres->l_flags & OCFS2_LOCK_PENDING) {
3434			mlog(ML_BASTS, "lockres %s, ReQ: Pending\n",
3435			     lockres->l_name);
3436			goto leave_requeue;
3437		}
3438
3439		ctl->requeue = 1;
3440		ret = ocfs2_prepare_cancel_convert(osb, lockres);
3441		spin_unlock_irqrestore(&lockres->l_lock, flags);
3442		if (ret) {
3443			ret = ocfs2_cancel_convert(osb, lockres);
3444			if (ret < 0)
3445				mlog_errno(ret);
3446		}
3447		goto leave;
3448	}
3449
3450	/*
3451	 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3452	 * set when the ast is received for an upconvert just before the
3453	 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3454	 * on the heels of the ast, we want to delay the downconvert just
3455	 * enough to allow the up requestor to do its task. Because this
3456	 * lock is in the blocked queue, the lock will be downconverted
3457	 * as soon as the requestor is done with the lock.
3458	 */
3459	if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
3460		goto leave_requeue;
3461
3462	/*
3463	 * How can we block and yet be at NL?  We were trying to upconvert
3464	 * from NL and got canceled.  The code comes back here, and now
3465	 * we notice and clear BLOCKING.
3466	 */
3467	if (lockres->l_level == DLM_LOCK_NL) {
3468		BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders);
3469		mlog(ML_BASTS, "lockres %s, Aborting dc\n", lockres->l_name);
3470		lockres->l_blocking = DLM_LOCK_NL;
3471		lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
3472		spin_unlock_irqrestore(&lockres->l_lock, flags);
3473		goto leave;
3474	}
3475
3476	/* if we're blocking an exclusive and we have *any* holders,
3477	 * then requeue. */
3478	if ((lockres->l_blocking == DLM_LOCK_EX)
3479	    && (lockres->l_ex_holders || lockres->l_ro_holders)) {
3480		mlog(ML_BASTS, "lockres %s, ReQ: EX/PR Holders %u,%u\n",
3481		     lockres->l_name, lockres->l_ex_holders,
3482		     lockres->l_ro_holders);
3483		goto leave_requeue;
3484	}
3485
3486	/* If it's a PR we're blocking, then only
3487	 * requeue if we've got any EX holders */
3488	if (lockres->l_blocking == DLM_LOCK_PR &&
3489	    lockres->l_ex_holders) {
3490		mlog(ML_BASTS, "lockres %s, ReQ: EX Holders %u\n",
3491		     lockres->l_name, lockres->l_ex_holders);
3492		goto leave_requeue;
3493	}
3494
3495	/*
3496	 * Can we get a lock in this state if the holder counts are
3497	 * zero? The meta data unblock code used to check this.
3498	 */
3499	if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
3500	    && (lockres->l_flags & OCFS2_LOCK_REFRESHING)) {
3501		mlog(ML_BASTS, "lockres %s, ReQ: Lock Refreshing\n",
3502		     lockres->l_name);
3503		goto leave_requeue;
3504	}
3505
3506	new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
3507
3508	if (lockres->l_ops->check_downconvert
3509	    && !lockres->l_ops->check_downconvert(lockres, new_level)) {
3510		mlog(ML_BASTS, "lockres %s, ReQ: Checkpointing\n",
3511		     lockres->l_name);
3512		goto leave_requeue;
3513	}
3514
3515	/* If we get here, then we know that there are no more
3516	 * incompatible holders (and anyone asking for an incompatible
3517	 * lock is blocked). We can now downconvert the lock */
3518	if (!lockres->l_ops->downconvert_worker)
3519		goto downconvert;
3520
3521	/* Some lockres types want to do a bit of work before
3522	 * downconverting a lock. Allow that here. The worker function
3523	 * may sleep, so we save off a copy of what we're blocking as
3524	 * it may change while we're not holding the spin lock. */
3525	blocking = lockres->l_blocking;
3526	level = lockres->l_level;
3527	spin_unlock_irqrestore(&lockres->l_lock, flags);
3528
3529	ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
3530
3531	if (ctl->unblock_action == UNBLOCK_STOP_POST) {
3532		mlog(ML_BASTS, "lockres %s, UNBLOCK_STOP_POST\n",
3533		     lockres->l_name);
3534		goto leave;
3535	}
3536
3537	spin_lock_irqsave(&lockres->l_lock, flags);
3538	if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) {
3539		/* If this changed underneath us, then we can't drop
3540		 * it just yet. */
3541		mlog(ML_BASTS, "lockres %s, block=%d:%d, level=%d:%d, "
3542		     "Recheck\n", lockres->l_name, blocking,
3543		     lockres->l_blocking, level, lockres->l_level);
3544		goto recheck;
3545	}
3546
3547downconvert:
3548	ctl->requeue = 0;
3549
3550	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3551		if (lockres->l_level == DLM_LOCK_EX)
3552			set_lvb = 1;
3553
3554		/*
3555		 * We only set the lvb if the lock has been fully
3556		 * refreshed - otherwise we risk setting stale
3557		 * data. Otherwise, there's no need to actually clear
3558		 * out the lvb here as it's value is still valid.
3559		 */
3560		if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3561			lockres->l_ops->set_lvb(lockres);
3562	}
3563
3564	gen = ocfs2_prepare_downconvert(lockres, new_level);
3565	spin_unlock_irqrestore(&lockres->l_lock, flags);
3566	ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
3567				     gen);
 
 
 
 
 
 
 
 
 
 
 
3568
3569leave:
3570	if (ret)
3571		mlog_errno(ret);
3572	return ret;
3573
3574leave_requeue:
3575	spin_unlock_irqrestore(&lockres->l_lock, flags);
3576	ctl->requeue = 1;
3577
3578	return 0;
3579}
3580
3581static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
3582				     int blocking)
3583{
3584	struct inode *inode;
3585	struct address_space *mapping;
3586	struct ocfs2_inode_info *oi;
3587
3588       	inode = ocfs2_lock_res_inode(lockres);
3589	mapping = inode->i_mapping;
3590
3591	if (S_ISDIR(inode->i_mode)) {
3592		oi = OCFS2_I(inode);
3593		oi->ip_dir_lock_gen++;
3594		mlog(0, "generation: %u\n", oi->ip_dir_lock_gen);
3595		goto out;
3596	}
3597
3598	if (!S_ISREG(inode->i_mode))
3599		goto out;
3600
3601	/*
3602	 * We need this before the filemap_fdatawrite() so that it can
3603	 * transfer the dirty bit from the PTE to the
3604	 * page. Unfortunately this means that even for EX->PR
3605	 * downconverts, we'll lose our mappings and have to build
3606	 * them up again.
3607	 */
3608	unmap_mapping_range(mapping, 0, 0, 0);
3609
3610	if (filemap_fdatawrite(mapping)) {
3611		mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
3612		     (unsigned long long)OCFS2_I(inode)->ip_blkno);
3613	}
3614	sync_mapping_buffers(mapping);
3615	if (blocking == DLM_LOCK_EX) {
3616		truncate_inode_pages(mapping, 0);
3617	} else {
3618		/* We only need to wait on the I/O if we're not also
3619		 * truncating pages because truncate_inode_pages waits
3620		 * for us above. We don't truncate pages if we're
3621		 * blocking anything < EXMODE because we want to keep
3622		 * them around in that case. */
3623		filemap_fdatawait(mapping);
3624	}
3625
 
3626	forget_all_cached_acls(inode);
3627
3628out:
3629	return UNBLOCK_CONTINUE;
3630}
3631
3632static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci,
3633				 struct ocfs2_lock_res *lockres,
3634				 int new_level)
3635{
3636	int checkpointed = ocfs2_ci_fully_checkpointed(ci);
3637
3638	BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR);
3639	BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed);
3640
3641	if (checkpointed)
3642		return 1;
3643
3644	ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci)));
3645	return 0;
3646}
3647
3648static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
3649					int new_level)
3650{
3651	struct inode *inode = ocfs2_lock_res_inode(lockres);
3652
3653	return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level);
3654}
3655
3656static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
3657{
3658	struct inode *inode = ocfs2_lock_res_inode(lockres);
3659
3660	__ocfs2_stuff_meta_lvb(inode);
3661}
3662
3663/*
3664 * Does the final reference drop on our dentry lock. Right now this
3665 * happens in the downconvert thread, but we could choose to simplify the
3666 * dlmglue API and push these off to the ocfs2_wq in the future.
3667 */
3668static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
3669				     struct ocfs2_lock_res *lockres)
3670{
3671	struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3672	ocfs2_dentry_lock_put(osb, dl);
3673}
3674
3675/*
3676 * d_delete() matching dentries before the lock downconvert.
3677 *
3678 * At this point, any process waiting to destroy the
3679 * dentry_lock due to last ref count is stopped by the
3680 * OCFS2_LOCK_QUEUED flag.
3681 *
3682 * We have two potential problems
3683 *
3684 * 1) If we do the last reference drop on our dentry_lock (via dput)
3685 *    we'll wind up in ocfs2_release_dentry_lock(), waiting on
3686 *    the downconvert to finish. Instead we take an elevated
3687 *    reference and push the drop until after we've completed our
3688 *    unblock processing.
3689 *
3690 * 2) There might be another process with a final reference,
3691 *    waiting on us to finish processing. If this is the case, we
3692 *    detect it and exit out - there's no more dentries anyway.
3693 */
3694static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
3695				       int blocking)
3696{
3697	struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3698	struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
3699	struct dentry *dentry;
3700	unsigned long flags;
3701	int extra_ref = 0;
3702
3703	/*
3704	 * This node is blocking another node from getting a read
3705	 * lock. This happens when we've renamed within a
3706	 * directory. We've forced the other nodes to d_delete(), but
3707	 * we never actually dropped our lock because it's still
3708	 * valid. The downconvert code will retain a PR for this node,
3709	 * so there's no further work to do.
3710	 */
3711	if (blocking == DLM_LOCK_PR)
3712		return UNBLOCK_CONTINUE;
3713
3714	/*
3715	 * Mark this inode as potentially orphaned. The code in
3716	 * ocfs2_delete_inode() will figure out whether it actually
3717	 * needs to be freed or not.
3718	 */
3719	spin_lock(&oi->ip_lock);
3720	oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
3721	spin_unlock(&oi->ip_lock);
3722
3723	/*
3724	 * Yuck. We need to make sure however that the check of
3725	 * OCFS2_LOCK_FREEING and the extra reference are atomic with
3726	 * respect to a reference decrement or the setting of that
3727	 * flag.
3728	 */
3729	spin_lock_irqsave(&lockres->l_lock, flags);
3730	spin_lock(&dentry_attach_lock);
3731	if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
3732	    && dl->dl_count) {
3733		dl->dl_count++;
3734		extra_ref = 1;
3735	}
3736	spin_unlock(&dentry_attach_lock);
3737	spin_unlock_irqrestore(&lockres->l_lock, flags);
3738
3739	mlog(0, "extra_ref = %d\n", extra_ref);
3740
3741	/*
3742	 * We have a process waiting on us in ocfs2_dentry_iput(),
3743	 * which means we can't have any more outstanding
3744	 * aliases. There's no need to do any more work.
3745	 */
3746	if (!extra_ref)
3747		return UNBLOCK_CONTINUE;
3748
3749	spin_lock(&dentry_attach_lock);
3750	while (1) {
3751		dentry = ocfs2_find_local_alias(dl->dl_inode,
3752						dl->dl_parent_blkno, 1);
3753		if (!dentry)
3754			break;
3755		spin_unlock(&dentry_attach_lock);
3756
3757		if (S_ISDIR(dl->dl_inode->i_mode))
3758			shrink_dcache_parent(dentry);
3759
3760		mlog(0, "d_delete(%pd);\n", dentry);
3761
3762		/*
3763		 * The following dcache calls may do an
3764		 * iput(). Normally we don't want that from the
3765		 * downconverting thread, but in this case it's ok
3766		 * because the requesting node already has an
3767		 * exclusive lock on the inode, so it can't be queued
3768		 * for a downconvert.
3769		 */
3770		d_delete(dentry);
3771		dput(dentry);
3772
3773		spin_lock(&dentry_attach_lock);
3774	}
3775	spin_unlock(&dentry_attach_lock);
3776
3777	/*
3778	 * If we are the last holder of this dentry lock, there is no
3779	 * reason to downconvert so skip straight to the unlock.
3780	 */
3781	if (dl->dl_count == 1)
3782		return UNBLOCK_STOP_POST;
3783
3784	return UNBLOCK_CONTINUE_POST;
3785}
3786
3787static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
3788					    int new_level)
3789{
3790	struct ocfs2_refcount_tree *tree =
3791				ocfs2_lock_res_refcount_tree(lockres);
3792
3793	return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level);
3794}
3795
3796static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
3797					 int blocking)
3798{
3799	struct ocfs2_refcount_tree *tree =
3800				ocfs2_lock_res_refcount_tree(lockres);
3801
3802	ocfs2_metadata_cache_purge(&tree->rf_ci);
3803
3804	return UNBLOCK_CONTINUE;
3805}
3806
3807static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres)
3808{
3809	struct ocfs2_qinfo_lvb *lvb;
3810	struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres);
3811	struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3812					    oinfo->dqi_gi.dqi_type);
3813
3814	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3815	lvb->lvb_version = OCFS2_QINFO_LVB_VERSION;
3816	lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace);
3817	lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace);
3818	lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms);
3819	lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks);
3820	lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk);
3821	lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry);
3822}
3823
3824void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3825{
3826	struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3827	struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3828	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3829
3830	if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
3831		ocfs2_cluster_unlock(osb, lockres, level);
3832}
3833
3834static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo)
3835{
3836	struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3837					    oinfo->dqi_gi.dqi_type);
3838	struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3839	struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3840	struct buffer_head *bh = NULL;
3841	struct ocfs2_global_disk_dqinfo *gdinfo;
3842	int status = 0;
3843
3844	if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
3845	    lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) {
3846		info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace);
3847		info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace);
3848		oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms);
3849		oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks);
3850		oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk);
3851		oinfo->dqi_gi.dqi_free_entry =
3852					be32_to_cpu(lvb->lvb_free_entry);
3853	} else {
3854		status = ocfs2_read_quota_phys_block(oinfo->dqi_gqinode,
3855						     oinfo->dqi_giblk, &bh);
3856		if (status) {
3857			mlog_errno(status);
3858			goto bail;
3859		}
3860		gdinfo = (struct ocfs2_global_disk_dqinfo *)
3861					(bh->b_data + OCFS2_GLOBAL_INFO_OFF);
3862		info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace);
3863		info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace);
3864		oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms);
3865		oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks);
3866		oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk);
3867		oinfo->dqi_gi.dqi_free_entry =
3868					le32_to_cpu(gdinfo->dqi_free_entry);
3869		brelse(bh);
3870		ocfs2_track_lock_refresh(lockres);
3871	}
3872
3873bail:
3874	return status;
3875}
3876
3877/* Lock quota info, this function expects at least shared lock on the quota file
3878 * so that we can safely refresh quota info from disk. */
3879int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3880{
3881	struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3882	struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3883	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3884	int status = 0;
3885
3886	/* On RO devices, locking really isn't needed... */
3887	if (ocfs2_is_hard_readonly(osb)) {
3888		if (ex)
3889			status = -EROFS;
3890		goto bail;
3891	}
3892	if (ocfs2_mount_local(osb))
3893		goto bail;
3894
3895	status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3896	if (status < 0) {
3897		mlog_errno(status);
3898		goto bail;
3899	}
3900	if (!ocfs2_should_refresh_lock_res(lockres))
3901		goto bail;
3902	/* OK, we have the lock but we need to refresh the quota info */
3903	status = ocfs2_refresh_qinfo(oinfo);
3904	if (status)
3905		ocfs2_qinfo_unlock(oinfo, ex);
3906	ocfs2_complete_lock_res_refresh(lockres, status);
3907bail:
3908	return status;
3909}
3910
3911int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex)
3912{
3913	int status;
3914	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3915	struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3916	struct ocfs2_super *osb = lockres->l_priv;
3917
3918
3919	if (ocfs2_is_hard_readonly(osb))
3920		return -EROFS;
3921
3922	if (ocfs2_mount_local(osb))
3923		return 0;
3924
3925	status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3926	if (status < 0)
3927		mlog_errno(status);
3928
3929	return status;
3930}
3931
3932void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex)
3933{
3934	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3935	struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3936	struct ocfs2_super *osb = lockres->l_priv;
3937
3938	if (!ocfs2_mount_local(osb))
3939		ocfs2_cluster_unlock(osb, lockres, level);
3940}
3941
3942static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3943				       struct ocfs2_lock_res *lockres)
3944{
3945	int status;
3946	struct ocfs2_unblock_ctl ctl = {0, 0,};
3947	unsigned long flags;
3948
3949	/* Our reference to the lockres in this function can be
3950	 * considered valid until we remove the OCFS2_LOCK_QUEUED
3951	 * flag. */
3952
3953	BUG_ON(!lockres);
3954	BUG_ON(!lockres->l_ops);
3955
3956	mlog(ML_BASTS, "lockres %s blocked\n", lockres->l_name);
3957
3958	/* Detect whether a lock has been marked as going away while
3959	 * the downconvert thread was processing other things. A lock can
3960	 * still be marked with OCFS2_LOCK_FREEING after this check,
3961	 * but short circuiting here will still save us some
3962	 * performance. */
3963	spin_lock_irqsave(&lockres->l_lock, flags);
3964	if (lockres->l_flags & OCFS2_LOCK_FREEING)
3965		goto unqueue;
3966	spin_unlock_irqrestore(&lockres->l_lock, flags);
3967
3968	status = ocfs2_unblock_lock(osb, lockres, &ctl);
3969	if (status < 0)
3970		mlog_errno(status);
3971
3972	spin_lock_irqsave(&lockres->l_lock, flags);
3973unqueue:
3974	if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
3975		lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
3976	} else
3977		ocfs2_schedule_blocked_lock(osb, lockres);
3978
3979	mlog(ML_BASTS, "lockres %s, requeue = %s.\n", lockres->l_name,
3980	     ctl.requeue ? "yes" : "no");
3981	spin_unlock_irqrestore(&lockres->l_lock, flags);
3982
3983	if (ctl.unblock_action != UNBLOCK_CONTINUE
3984	    && lockres->l_ops->post_unlock)
3985		lockres->l_ops->post_unlock(osb, lockres);
3986}
3987
3988static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
3989					struct ocfs2_lock_res *lockres)
3990{
3991	unsigned long flags;
3992
3993	assert_spin_locked(&lockres->l_lock);
3994
3995	if (lockres->l_flags & OCFS2_LOCK_FREEING) {
3996		/* Do not schedule a lock for downconvert when it's on
3997		 * the way to destruction - any nodes wanting access
3998		 * to the resource will get it soon. */
3999		mlog(ML_BASTS, "lockres %s won't be scheduled: flags 0x%lx\n",
4000		     lockres->l_name, lockres->l_flags);
4001		return;
4002	}
4003
4004	lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
4005
4006	spin_lock_irqsave(&osb->dc_task_lock, flags);
4007	if (list_empty(&lockres->l_blocked_list)) {
4008		list_add_tail(&lockres->l_blocked_list,
4009			      &osb->blocked_lock_list);
4010		osb->blocked_lock_count++;
4011	}
4012	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4013}
4014
4015static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb)
4016{
4017	unsigned long processed;
4018	unsigned long flags;
4019	struct ocfs2_lock_res *lockres;
4020
4021	spin_lock_irqsave(&osb->dc_task_lock, flags);
4022	/* grab this early so we know to try again if a state change and
4023	 * wake happens part-way through our work  */
4024	osb->dc_work_sequence = osb->dc_wake_sequence;
4025
4026	processed = osb->blocked_lock_count;
4027	/*
4028	 * blocked lock processing in this loop might call iput which can
4029	 * remove items off osb->blocked_lock_list. Downconvert up to
4030	 * 'processed' number of locks, but stop short if we had some
4031	 * removed in ocfs2_mark_lockres_freeing when downconverting.
4032	 */
4033	while (processed && !list_empty(&osb->blocked_lock_list)) {
4034		lockres = list_entry(osb->blocked_lock_list.next,
4035				     struct ocfs2_lock_res, l_blocked_list);
4036		list_del_init(&lockres->l_blocked_list);
4037		osb->blocked_lock_count--;
4038		spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4039
4040		BUG_ON(!processed);
4041		processed--;
4042
4043		ocfs2_process_blocked_lock(osb, lockres);
4044
4045		spin_lock_irqsave(&osb->dc_task_lock, flags);
4046	}
4047	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4048}
4049
4050static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb)
4051{
4052	int empty = 0;
4053	unsigned long flags;
4054
4055	spin_lock_irqsave(&osb->dc_task_lock, flags);
4056	if (list_empty(&osb->blocked_lock_list))
4057		empty = 1;
4058
4059	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4060	return empty;
4061}
4062
4063static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb)
4064{
4065	int should_wake = 0;
4066	unsigned long flags;
4067
4068	spin_lock_irqsave(&osb->dc_task_lock, flags);
4069	if (osb->dc_work_sequence != osb->dc_wake_sequence)
4070		should_wake = 1;
4071	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4072
4073	return should_wake;
4074}
4075
4076static int ocfs2_downconvert_thread(void *arg)
4077{
4078	int status = 0;
4079	struct ocfs2_super *osb = arg;
4080
4081	/* only quit once we've been asked to stop and there is no more
4082	 * work available */
4083	while (!(kthread_should_stop() &&
4084		ocfs2_downconvert_thread_lists_empty(osb))) {
4085
4086		wait_event_interruptible(osb->dc_event,
4087					 ocfs2_downconvert_thread_should_wake(osb) ||
4088					 kthread_should_stop());
4089
4090		mlog(0, "downconvert_thread: awoken\n");
4091
4092		ocfs2_downconvert_thread_do_work(osb);
4093	}
4094
4095	osb->dc_task = NULL;
4096	return status;
4097}
4098
4099void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb)
4100{
4101	unsigned long flags;
4102
4103	spin_lock_irqsave(&osb->dc_task_lock, flags);
4104	/* make sure the voting thread gets a swipe at whatever changes
4105	 * the caller may have made to the voting state */
4106	osb->dc_wake_sequence++;
4107	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4108	wake_up(&osb->dc_event);
4109}
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
 
   3 * dlmglue.c
   4 *
   5 * Code which implements an OCFS2 specific interface to our DLM.
   6 *
   7 * Copyright (C) 2003, 2004 Oracle.  All rights reserved.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   8 */
   9
  10#include <linux/types.h>
  11#include <linux/slab.h>
  12#include <linux/highmem.h>
  13#include <linux/mm.h>
  14#include <linux/kthread.h>
  15#include <linux/pagemap.h>
  16#include <linux/debugfs.h>
  17#include <linux/seq_file.h>
  18#include <linux/time.h>
  19#include <linux/delay.h>
  20#include <linux/quotaops.h>
  21#include <linux/sched/signal.h>
  22
  23#define MLOG_MASK_PREFIX ML_DLM_GLUE
  24#include <cluster/masklog.h>
  25
  26#include "ocfs2.h"
  27#include "ocfs2_lockingver.h"
  28
  29#include "alloc.h"
  30#include "dcache.h"
  31#include "dlmglue.h"
  32#include "extent_map.h"
  33#include "file.h"
  34#include "heartbeat.h"
  35#include "inode.h"
  36#include "journal.h"
  37#include "stackglue.h"
  38#include "slot_map.h"
  39#include "super.h"
  40#include "uptodate.h"
  41#include "quota.h"
  42#include "refcounttree.h"
  43#include "acl.h"
  44
  45#include "buffer_head_io.h"
  46
  47struct ocfs2_mask_waiter {
  48	struct list_head	mw_item;
  49	int			mw_status;
  50	struct completion	mw_complete;
  51	unsigned long		mw_mask;
  52	unsigned long		mw_goal;
  53#ifdef CONFIG_OCFS2_FS_STATS
  54	ktime_t			mw_lock_start;
  55#endif
  56};
  57
  58static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres);
  59static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres);
  60static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres);
  61static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres);
  62
  63/*
  64 * Return value from ->downconvert_worker functions.
  65 *
  66 * These control the precise actions of ocfs2_unblock_lock()
  67 * and ocfs2_process_blocked_lock()
  68 *
  69 */
  70enum ocfs2_unblock_action {
  71	UNBLOCK_CONTINUE	= 0, /* Continue downconvert */
  72	UNBLOCK_CONTINUE_POST	= 1, /* Continue downconvert, fire
  73				      * ->post_unlock callback */
  74	UNBLOCK_STOP_POST	= 2, /* Do not downconvert, fire
  75				      * ->post_unlock() callback. */
  76};
  77
  78struct ocfs2_unblock_ctl {
  79	int requeue;
  80	enum ocfs2_unblock_action unblock_action;
  81};
  82
  83/* Lockdep class keys */
  84#ifdef CONFIG_DEBUG_LOCK_ALLOC
  85static struct lock_class_key lockdep_keys[OCFS2_NUM_LOCK_TYPES];
  86#endif
  87
  88static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
  89					int new_level);
  90static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres);
  91
  92static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
  93				     int blocking);
  94
  95static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
  96				       int blocking);
  97
  98static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
  99				     struct ocfs2_lock_res *lockres);
 100
 101static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres);
 102
 103static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
 104					    int new_level);
 105static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
 106					 int blocking);
 107
 108#define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
 109
 110/* This aids in debugging situations where a bad LVB might be involved. */
 111static void ocfs2_dump_meta_lvb_info(u64 level,
 112				     const char *function,
 113				     unsigned int line,
 114				     struct ocfs2_lock_res *lockres)
 115{
 116	struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
 117
 118	mlog(level, "LVB information for %s (called from %s:%u):\n",
 119	     lockres->l_name, function, line);
 120	mlog(level, "version: %u, clusters: %u, generation: 0x%x\n",
 121	     lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters),
 122	     be32_to_cpu(lvb->lvb_igeneration));
 123	mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n",
 124	     (unsigned long long)be64_to_cpu(lvb->lvb_isize),
 125	     be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid),
 126	     be16_to_cpu(lvb->lvb_imode));
 127	mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
 128	     "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink),
 129	     (long long)be64_to_cpu(lvb->lvb_iatime_packed),
 130	     (long long)be64_to_cpu(lvb->lvb_ictime_packed),
 131	     (long long)be64_to_cpu(lvb->lvb_imtime_packed),
 132	     be32_to_cpu(lvb->lvb_iattr));
 133}
 134
 135
 136/*
 137 * OCFS2 Lock Resource Operations
 138 *
 139 * These fine tune the behavior of the generic dlmglue locking infrastructure.
 140 *
 141 * The most basic of lock types can point ->l_priv to their respective
 142 * struct ocfs2_super and allow the default actions to manage things.
 143 *
 144 * Right now, each lock type also needs to implement an init function,
 145 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
 146 * should be called when the lock is no longer needed (i.e., object
 147 * destruction time).
 148 */
 149struct ocfs2_lock_res_ops {
 150	/*
 151	 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
 152	 * this callback if ->l_priv is not an ocfs2_super pointer
 153	 */
 154	struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *);
 155
 156	/*
 157	 * Optionally called in the downconvert thread after a
 158	 * successful downconvert. The lockres will not be referenced
 159	 * after this callback is called, so it is safe to free
 160	 * memory, etc.
 161	 *
 162	 * The exact semantics of when this is called are controlled
 163	 * by ->downconvert_worker()
 164	 */
 165	void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *);
 166
 167	/*
 168	 * Allow a lock type to add checks to determine whether it is
 169	 * safe to downconvert a lock. Return 0 to re-queue the
 170	 * downconvert at a later time, nonzero to continue.
 171	 *
 172	 * For most locks, the default checks that there are no
 173	 * incompatible holders are sufficient.
 174	 *
 175	 * Called with the lockres spinlock held.
 176	 */
 177	int (*check_downconvert)(struct ocfs2_lock_res *, int);
 178
 179	/*
 180	 * Allows a lock type to populate the lock value block. This
 181	 * is called on downconvert, and when we drop a lock.
 182	 *
 183	 * Locks that want to use this should set LOCK_TYPE_USES_LVB
 184	 * in the flags field.
 185	 *
 186	 * Called with the lockres spinlock held.
 187	 */
 188	void (*set_lvb)(struct ocfs2_lock_res *);
 189
 190	/*
 191	 * Called from the downconvert thread when it is determined
 192	 * that a lock will be downconverted. This is called without
 193	 * any locks held so the function can do work that might
 194	 * schedule (syncing out data, etc).
 195	 *
 196	 * This should return any one of the ocfs2_unblock_action
 197	 * values, depending on what it wants the thread to do.
 198	 */
 199	int (*downconvert_worker)(struct ocfs2_lock_res *, int);
 200
 201	/*
 202	 * LOCK_TYPE_* flags which describe the specific requirements
 203	 * of a lock type. Descriptions of each individual flag follow.
 204	 */
 205	int flags;
 206};
 207
 208/*
 209 * Some locks want to "refresh" potentially stale data when a
 210 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
 211 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
 212 * individual lockres l_flags member from the ast function. It is
 213 * expected that the locking wrapper will clear the
 214 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
 215 */
 216#define LOCK_TYPE_REQUIRES_REFRESH 0x1
 217
 218/*
 219 * Indicate that a lock type makes use of the lock value block. The
 220 * ->set_lvb lock type callback must be defined.
 221 */
 222#define LOCK_TYPE_USES_LVB		0x2
 223
 224static const struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = {
 225	.get_osb	= ocfs2_get_inode_osb,
 226	.flags		= 0,
 227};
 228
 229static const struct ocfs2_lock_res_ops ocfs2_inode_inode_lops = {
 230	.get_osb	= ocfs2_get_inode_osb,
 231	.check_downconvert = ocfs2_check_meta_downconvert,
 232	.set_lvb	= ocfs2_set_meta_lvb,
 233	.downconvert_worker = ocfs2_data_convert_worker,
 234	.flags		= LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
 235};
 236
 237static const struct ocfs2_lock_res_ops ocfs2_super_lops = {
 238	.flags		= LOCK_TYPE_REQUIRES_REFRESH,
 239};
 240
 241static const struct ocfs2_lock_res_ops ocfs2_rename_lops = {
 242	.flags		= 0,
 243};
 244
 245static const struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops = {
 246	.flags		= 0,
 247};
 248
 249static const struct ocfs2_lock_res_ops ocfs2_trim_fs_lops = {
 250	.flags		= LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
 251};
 252
 253static const struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops = {
 254	.flags		= LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
 255};
 256
 257static const struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
 258	.get_osb	= ocfs2_get_dentry_osb,
 259	.post_unlock	= ocfs2_dentry_post_unlock,
 260	.downconvert_worker = ocfs2_dentry_convert_worker,
 261	.flags		= 0,
 262};
 263
 264static const struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
 265	.get_osb	= ocfs2_get_inode_osb,
 266	.flags		= 0,
 267};
 268
 269static const struct ocfs2_lock_res_ops ocfs2_flock_lops = {
 270	.get_osb	= ocfs2_get_file_osb,
 271	.flags		= 0,
 272};
 273
 274static const struct ocfs2_lock_res_ops ocfs2_qinfo_lops = {
 275	.set_lvb	= ocfs2_set_qinfo_lvb,
 276	.get_osb	= ocfs2_get_qinfo_osb,
 277	.flags		= LOCK_TYPE_REQUIRES_REFRESH | LOCK_TYPE_USES_LVB,
 278};
 279
 280static const struct ocfs2_lock_res_ops ocfs2_refcount_block_lops = {
 281	.check_downconvert = ocfs2_check_refcount_downconvert,
 282	.downconvert_worker = ocfs2_refcount_convert_worker,
 283	.flags		= 0,
 284};
 285
 286static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
 287{
 288	return lockres->l_type == OCFS2_LOCK_TYPE_META ||
 289		lockres->l_type == OCFS2_LOCK_TYPE_RW ||
 290		lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
 291}
 292
 293static inline struct ocfs2_lock_res *ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb)
 294{
 295	return container_of(lksb, struct ocfs2_lock_res, l_lksb);
 296}
 297
 298static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
 299{
 300	BUG_ON(!ocfs2_is_inode_lock(lockres));
 301
 302	return (struct inode *) lockres->l_priv;
 303}
 304
 305static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres)
 306{
 307	BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY);
 308
 309	return (struct ocfs2_dentry_lock *)lockres->l_priv;
 310}
 311
 312static inline struct ocfs2_mem_dqinfo *ocfs2_lock_res_qinfo(struct ocfs2_lock_res *lockres)
 313{
 314	BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_QINFO);
 315
 316	return (struct ocfs2_mem_dqinfo *)lockres->l_priv;
 317}
 318
 319static inline struct ocfs2_refcount_tree *
 320ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res *res)
 321{
 322	return container_of(res, struct ocfs2_refcount_tree, rf_lockres);
 323}
 324
 325static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres)
 326{
 327	if (lockres->l_ops->get_osb)
 328		return lockres->l_ops->get_osb(lockres);
 329
 330	return (struct ocfs2_super *)lockres->l_priv;
 331}
 332
 333static int ocfs2_lock_create(struct ocfs2_super *osb,
 334			     struct ocfs2_lock_res *lockres,
 335			     int level,
 336			     u32 dlm_flags);
 337static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
 338						     int wanted);
 339static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
 340				   struct ocfs2_lock_res *lockres,
 341				   int level, unsigned long caller_ip);
 342static inline void ocfs2_cluster_unlock(struct ocfs2_super *osb,
 343					struct ocfs2_lock_res *lockres,
 344					int level)
 345{
 346	__ocfs2_cluster_unlock(osb, lockres, level, _RET_IP_);
 347}
 348
 349static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres);
 350static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres);
 351static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres);
 352static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level);
 353static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
 354					struct ocfs2_lock_res *lockres);
 355static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
 356						int convert);
 357#define ocfs2_log_dlm_error(_func, _err, _lockres) do {					\
 358	if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY)				\
 359		mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n",	\
 360		     _err, _func, _lockres->l_name);					\
 361	else										\
 362		mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n",	\
 363		     _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name,	\
 364		     (unsigned int)ocfs2_get_dentry_lock_ino(_lockres));		\
 365} while (0)
 366static int ocfs2_downconvert_thread(void *arg);
 367static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
 368					struct ocfs2_lock_res *lockres);
 369static int ocfs2_inode_lock_update(struct inode *inode,
 370				  struct buffer_head **bh);
 371static void ocfs2_drop_osb_locks(struct ocfs2_super *osb);
 372static inline int ocfs2_highest_compat_lock_level(int level);
 373static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
 374					      int new_level);
 375static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
 376				  struct ocfs2_lock_res *lockres,
 377				  int new_level,
 378				  int lvb,
 379				  unsigned int generation);
 380static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
 381				        struct ocfs2_lock_res *lockres);
 382static int ocfs2_cancel_convert(struct ocfs2_super *osb,
 383				struct ocfs2_lock_res *lockres);
 384
 385
 386static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
 387				  u64 blkno,
 388				  u32 generation,
 389				  char *name)
 390{
 391	int len;
 392
 393	BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
 394
 395	len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
 396		       ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
 397		       (long long)blkno, generation);
 398
 399	BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
 400
 401	mlog(0, "built lock resource with name: %s\n", name);
 402}
 403
 404static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
 405
 406static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
 407				       struct ocfs2_dlm_debug *dlm_debug)
 408{
 409	mlog(0, "Add tracking for lockres %s\n", res->l_name);
 410
 411	spin_lock(&ocfs2_dlm_tracking_lock);
 412	list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
 413	spin_unlock(&ocfs2_dlm_tracking_lock);
 414}
 415
 416static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
 417{
 418	spin_lock(&ocfs2_dlm_tracking_lock);
 419	if (!list_empty(&res->l_debug_list))
 420		list_del_init(&res->l_debug_list);
 421	spin_unlock(&ocfs2_dlm_tracking_lock);
 422}
 423
 424#ifdef CONFIG_OCFS2_FS_STATS
 425static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
 426{
 427	res->l_lock_refresh = 0;
 428	res->l_lock_wait = 0;
 429	memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats));
 430	memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats));
 431}
 432
 433static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
 434				    struct ocfs2_mask_waiter *mw, int ret)
 435{
 436	u32 usec;
 437	ktime_t kt;
 438	struct ocfs2_lock_stats *stats;
 439
 440	if (level == LKM_PRMODE)
 441		stats = &res->l_lock_prmode;
 442	else if (level == LKM_EXMODE)
 443		stats = &res->l_lock_exmode;
 444	else
 445		return;
 446
 447	kt = ktime_sub(ktime_get(), mw->mw_lock_start);
 448	usec = ktime_to_us(kt);
 449
 450	stats->ls_gets++;
 451	stats->ls_total += ktime_to_ns(kt);
 452	/* overflow */
 453	if (unlikely(stats->ls_gets == 0)) {
 454		stats->ls_gets++;
 455		stats->ls_total = ktime_to_ns(kt);
 456	}
 457
 458	if (stats->ls_max < usec)
 459		stats->ls_max = usec;
 460
 461	if (ret)
 462		stats->ls_fail++;
 463
 464	stats->ls_last = ktime_to_us(ktime_get_real());
 465}
 466
 467static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
 468{
 469	lockres->l_lock_refresh++;
 470}
 471
 472static inline void ocfs2_track_lock_wait(struct ocfs2_lock_res *lockres)
 473{
 474	struct ocfs2_mask_waiter *mw;
 475
 476	if (list_empty(&lockres->l_mask_waiters)) {
 477		lockres->l_lock_wait = 0;
 478		return;
 479	}
 480
 481	mw = list_first_entry(&lockres->l_mask_waiters,
 482				struct ocfs2_mask_waiter, mw_item);
 483	lockres->l_lock_wait =
 484			ktime_to_us(ktime_mono_to_real(mw->mw_lock_start));
 485}
 486
 487static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
 488{
 489	mw->mw_lock_start = ktime_get();
 490}
 491#else
 492static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
 493{
 494}
 495static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res,
 496			   int level, struct ocfs2_mask_waiter *mw, int ret)
 497{
 498}
 499static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
 500{
 501}
 502static inline void ocfs2_track_lock_wait(struct ocfs2_lock_res *lockres)
 503{
 504}
 505static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
 506{
 507}
 508#endif
 509
 510static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
 511				       struct ocfs2_lock_res *res,
 512				       enum ocfs2_lock_type type,
 513				       const struct ocfs2_lock_res_ops *ops,
 514				       void *priv)
 515{
 516	res->l_type          = type;
 517	res->l_ops           = ops;
 518	res->l_priv          = priv;
 519
 520	res->l_level         = DLM_LOCK_IV;
 521	res->l_requested     = DLM_LOCK_IV;
 522	res->l_blocking      = DLM_LOCK_IV;
 523	res->l_action        = OCFS2_AST_INVALID;
 524	res->l_unlock_action = OCFS2_UNLOCK_INVALID;
 525
 526	res->l_flags         = OCFS2_LOCK_INITIALIZED;
 527
 528	ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
 529
 530	ocfs2_init_lock_stats(res);
 531#ifdef CONFIG_DEBUG_LOCK_ALLOC
 532	if (type != OCFS2_LOCK_TYPE_OPEN)
 533		lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type],
 534				 &lockdep_keys[type], 0);
 535	else
 536		res->l_lockdep_map.key = NULL;
 537#endif
 538}
 539
 540void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
 541{
 542	/* This also clears out the lock status block */
 543	memset(res, 0, sizeof(struct ocfs2_lock_res));
 544	spin_lock_init(&res->l_lock);
 545	init_waitqueue_head(&res->l_event);
 546	INIT_LIST_HEAD(&res->l_blocked_list);
 547	INIT_LIST_HEAD(&res->l_mask_waiters);
 548	INIT_LIST_HEAD(&res->l_holders);
 549}
 550
 551void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
 552			       enum ocfs2_lock_type type,
 553			       unsigned int generation,
 554			       struct inode *inode)
 555{
 556	const struct ocfs2_lock_res_ops *ops;
 557
 558	switch(type) {
 559		case OCFS2_LOCK_TYPE_RW:
 560			ops = &ocfs2_inode_rw_lops;
 561			break;
 562		case OCFS2_LOCK_TYPE_META:
 563			ops = &ocfs2_inode_inode_lops;
 564			break;
 565		case OCFS2_LOCK_TYPE_OPEN:
 566			ops = &ocfs2_inode_open_lops;
 567			break;
 568		default:
 569			mlog_bug_on_msg(1, "type: %d\n", type);
 570			ops = NULL; /* thanks, gcc */
 571			break;
 572	}
 573
 574	ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
 575			      generation, res->l_name);
 576	ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
 577}
 578
 579static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
 580{
 581	struct inode *inode = ocfs2_lock_res_inode(lockres);
 582
 583	return OCFS2_SB(inode->i_sb);
 584}
 585
 586static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres)
 587{
 588	struct ocfs2_mem_dqinfo *info = lockres->l_priv;
 589
 590	return OCFS2_SB(info->dqi_gi.dqi_sb);
 591}
 592
 593static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres)
 594{
 595	struct ocfs2_file_private *fp = lockres->l_priv;
 596
 597	return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb);
 598}
 599
 600static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
 601{
 602	__be64 inode_blkno_be;
 603
 604	memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
 605	       sizeof(__be64));
 606
 607	return be64_to_cpu(inode_blkno_be);
 608}
 609
 610static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
 611{
 612	struct ocfs2_dentry_lock *dl = lockres->l_priv;
 613
 614	return OCFS2_SB(dl->dl_inode->i_sb);
 615}
 616
 617void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
 618				u64 parent, struct inode *inode)
 619{
 620	int len;
 621	u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
 622	__be64 inode_blkno_be = cpu_to_be64(inode_blkno);
 623	struct ocfs2_lock_res *lockres = &dl->dl_lockres;
 624
 625	ocfs2_lock_res_init_once(lockres);
 626
 627	/*
 628	 * Unfortunately, the standard lock naming scheme won't work
 629	 * here because we have two 16 byte values to use. Instead,
 630	 * we'll stuff the inode number as a binary value. We still
 631	 * want error prints to show something without garbling the
 632	 * display, so drop a null byte in there before the inode
 633	 * number. A future version of OCFS2 will likely use all
 634	 * binary lock names. The stringified names have been a
 635	 * tremendous aid in debugging, but now that the debugfs
 636	 * interface exists, we can mangle things there if need be.
 637	 *
 638	 * NOTE: We also drop the standard "pad" value (the total lock
 639	 * name size stays the same though - the last part is all
 640	 * zeros due to the memset in ocfs2_lock_res_init_once()
 641	 */
 642	len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
 643		       "%c%016llx",
 644		       ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
 645		       (long long)parent);
 646
 647	BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
 648
 649	memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
 650	       sizeof(__be64));
 651
 652	ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
 653				   OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
 654				   dl);
 655}
 656
 657static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
 658				      struct ocfs2_super *osb)
 659{
 660	/* Superblock lockres doesn't come from a slab so we call init
 661	 * once on it manually.  */
 662	ocfs2_lock_res_init_once(res);
 663	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
 664			      0, res->l_name);
 665	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
 666				   &ocfs2_super_lops, osb);
 667}
 668
 669static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
 670				       struct ocfs2_super *osb)
 671{
 672	/* Rename lockres doesn't come from a slab so we call init
 673	 * once on it manually.  */
 674	ocfs2_lock_res_init_once(res);
 675	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
 676	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
 677				   &ocfs2_rename_lops, osb);
 678}
 679
 680static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res,
 681					 struct ocfs2_super *osb)
 682{
 683	/* nfs_sync lockres doesn't come from a slab so we call init
 684	 * once on it manually.  */
 685	ocfs2_lock_res_init_once(res);
 686	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name);
 687	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC,
 688				   &ocfs2_nfs_sync_lops, osb);
 689}
 690
 691static void ocfs2_nfs_sync_lock_init(struct ocfs2_super *osb)
 692{
 693	ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb);
 694	init_rwsem(&osb->nfs_sync_rwlock);
 695}
 696
 697void ocfs2_trim_fs_lock_res_init(struct ocfs2_super *osb)
 698{
 699	struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
 700
 701	/* Only one trimfs thread are allowed to work at the same time. */
 702	mutex_lock(&osb->obs_trim_fs_mutex);
 703
 704	ocfs2_lock_res_init_once(lockres);
 705	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_TRIM_FS, 0, 0, lockres->l_name);
 706	ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_TRIM_FS,
 707				   &ocfs2_trim_fs_lops, osb);
 708}
 709
 710void ocfs2_trim_fs_lock_res_uninit(struct ocfs2_super *osb)
 711{
 712	struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
 713
 714	ocfs2_simple_drop_lockres(osb, lockres);
 715	ocfs2_lock_res_free(lockres);
 716
 717	mutex_unlock(&osb->obs_trim_fs_mutex);
 718}
 719
 720static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res,
 721					    struct ocfs2_super *osb)
 722{
 723	ocfs2_lock_res_init_once(res);
 724	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name);
 725	ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN,
 726				   &ocfs2_orphan_scan_lops, osb);
 727}
 728
 729void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres,
 730			      struct ocfs2_file_private *fp)
 731{
 732	struct inode *inode = fp->fp_file->f_mapping->host;
 733	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 734
 735	ocfs2_lock_res_init_once(lockres);
 736	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno,
 737			      inode->i_generation, lockres->l_name);
 738	ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
 739				   OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops,
 740				   fp);
 741	lockres->l_flags |= OCFS2_LOCK_NOCACHE;
 742}
 743
 744void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres,
 745			       struct ocfs2_mem_dqinfo *info)
 746{
 747	ocfs2_lock_res_init_once(lockres);
 748	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type,
 749			      0, lockres->l_name);
 750	ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres,
 751				   OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops,
 752				   info);
 753}
 754
 755void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres,
 756				  struct ocfs2_super *osb, u64 ref_blkno,
 757				  unsigned int generation)
 758{
 759	ocfs2_lock_res_init_once(lockres);
 760	ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno,
 761			      generation, lockres->l_name);
 762	ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT,
 763				   &ocfs2_refcount_block_lops, osb);
 764}
 765
 766void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
 767{
 768	if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
 769		return;
 770
 771	ocfs2_remove_lockres_tracking(res);
 772
 773	mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
 774			"Lockres %s is on the blocked list\n",
 775			res->l_name);
 776	mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
 777			"Lockres %s has mask waiters pending\n",
 778			res->l_name);
 779	mlog_bug_on_msg(spin_is_locked(&res->l_lock),
 780			"Lockres %s is locked\n",
 781			res->l_name);
 782	mlog_bug_on_msg(res->l_ro_holders,
 783			"Lockres %s has %u ro holders\n",
 784			res->l_name, res->l_ro_holders);
 785	mlog_bug_on_msg(res->l_ex_holders,
 786			"Lockres %s has %u ex holders\n",
 787			res->l_name, res->l_ex_holders);
 788
 789	/* Need to clear out the lock status block for the dlm */
 790	memset(&res->l_lksb, 0, sizeof(res->l_lksb));
 791
 792	res->l_flags = 0UL;
 793}
 794
 795/*
 796 * Keep a list of processes who have interest in a lockres.
 797 * Note: this is now only uesed for check recursive cluster locking.
 798 */
 799static inline void ocfs2_add_holder(struct ocfs2_lock_res *lockres,
 800				   struct ocfs2_lock_holder *oh)
 801{
 802	INIT_LIST_HEAD(&oh->oh_list);
 803	oh->oh_owner_pid = get_pid(task_pid(current));
 804
 805	spin_lock(&lockres->l_lock);
 806	list_add_tail(&oh->oh_list, &lockres->l_holders);
 807	spin_unlock(&lockres->l_lock);
 808}
 809
 810static struct ocfs2_lock_holder *
 811ocfs2_pid_holder(struct ocfs2_lock_res *lockres,
 812		struct pid *pid)
 813{
 814	struct ocfs2_lock_holder *oh;
 815
 816	spin_lock(&lockres->l_lock);
 817	list_for_each_entry(oh, &lockres->l_holders, oh_list) {
 818		if (oh->oh_owner_pid == pid) {
 819			spin_unlock(&lockres->l_lock);
 820			return oh;
 821		}
 822	}
 823	spin_unlock(&lockres->l_lock);
 824	return NULL;
 825}
 826
 827static inline void ocfs2_remove_holder(struct ocfs2_lock_res *lockres,
 828				       struct ocfs2_lock_holder *oh)
 829{
 830	spin_lock(&lockres->l_lock);
 831	list_del(&oh->oh_list);
 832	spin_unlock(&lockres->l_lock);
 833
 834	put_pid(oh->oh_owner_pid);
 835}
 836
 837
 838static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
 839				     int level)
 840{
 841	BUG_ON(!lockres);
 842
 843	switch(level) {
 844	case DLM_LOCK_EX:
 845		lockres->l_ex_holders++;
 846		break;
 847	case DLM_LOCK_PR:
 848		lockres->l_ro_holders++;
 849		break;
 850	default:
 851		BUG();
 852	}
 853}
 854
 855static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
 856				     int level)
 857{
 858	BUG_ON(!lockres);
 859
 860	switch(level) {
 861	case DLM_LOCK_EX:
 862		BUG_ON(!lockres->l_ex_holders);
 863		lockres->l_ex_holders--;
 864		break;
 865	case DLM_LOCK_PR:
 866		BUG_ON(!lockres->l_ro_holders);
 867		lockres->l_ro_holders--;
 868		break;
 869	default:
 870		BUG();
 871	}
 872}
 873
 874/* WARNING: This function lives in a world where the only three lock
 875 * levels are EX, PR, and NL. It *will* have to be adjusted when more
 876 * lock types are added. */
 877static inline int ocfs2_highest_compat_lock_level(int level)
 878{
 879	int new_level = DLM_LOCK_EX;
 880
 881	if (level == DLM_LOCK_EX)
 882		new_level = DLM_LOCK_NL;
 883	else if (level == DLM_LOCK_PR)
 884		new_level = DLM_LOCK_PR;
 885	return new_level;
 886}
 887
 888static void lockres_set_flags(struct ocfs2_lock_res *lockres,
 889			      unsigned long newflags)
 890{
 891	struct ocfs2_mask_waiter *mw, *tmp;
 892
 893 	assert_spin_locked(&lockres->l_lock);
 894
 895	lockres->l_flags = newflags;
 896
 897	list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) {
 898		if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
 899			continue;
 900
 901		list_del_init(&mw->mw_item);
 902		mw->mw_status = 0;
 903		complete(&mw->mw_complete);
 904		ocfs2_track_lock_wait(lockres);
 905	}
 906}
 907static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
 908{
 909	lockres_set_flags(lockres, lockres->l_flags | or);
 910}
 911static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
 912				unsigned long clear)
 913{
 914	lockres_set_flags(lockres, lockres->l_flags & ~clear);
 915}
 916
 917static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
 918{
 919	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
 920	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
 921	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
 922	BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
 923
 924	lockres->l_level = lockres->l_requested;
 925	if (lockres->l_level <=
 926	    ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
 927		lockres->l_blocking = DLM_LOCK_NL;
 928		lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
 929	}
 930	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
 931}
 932
 933static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
 934{
 935	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
 936	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
 937
 938	/* Convert from RO to EX doesn't really need anything as our
 939	 * information is already up to data. Convert from NL to
 940	 * *anything* however should mark ourselves as needing an
 941	 * update */
 942	if (lockres->l_level == DLM_LOCK_NL &&
 943	    lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
 944		lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
 945
 946	lockres->l_level = lockres->l_requested;
 947
 948	/*
 949	 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
 950	 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
 951	 * downconverting the lock before the upconvert has fully completed.
 952	 * Do not prevent the dc thread from downconverting if NONBLOCK lock
 953	 * had already returned.
 954	 */
 955	if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED))
 956		lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
 957	else
 958		lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED);
 959
 960	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
 961}
 962
 963static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
 964{
 965	BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY)));
 966	BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
 967
 968	if (lockres->l_requested > DLM_LOCK_NL &&
 969	    !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
 970	    lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
 971		lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
 972
 973	lockres->l_level = lockres->l_requested;
 974	lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
 975	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
 976}
 977
 978static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
 979				     int level)
 980{
 981	int needs_downconvert = 0;
 982
 983	assert_spin_locked(&lockres->l_lock);
 984
 985	if (level > lockres->l_blocking) {
 986		/* only schedule a downconvert if we haven't already scheduled
 987		 * one that goes low enough to satisfy the level we're
 988		 * blocking.  this also catches the case where we get
 989		 * duplicate BASTs */
 990		if (ocfs2_highest_compat_lock_level(level) <
 991		    ocfs2_highest_compat_lock_level(lockres->l_blocking))
 992			needs_downconvert = 1;
 993
 994		lockres->l_blocking = level;
 995	}
 996
 997	mlog(ML_BASTS, "lockres %s, block %d, level %d, l_block %d, dwn %d\n",
 998	     lockres->l_name, level, lockres->l_level, lockres->l_blocking,
 999	     needs_downconvert);
1000
1001	if (needs_downconvert)
1002		lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
1003	mlog(0, "needs_downconvert = %d\n", needs_downconvert);
1004	return needs_downconvert;
1005}
1006
1007/*
1008 * OCFS2_LOCK_PENDING and l_pending_gen.
1009 *
1010 * Why does OCFS2_LOCK_PENDING exist?  To close a race between setting
1011 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock().  See ocfs2_unblock_lock()
1012 * for more details on the race.
1013 *
1014 * OCFS2_LOCK_PENDING closes the race quite nicely.  However, it introduces
1015 * a race on itself.  In o2dlm, we can get the ast before ocfs2_dlm_lock()
1016 * returns.  The ast clears OCFS2_LOCK_BUSY, and must therefore clear
1017 * OCFS2_LOCK_PENDING at the same time.  When ocfs2_dlm_lock() returns,
1018 * the caller is going to try to clear PENDING again.  If nothing else is
1019 * happening, __lockres_clear_pending() sees PENDING is unset and does
1020 * nothing.
1021 *
1022 * But what if another path (eg downconvert thread) has just started a
1023 * new locking action?  The other path has re-set PENDING.  Our path
1024 * cannot clear PENDING, because that will re-open the original race
1025 * window.
1026 *
1027 * [Example]
1028 *
1029 * ocfs2_meta_lock()
1030 *  ocfs2_cluster_lock()
1031 *   set BUSY
1032 *   set PENDING
1033 *   drop l_lock
1034 *   ocfs2_dlm_lock()
1035 *    ocfs2_locking_ast()		ocfs2_downconvert_thread()
1036 *     clear PENDING			 ocfs2_unblock_lock()
1037 *					  take_l_lock
1038 *					  !BUSY
1039 *					  ocfs2_prepare_downconvert()
1040 *					   set BUSY
1041 *					   set PENDING
1042 *					  drop l_lock
1043 *   take l_lock
1044 *   clear PENDING
1045 *   drop l_lock
1046 *			<window>
1047 *					  ocfs2_dlm_lock()
1048 *
1049 * So as you can see, we now have a window where l_lock is not held,
1050 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
1051 *
1052 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
1053 * set by ocfs2_prepare_downconvert().  That wasn't nice.
1054 *
1055 * To solve this we introduce l_pending_gen.  A call to
1056 * lockres_clear_pending() will only do so when it is passed a generation
1057 * number that matches the lockres.  lockres_set_pending() will return the
1058 * current generation number.  When ocfs2_cluster_lock() goes to clear
1059 * PENDING, it passes the generation it got from set_pending().  In our
1060 * example above, the generation numbers will *not* match.  Thus,
1061 * ocfs2_cluster_lock() will not clear the PENDING set by
1062 * ocfs2_prepare_downconvert().
1063 */
1064
1065/* Unlocked version for ocfs2_locking_ast() */
1066static void __lockres_clear_pending(struct ocfs2_lock_res *lockres,
1067				    unsigned int generation,
1068				    struct ocfs2_super *osb)
1069{
1070	assert_spin_locked(&lockres->l_lock);
1071
1072	/*
1073	 * The ast and locking functions can race us here.  The winner
1074	 * will clear pending, the loser will not.
1075	 */
1076	if (!(lockres->l_flags & OCFS2_LOCK_PENDING) ||
1077	    (lockres->l_pending_gen != generation))
1078		return;
1079
1080	lockres_clear_flags(lockres, OCFS2_LOCK_PENDING);
1081	lockres->l_pending_gen++;
1082
1083	/*
1084	 * The downconvert thread may have skipped us because we
1085	 * were PENDING.  Wake it up.
1086	 */
1087	if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1088		ocfs2_wake_downconvert_thread(osb);
1089}
1090
1091/* Locked version for callers of ocfs2_dlm_lock() */
1092static void lockres_clear_pending(struct ocfs2_lock_res *lockres,
1093				  unsigned int generation,
1094				  struct ocfs2_super *osb)
1095{
1096	unsigned long flags;
1097
1098	spin_lock_irqsave(&lockres->l_lock, flags);
1099	__lockres_clear_pending(lockres, generation, osb);
1100	spin_unlock_irqrestore(&lockres->l_lock, flags);
1101}
1102
1103static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres)
1104{
1105	assert_spin_locked(&lockres->l_lock);
1106	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
1107
1108	lockres_or_flags(lockres, OCFS2_LOCK_PENDING);
1109
1110	return lockres->l_pending_gen;
1111}
1112
1113static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level)
1114{
1115	struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1116	struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1117	int needs_downconvert;
1118	unsigned long flags;
1119
1120	BUG_ON(level <= DLM_LOCK_NL);
1121
1122	mlog(ML_BASTS, "BAST fired for lockres %s, blocking %d, level %d, "
1123	     "type %s\n", lockres->l_name, level, lockres->l_level,
1124	     ocfs2_lock_type_string(lockres->l_type));
1125
1126	/*
1127	 * We can skip the bast for locks which don't enable caching -
1128	 * they'll be dropped at the earliest possible time anyway.
1129	 */
1130	if (lockres->l_flags & OCFS2_LOCK_NOCACHE)
1131		return;
1132
1133	spin_lock_irqsave(&lockres->l_lock, flags);
1134	needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
1135	if (needs_downconvert)
1136		ocfs2_schedule_blocked_lock(osb, lockres);
1137	spin_unlock_irqrestore(&lockres->l_lock, flags);
1138
1139	wake_up(&lockres->l_event);
1140
1141	ocfs2_wake_downconvert_thread(osb);
1142}
1143
1144static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb)
1145{
1146	struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1147	struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1148	unsigned long flags;
1149	int status;
1150
1151	spin_lock_irqsave(&lockres->l_lock, flags);
1152
1153	status = ocfs2_dlm_lock_status(&lockres->l_lksb);
1154
1155	if (status == -EAGAIN) {
1156		lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1157		goto out;
1158	}
1159
1160	if (status) {
1161		mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n",
1162		     lockres->l_name, status);
1163		spin_unlock_irqrestore(&lockres->l_lock, flags);
1164		return;
1165	}
1166
1167	mlog(ML_BASTS, "AST fired for lockres %s, action %d, unlock %d, "
1168	     "level %d => %d\n", lockres->l_name, lockres->l_action,
1169	     lockres->l_unlock_action, lockres->l_level, lockres->l_requested);
1170
1171	switch(lockres->l_action) {
1172	case OCFS2_AST_ATTACH:
1173		ocfs2_generic_handle_attach_action(lockres);
1174		lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
1175		break;
1176	case OCFS2_AST_CONVERT:
1177		ocfs2_generic_handle_convert_action(lockres);
1178		break;
1179	case OCFS2_AST_DOWNCONVERT:
1180		ocfs2_generic_handle_downconvert_action(lockres);
1181		break;
1182	default:
1183		mlog(ML_ERROR, "lockres %s: AST fired with invalid action: %u, "
1184		     "flags 0x%lx, unlock: %u\n",
1185		     lockres->l_name, lockres->l_action, lockres->l_flags,
1186		     lockres->l_unlock_action);
1187		BUG();
1188	}
1189out:
1190	/* set it to something invalid so if we get called again we
1191	 * can catch it. */
1192	lockres->l_action = OCFS2_AST_INVALID;
1193
1194	/* Did we try to cancel this lock?  Clear that state */
1195	if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT)
1196		lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1197
1198	/*
1199	 * We may have beaten the locking functions here.  We certainly
1200	 * know that dlm_lock() has been called :-)
1201	 * Because we can't have two lock calls in flight at once, we
1202	 * can use lockres->l_pending_gen.
1203	 */
1204	__lockres_clear_pending(lockres, lockres->l_pending_gen,  osb);
1205
1206	wake_up(&lockres->l_event);
1207	spin_unlock_irqrestore(&lockres->l_lock, flags);
1208}
1209
1210static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error)
1211{
1212	struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1213	unsigned long flags;
1214
1215	mlog(ML_BASTS, "UNLOCK AST fired for lockres %s, action = %d\n",
1216	     lockres->l_name, lockres->l_unlock_action);
1217
1218	spin_lock_irqsave(&lockres->l_lock, flags);
1219	if (error) {
1220		mlog(ML_ERROR, "Dlm passes error %d for lock %s, "
1221		     "unlock_action %d\n", error, lockres->l_name,
1222		     lockres->l_unlock_action);
1223		spin_unlock_irqrestore(&lockres->l_lock, flags);
1224		return;
1225	}
1226
1227	switch(lockres->l_unlock_action) {
1228	case OCFS2_UNLOCK_CANCEL_CONVERT:
1229		mlog(0, "Cancel convert success for %s\n", lockres->l_name);
1230		lockres->l_action = OCFS2_AST_INVALID;
1231		/* Downconvert thread may have requeued this lock, we
1232		 * need to wake it. */
1233		if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1234			ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres));
1235		break;
1236	case OCFS2_UNLOCK_DROP_LOCK:
1237		lockres->l_level = DLM_LOCK_IV;
1238		break;
1239	default:
1240		BUG();
1241	}
1242
1243	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1244	lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1245	wake_up(&lockres->l_event);
1246	spin_unlock_irqrestore(&lockres->l_lock, flags);
1247}
1248
1249/*
1250 * This is the filesystem locking protocol.  It provides the lock handling
1251 * hooks for the underlying DLM.  It has a maximum version number.
1252 * The version number allows interoperability with systems running at
1253 * the same major number and an equal or smaller minor number.
1254 *
1255 * Whenever the filesystem does new things with locks (adds or removes a
1256 * lock, orders them differently, does different things underneath a lock),
1257 * the version must be changed.  The protocol is negotiated when joining
1258 * the dlm domain.  A node may join the domain if its major version is
1259 * identical to all other nodes and its minor version is greater than
1260 * or equal to all other nodes.  When its minor version is greater than
1261 * the other nodes, it will run at the minor version specified by the
1262 * other nodes.
1263 *
1264 * If a locking change is made that will not be compatible with older
1265 * versions, the major number must be increased and the minor version set
1266 * to zero.  If a change merely adds a behavior that can be disabled when
1267 * speaking to older versions, the minor version must be increased.  If a
1268 * change adds a fully backwards compatible change (eg, LVB changes that
1269 * are just ignored by older versions), the version does not need to be
1270 * updated.
1271 */
1272static struct ocfs2_locking_protocol lproto = {
1273	.lp_max_version = {
1274		.pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
1275		.pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
1276	},
1277	.lp_lock_ast		= ocfs2_locking_ast,
1278	.lp_blocking_ast	= ocfs2_blocking_ast,
1279	.lp_unlock_ast		= ocfs2_unlock_ast,
1280};
1281
1282void ocfs2_set_locking_protocol(void)
1283{
1284	ocfs2_stack_glue_set_max_proto_version(&lproto.lp_max_version);
1285}
1286
1287static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
1288						int convert)
1289{
1290	unsigned long flags;
1291
1292	spin_lock_irqsave(&lockres->l_lock, flags);
1293	lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1294	lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1295	if (convert)
1296		lockres->l_action = OCFS2_AST_INVALID;
1297	else
1298		lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1299	spin_unlock_irqrestore(&lockres->l_lock, flags);
1300
1301	wake_up(&lockres->l_event);
1302}
1303
1304/* Note: If we detect another process working on the lock (i.e.,
1305 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1306 * to do the right thing in that case.
1307 */
1308static int ocfs2_lock_create(struct ocfs2_super *osb,
1309			     struct ocfs2_lock_res *lockres,
1310			     int level,
1311			     u32 dlm_flags)
1312{
1313	int ret = 0;
1314	unsigned long flags;
1315	unsigned int gen;
1316
1317	mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level,
1318	     dlm_flags);
1319
1320	spin_lock_irqsave(&lockres->l_lock, flags);
1321	if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
1322	    (lockres->l_flags & OCFS2_LOCK_BUSY)) {
1323		spin_unlock_irqrestore(&lockres->l_lock, flags);
1324		goto bail;
1325	}
1326
1327	lockres->l_action = OCFS2_AST_ATTACH;
1328	lockres->l_requested = level;
1329	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1330	gen = lockres_set_pending(lockres);
1331	spin_unlock_irqrestore(&lockres->l_lock, flags);
1332
1333	ret = ocfs2_dlm_lock(osb->cconn,
1334			     level,
1335			     &lockres->l_lksb,
1336			     dlm_flags,
1337			     lockres->l_name,
1338			     OCFS2_LOCK_ID_MAX_LEN - 1);
1339	lockres_clear_pending(lockres, gen, osb);
1340	if (ret) {
1341		ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1342		ocfs2_recover_from_dlm_error(lockres, 1);
1343	}
1344
1345	mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name);
1346
1347bail:
1348	return ret;
1349}
1350
1351static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
1352					int flag)
1353{
1354	unsigned long flags;
1355	int ret;
1356
1357	spin_lock_irqsave(&lockres->l_lock, flags);
1358	ret = lockres->l_flags & flag;
1359	spin_unlock_irqrestore(&lockres->l_lock, flags);
1360
1361	return ret;
1362}
1363
1364static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
1365
1366{
1367	wait_event(lockres->l_event,
1368		   !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
1369}
1370
1371static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
1372
1373{
1374	wait_event(lockres->l_event,
1375		   !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
1376}
1377
1378/* predict what lock level we'll be dropping down to on behalf
1379 * of another node, and return true if the currently wanted
1380 * level will be compatible with it. */
1381static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
1382						     int wanted)
1383{
1384	BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
1385
1386	return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
1387}
1388
1389static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
1390{
1391	INIT_LIST_HEAD(&mw->mw_item);
1392	init_completion(&mw->mw_complete);
1393	ocfs2_init_start_time(mw);
1394}
1395
1396static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
1397{
1398	wait_for_completion(&mw->mw_complete);
1399	/* Re-arm the completion in case we want to wait on it again */
1400	reinit_completion(&mw->mw_complete);
1401	return mw->mw_status;
1402}
1403
1404static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
1405				    struct ocfs2_mask_waiter *mw,
1406				    unsigned long mask,
1407				    unsigned long goal)
1408{
1409	BUG_ON(!list_empty(&mw->mw_item));
1410
1411	assert_spin_locked(&lockres->l_lock);
1412
1413	list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
1414	mw->mw_mask = mask;
1415	mw->mw_goal = goal;
1416	ocfs2_track_lock_wait(lockres);
1417}
1418
1419/* returns 0 if the mw that was removed was already satisfied, -EBUSY
1420 * if the mask still hadn't reached its goal */
1421static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1422				      struct ocfs2_mask_waiter *mw)
1423{
1424	int ret = 0;
1425
1426	assert_spin_locked(&lockres->l_lock);
1427	if (!list_empty(&mw->mw_item)) {
1428		if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
1429			ret = -EBUSY;
1430
1431		list_del_init(&mw->mw_item);
1432		init_completion(&mw->mw_complete);
1433		ocfs2_track_lock_wait(lockres);
1434	}
1435
1436	return ret;
1437}
1438
1439static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1440				      struct ocfs2_mask_waiter *mw)
1441{
1442	unsigned long flags;
1443	int ret = 0;
1444
1445	spin_lock_irqsave(&lockres->l_lock, flags);
1446	ret = __lockres_remove_mask_waiter(lockres, mw);
1447	spin_unlock_irqrestore(&lockres->l_lock, flags);
1448
1449	return ret;
1450
1451}
1452
1453static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw,
1454					     struct ocfs2_lock_res *lockres)
1455{
1456	int ret;
1457
1458	ret = wait_for_completion_interruptible(&mw->mw_complete);
1459	if (ret)
1460		lockres_remove_mask_waiter(lockres, mw);
1461	else
1462		ret = mw->mw_status;
1463	/* Re-arm the completion in case we want to wait on it again */
1464	reinit_completion(&mw->mw_complete);
1465	return ret;
1466}
1467
1468static int __ocfs2_cluster_lock(struct ocfs2_super *osb,
1469				struct ocfs2_lock_res *lockres,
1470				int level,
1471				u32 lkm_flags,
1472				int arg_flags,
1473				int l_subclass,
1474				unsigned long caller_ip)
1475{
1476	struct ocfs2_mask_waiter mw;
1477	int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
1478	int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1479	unsigned long flags;
1480	unsigned int gen;
1481	int noqueue_attempted = 0;
1482	int dlm_locked = 0;
1483	int kick_dc = 0;
1484
1485	if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED)) {
1486		mlog_errno(-EINVAL);
1487		return -EINVAL;
1488	}
1489
1490	ocfs2_init_mask_waiter(&mw);
1491
1492	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
1493		lkm_flags |= DLM_LKF_VALBLK;
1494
1495again:
1496	wait = 0;
1497
1498	spin_lock_irqsave(&lockres->l_lock, flags);
1499
1500	if (catch_signals && signal_pending(current)) {
1501		ret = -ERESTARTSYS;
1502		goto unlock;
1503	}
1504
1505	mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
1506			"Cluster lock called on freeing lockres %s! flags "
1507			"0x%lx\n", lockres->l_name, lockres->l_flags);
1508
1509	/* We only compare against the currently granted level
1510	 * here. If the lock is blocked waiting on a downconvert,
1511	 * we'll get caught below. */
1512	if (lockres->l_flags & OCFS2_LOCK_BUSY &&
1513	    level > lockres->l_level) {
1514		/* is someone sitting in dlm_lock? If so, wait on
1515		 * them. */
1516		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1517		wait = 1;
1518		goto unlock;
1519	}
1520
1521	if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
1522		/*
1523		 * We've upconverted. If the lock now has a level we can
1524		 * work with, we take it. If, however, the lock is not at the
1525		 * required level, we go thru the full cycle. One way this could
1526		 * happen is if a process requesting an upconvert to PR is
1527		 * closely followed by another requesting upconvert to an EX.
1528		 * If the process requesting EX lands here, we want it to
1529		 * continue attempting to upconvert and let the process
1530		 * requesting PR take the lock.
1531		 * If multiple processes request upconvert to PR, the first one
1532		 * here will take the lock. The others will have to go thru the
1533		 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1534		 * downconvert request.
1535		 */
1536		if (level <= lockres->l_level)
1537			goto update_holders;
1538	}
1539
1540	if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
1541	    !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
1542		/* is the lock is currently blocked on behalf of
1543		 * another node */
1544		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
1545		wait = 1;
1546		goto unlock;
1547	}
1548
1549	if (level > lockres->l_level) {
1550		if (noqueue_attempted > 0) {
1551			ret = -EAGAIN;
1552			goto unlock;
1553		}
1554		if (lkm_flags & DLM_LKF_NOQUEUE)
1555			noqueue_attempted = 1;
1556
1557		if (lockres->l_action != OCFS2_AST_INVALID)
1558			mlog(ML_ERROR, "lockres %s has action %u pending\n",
1559			     lockres->l_name, lockres->l_action);
1560
1561		if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1562			lockres->l_action = OCFS2_AST_ATTACH;
1563			lkm_flags &= ~DLM_LKF_CONVERT;
1564		} else {
1565			lockres->l_action = OCFS2_AST_CONVERT;
1566			lkm_flags |= DLM_LKF_CONVERT;
1567		}
1568
1569		lockres->l_requested = level;
1570		lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1571		gen = lockres_set_pending(lockres);
1572		spin_unlock_irqrestore(&lockres->l_lock, flags);
1573
1574		BUG_ON(level == DLM_LOCK_IV);
1575		BUG_ON(level == DLM_LOCK_NL);
1576
1577		mlog(ML_BASTS, "lockres %s, convert from %d to %d\n",
1578		     lockres->l_name, lockres->l_level, level);
1579
1580		/* call dlm_lock to upgrade lock now */
1581		ret = ocfs2_dlm_lock(osb->cconn,
1582				     level,
1583				     &lockres->l_lksb,
1584				     lkm_flags,
1585				     lockres->l_name,
1586				     OCFS2_LOCK_ID_MAX_LEN - 1);
1587		lockres_clear_pending(lockres, gen, osb);
1588		if (ret) {
1589			if (!(lkm_flags & DLM_LKF_NOQUEUE) ||
1590			    (ret != -EAGAIN)) {
1591				ocfs2_log_dlm_error("ocfs2_dlm_lock",
1592						    ret, lockres);
1593			}
1594			ocfs2_recover_from_dlm_error(lockres, 1);
1595			goto out;
1596		}
1597		dlm_locked = 1;
1598
1599		mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1600		     lockres->l_name);
1601
1602		/* At this point we've gone inside the dlm and need to
1603		 * complete our work regardless. */
1604		catch_signals = 0;
1605
1606		/* wait for busy to clear and carry on */
1607		goto again;
1608	}
1609
1610update_holders:
1611	/* Ok, if we get here then we're good to go. */
1612	ocfs2_inc_holders(lockres, level);
1613
1614	ret = 0;
1615unlock:
1616	lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1617
1618	/* ocfs2_unblock_lock request on seeing OCFS2_LOCK_UPCONVERT_FINISHING */
1619	kick_dc = (lockres->l_flags & OCFS2_LOCK_BLOCKED);
1620
1621	spin_unlock_irqrestore(&lockres->l_lock, flags);
1622	if (kick_dc)
1623		ocfs2_wake_downconvert_thread(osb);
1624out:
1625	/*
1626	 * This is helping work around a lock inversion between the page lock
1627	 * and dlm locks.  One path holds the page lock while calling aops
1628	 * which block acquiring dlm locks.  The voting thread holds dlm
1629	 * locks while acquiring page locks while down converting data locks.
1630	 * This block is helping an aop path notice the inversion and back
1631	 * off to unlock its page lock before trying the dlm lock again.
1632	 */
1633	if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1634	    mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1635		wait = 0;
1636		spin_lock_irqsave(&lockres->l_lock, flags);
1637		if (__lockres_remove_mask_waiter(lockres, &mw)) {
1638			if (dlm_locked)
1639				lockres_or_flags(lockres,
1640					OCFS2_LOCK_NONBLOCK_FINISHED);
1641			spin_unlock_irqrestore(&lockres->l_lock, flags);
1642			ret = -EAGAIN;
1643		} else {
1644			spin_unlock_irqrestore(&lockres->l_lock, flags);
1645			goto again;
1646		}
1647	}
1648	if (wait) {
1649		ret = ocfs2_wait_for_mask(&mw);
1650		if (ret == 0)
1651			goto again;
1652		mlog_errno(ret);
1653	}
1654	ocfs2_update_lock_stats(lockres, level, &mw, ret);
1655
1656#ifdef CONFIG_DEBUG_LOCK_ALLOC
1657	if (!ret && lockres->l_lockdep_map.key != NULL) {
1658		if (level == DLM_LOCK_PR)
1659			rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass,
1660				!!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1661				caller_ip);
1662		else
1663			rwsem_acquire(&lockres->l_lockdep_map, l_subclass,
1664				!!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1665				caller_ip);
1666	}
1667#endif
1668	return ret;
1669}
1670
1671static inline int ocfs2_cluster_lock(struct ocfs2_super *osb,
1672				     struct ocfs2_lock_res *lockres,
1673				     int level,
1674				     u32 lkm_flags,
1675				     int arg_flags)
1676{
1677	return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags,
1678				    0, _RET_IP_);
1679}
1680
1681
1682static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
1683				   struct ocfs2_lock_res *lockres,
1684				   int level,
1685				   unsigned long caller_ip)
1686{
1687	unsigned long flags;
1688
1689	spin_lock_irqsave(&lockres->l_lock, flags);
1690	ocfs2_dec_holders(lockres, level);
1691	ocfs2_downconvert_on_unlock(osb, lockres);
1692	spin_unlock_irqrestore(&lockres->l_lock, flags);
1693#ifdef CONFIG_DEBUG_LOCK_ALLOC
1694	if (lockres->l_lockdep_map.key != NULL)
1695		rwsem_release(&lockres->l_lockdep_map, caller_ip);
1696#endif
1697}
1698
1699static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1700				 struct ocfs2_lock_res *lockres,
1701				 int ex,
1702				 int local)
1703{
1704	int level =  ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1705	unsigned long flags;
1706	u32 lkm_flags = local ? DLM_LKF_LOCAL : 0;
1707
1708	spin_lock_irqsave(&lockres->l_lock, flags);
1709	BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1710	lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1711	spin_unlock_irqrestore(&lockres->l_lock, flags);
1712
1713	return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1714}
1715
1716/* Grants us an EX lock on the data and metadata resources, skipping
1717 * the normal cluster directory lookup. Use this ONLY on newly created
1718 * inodes which other nodes can't possibly see, and which haven't been
1719 * hashed in the inode hash yet. This can give us a good performance
1720 * increase as it'll skip the network broadcast normally associated
1721 * with creating a new lock resource. */
1722int ocfs2_create_new_inode_locks(struct inode *inode)
1723{
1724	int ret;
1725	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1726
1727	BUG_ON(!ocfs2_inode_is_new(inode));
1728
1729	mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1730
1731	/* NOTE: That we don't increment any of the holder counts, nor
1732	 * do we add anything to a journal handle. Since this is
1733	 * supposed to be a new inode which the cluster doesn't know
1734	 * about yet, there is no need to.  As far as the LVB handling
1735	 * is concerned, this is basically like acquiring an EX lock
1736	 * on a resource which has an invalid one -- we'll set it
1737	 * valid when we release the EX. */
1738
1739	ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1740	if (ret) {
1741		mlog_errno(ret);
1742		goto bail;
1743	}
1744
1745	/*
1746	 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1747	 * don't use a generation in their lock names.
1748	 */
1749	ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0);
1750	if (ret) {
1751		mlog_errno(ret);
1752		goto bail;
1753	}
1754
1755	ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1756	if (ret)
1757		mlog_errno(ret);
1758
1759bail:
1760	return ret;
1761}
1762
1763int ocfs2_rw_lock(struct inode *inode, int write)
1764{
1765	int status, level;
1766	struct ocfs2_lock_res *lockres;
1767	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1768
1769	mlog(0, "inode %llu take %s RW lock\n",
1770	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1771	     write ? "EXMODE" : "PRMODE");
1772
1773	if (ocfs2_mount_local(osb))
1774		return 0;
1775
1776	lockres = &OCFS2_I(inode)->ip_rw_lockres;
1777
1778	level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1779
1780	status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
 
1781	if (status < 0)
1782		mlog_errno(status);
1783
1784	return status;
1785}
1786
1787int ocfs2_try_rw_lock(struct inode *inode, int write)
1788{
1789	int status, level;
1790	struct ocfs2_lock_res *lockres;
1791	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1792
1793	mlog(0, "inode %llu try to take %s RW lock\n",
1794	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1795	     write ? "EXMODE" : "PRMODE");
1796
1797	if (ocfs2_mount_local(osb))
1798		return 0;
1799
1800	lockres = &OCFS2_I(inode)->ip_rw_lockres;
1801
1802	level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1803
1804	status = ocfs2_cluster_lock(osb, lockres, level, DLM_LKF_NOQUEUE, 0);
1805	return status;
1806}
1807
1808void ocfs2_rw_unlock(struct inode *inode, int write)
1809{
1810	int level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1811	struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1812	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1813
1814	mlog(0, "inode %llu drop %s RW lock\n",
1815	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1816	     write ? "EXMODE" : "PRMODE");
1817
1818	if (!ocfs2_mount_local(osb))
1819		ocfs2_cluster_unlock(osb, lockres, level);
1820}
1821
1822/*
1823 * ocfs2_open_lock always get PR mode lock.
1824 */
1825int ocfs2_open_lock(struct inode *inode)
1826{
1827	int status = 0;
1828	struct ocfs2_lock_res *lockres;
1829	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1830
1831	mlog(0, "inode %llu take PRMODE open lock\n",
1832	     (unsigned long long)OCFS2_I(inode)->ip_blkno);
1833
1834	if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1835		goto out;
1836
1837	lockres = &OCFS2_I(inode)->ip_open_lockres;
1838
1839	status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_PR, 0, 0);
 
1840	if (status < 0)
1841		mlog_errno(status);
1842
1843out:
1844	return status;
1845}
1846
1847int ocfs2_try_open_lock(struct inode *inode, int write)
1848{
1849	int status = 0, level;
1850	struct ocfs2_lock_res *lockres;
1851	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1852
1853	mlog(0, "inode %llu try to take %s open lock\n",
1854	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1855	     write ? "EXMODE" : "PRMODE");
1856
1857	if (ocfs2_is_hard_readonly(osb)) {
1858		if (write)
1859			status = -EROFS;
1860		goto out;
1861	}
1862
1863	if (ocfs2_mount_local(osb))
1864		goto out;
1865
1866	lockres = &OCFS2_I(inode)->ip_open_lockres;
1867
1868	level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1869
1870	/*
1871	 * The file system may already holding a PRMODE/EXMODE open lock.
1872	 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1873	 * other nodes and the -EAGAIN will indicate to the caller that
1874	 * this inode is still in use.
1875	 */
1876	status = ocfs2_cluster_lock(osb, lockres, level, DLM_LKF_NOQUEUE, 0);
 
1877
1878out:
1879	return status;
1880}
1881
1882/*
1883 * ocfs2_open_unlock unlock PR and EX mode open locks.
1884 */
1885void ocfs2_open_unlock(struct inode *inode)
1886{
1887	struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1888	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1889
1890	mlog(0, "inode %llu drop open lock\n",
1891	     (unsigned long long)OCFS2_I(inode)->ip_blkno);
1892
1893	if (ocfs2_mount_local(osb))
1894		goto out;
1895
1896	if(lockres->l_ro_holders)
1897		ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_PR);
 
1898	if(lockres->l_ex_holders)
1899		ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
 
1900
1901out:
1902	return;
1903}
1904
1905static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres,
1906				     int level)
1907{
1908	int ret;
1909	struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1910	unsigned long flags;
1911	struct ocfs2_mask_waiter mw;
1912
1913	ocfs2_init_mask_waiter(&mw);
1914
1915retry_cancel:
1916	spin_lock_irqsave(&lockres->l_lock, flags);
1917	if (lockres->l_flags & OCFS2_LOCK_BUSY) {
1918		ret = ocfs2_prepare_cancel_convert(osb, lockres);
1919		if (ret) {
1920			spin_unlock_irqrestore(&lockres->l_lock, flags);
1921			ret = ocfs2_cancel_convert(osb, lockres);
1922			if (ret < 0) {
1923				mlog_errno(ret);
1924				goto out;
1925			}
1926			goto retry_cancel;
1927		}
1928		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1929		spin_unlock_irqrestore(&lockres->l_lock, flags);
1930
1931		ocfs2_wait_for_mask(&mw);
1932		goto retry_cancel;
1933	}
1934
1935	ret = -ERESTARTSYS;
1936	/*
1937	 * We may still have gotten the lock, in which case there's no
1938	 * point to restarting the syscall.
1939	 */
1940	if (lockres->l_level == level)
1941		ret = 0;
1942
1943	mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret,
1944	     lockres->l_flags, lockres->l_level, lockres->l_action);
1945
1946	spin_unlock_irqrestore(&lockres->l_lock, flags);
1947
1948out:
1949	return ret;
1950}
1951
1952/*
1953 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1954 * flock() calls. The locking approach this requires is sufficiently
1955 * different from all other cluster lock types that we implement a
1956 * separate path to the "low-level" dlm calls. In particular:
1957 *
1958 * - No optimization of lock levels is done - we take at exactly
1959 *   what's been requested.
1960 *
1961 * - No lock caching is employed. We immediately downconvert to
1962 *   no-lock at unlock time. This also means flock locks never go on
1963 *   the blocking list).
1964 *
1965 * - Since userspace can trivially deadlock itself with flock, we make
1966 *   sure to allow cancellation of a misbehaving applications flock()
1967 *   request.
1968 *
1969 * - Access to any flock lockres doesn't require concurrency, so we
1970 *   can simplify the code by requiring the caller to guarantee
1971 *   serialization of dlmglue flock calls.
1972 */
1973int ocfs2_file_lock(struct file *file, int ex, int trylock)
1974{
1975	int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1976	unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
1977	unsigned long flags;
1978	struct ocfs2_file_private *fp = file->private_data;
1979	struct ocfs2_lock_res *lockres = &fp->fp_flock;
1980	struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1981	struct ocfs2_mask_waiter mw;
1982
1983	ocfs2_init_mask_waiter(&mw);
1984
1985	if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
1986	    (lockres->l_level > DLM_LOCK_NL)) {
1987		mlog(ML_ERROR,
1988		     "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1989		     "level: %u\n", lockres->l_name, lockres->l_flags,
1990		     lockres->l_level);
1991		return -EINVAL;
1992	}
1993
1994	spin_lock_irqsave(&lockres->l_lock, flags);
1995	if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1996		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1997		spin_unlock_irqrestore(&lockres->l_lock, flags);
1998
1999		/*
2000		 * Get the lock at NLMODE to start - that way we
2001		 * can cancel the upconvert request if need be.
2002		 */
2003		ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
2004		if (ret < 0) {
2005			mlog_errno(ret);
2006			goto out;
2007		}
2008
2009		ret = ocfs2_wait_for_mask(&mw);
2010		if (ret) {
2011			mlog_errno(ret);
2012			goto out;
2013		}
2014		spin_lock_irqsave(&lockres->l_lock, flags);
2015	}
2016
2017	lockres->l_action = OCFS2_AST_CONVERT;
2018	lkm_flags |= DLM_LKF_CONVERT;
2019	lockres->l_requested = level;
2020	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
2021
2022	lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
2023	spin_unlock_irqrestore(&lockres->l_lock, flags);
2024
2025	ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
2026			     lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1);
2027	if (ret) {
2028		if (!trylock || (ret != -EAGAIN)) {
2029			ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
2030			ret = -EINVAL;
2031		}
2032
2033		ocfs2_recover_from_dlm_error(lockres, 1);
2034		lockres_remove_mask_waiter(lockres, &mw);
2035		goto out;
2036	}
2037
2038	ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
2039	if (ret == -ERESTARTSYS) {
2040		/*
2041		 * Userspace can cause deadlock itself with
2042		 * flock(). Current behavior locally is to allow the
2043		 * deadlock, but abort the system call if a signal is
2044		 * received. We follow this example, otherwise a
2045		 * poorly written program could sit in kernel until
2046		 * reboot.
2047		 *
2048		 * Handling this is a bit more complicated for Ocfs2
2049		 * though. We can't exit this function with an
2050		 * outstanding lock request, so a cancel convert is
2051		 * required. We intentionally overwrite 'ret' - if the
2052		 * cancel fails and the lock was granted, it's easier
2053		 * to just bubble success back up to the user.
2054		 */
2055		ret = ocfs2_flock_handle_signal(lockres, level);
2056	} else if (!ret && (level > lockres->l_level)) {
2057		/* Trylock failed asynchronously */
2058		BUG_ON(!trylock);
2059		ret = -EAGAIN;
2060	}
2061
2062out:
2063
2064	mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
2065	     lockres->l_name, ex, trylock, ret);
2066	return ret;
2067}
2068
2069void ocfs2_file_unlock(struct file *file)
2070{
2071	int ret;
2072	unsigned int gen;
2073	unsigned long flags;
2074	struct ocfs2_file_private *fp = file->private_data;
2075	struct ocfs2_lock_res *lockres = &fp->fp_flock;
2076	struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
2077	struct ocfs2_mask_waiter mw;
2078
2079	ocfs2_init_mask_waiter(&mw);
2080
2081	if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
2082		return;
2083
2084	if (lockres->l_level == DLM_LOCK_NL)
2085		return;
2086
2087	mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
2088	     lockres->l_name, lockres->l_flags, lockres->l_level,
2089	     lockres->l_action);
2090
2091	spin_lock_irqsave(&lockres->l_lock, flags);
2092	/*
2093	 * Fake a blocking ast for the downconvert code.
2094	 */
2095	lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
2096	lockres->l_blocking = DLM_LOCK_EX;
2097
2098	gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
2099	lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
2100	spin_unlock_irqrestore(&lockres->l_lock, flags);
2101
2102	ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
2103	if (ret) {
2104		mlog_errno(ret);
2105		return;
2106	}
2107
2108	ret = ocfs2_wait_for_mask(&mw);
2109	if (ret)
2110		mlog_errno(ret);
2111}
2112
2113static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
2114					struct ocfs2_lock_res *lockres)
2115{
2116	int kick = 0;
2117
2118	/* If we know that another node is waiting on our lock, kick
2119	 * the downconvert thread * pre-emptively when we reach a release
2120	 * condition. */
2121	if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
2122		switch(lockres->l_blocking) {
2123		case DLM_LOCK_EX:
2124			if (!lockres->l_ex_holders && !lockres->l_ro_holders)
2125				kick = 1;
2126			break;
2127		case DLM_LOCK_PR:
2128			if (!lockres->l_ex_holders)
2129				kick = 1;
2130			break;
2131		default:
2132			BUG();
2133		}
2134	}
2135
2136	if (kick)
2137		ocfs2_wake_downconvert_thread(osb);
2138}
2139
2140#define OCFS2_SEC_BITS   34
2141#define OCFS2_SEC_SHIFT  (64 - OCFS2_SEC_BITS)
2142#define OCFS2_NSEC_MASK  ((1ULL << OCFS2_SEC_SHIFT) - 1)
2143
2144/* LVB only has room for 64 bits of time here so we pack it for
2145 * now. */
2146static u64 ocfs2_pack_timespec(struct timespec64 *spec)
2147{
2148	u64 res;
2149	u64 sec = clamp_t(time64_t, spec->tv_sec, 0, 0x3ffffffffull);
2150	u32 nsec = spec->tv_nsec;
2151
2152	res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
2153
2154	return res;
2155}
2156
2157/* Call this with the lockres locked. I am reasonably sure we don't
2158 * need ip_lock in this function as anyone who would be changing those
2159 * values is supposed to be blocked in ocfs2_inode_lock right now. */
2160static void __ocfs2_stuff_meta_lvb(struct inode *inode)
2161{
2162	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2163	struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2164	struct ocfs2_meta_lvb *lvb;
2165	struct timespec64 ts;
2166
2167	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2168
2169	/*
2170	 * Invalidate the LVB of a deleted inode - this way other
2171	 * nodes are forced to go to disk and discover the new inode
2172	 * status.
2173	 */
2174	if (oi->ip_flags & OCFS2_INODE_DELETED) {
2175		lvb->lvb_version = 0;
2176		goto out;
2177	}
2178
2179	lvb->lvb_version   = OCFS2_LVB_VERSION;
2180	lvb->lvb_isize	   = cpu_to_be64(i_size_read(inode));
2181	lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
2182	lvb->lvb_iuid      = cpu_to_be32(i_uid_read(inode));
2183	lvb->lvb_igid      = cpu_to_be32(i_gid_read(inode));
2184	lvb->lvb_imode     = cpu_to_be16(inode->i_mode);
2185	lvb->lvb_inlink    = cpu_to_be16(inode->i_nlink);
2186	ts = inode_get_atime(inode);
2187	lvb->lvb_iatime_packed = cpu_to_be64(ocfs2_pack_timespec(&ts));
2188	ts = inode_get_ctime(inode);
2189	lvb->lvb_ictime_packed = cpu_to_be64(ocfs2_pack_timespec(&ts));
2190	ts = inode_get_mtime(inode);
2191	lvb->lvb_imtime_packed = cpu_to_be64(ocfs2_pack_timespec(&ts));
2192	lvb->lvb_iattr    = cpu_to_be32(oi->ip_attr);
2193	lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features);
2194	lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
2195
2196out:
2197	mlog_meta_lvb(0, lockres);
2198}
2199
2200static void ocfs2_unpack_timespec(struct timespec64 *spec,
2201				  u64 packed_time)
2202{
2203	spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
2204	spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
2205}
2206
2207static int ocfs2_refresh_inode_from_lvb(struct inode *inode)
2208{
2209	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2210	struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2211	struct ocfs2_meta_lvb *lvb;
2212	struct timespec64 ts;
2213
2214	mlog_meta_lvb(0, lockres);
2215
2216	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2217	if (inode_wrong_type(inode, be16_to_cpu(lvb->lvb_imode)))
2218		return -ESTALE;
2219
2220	/* We're safe here without the lockres lock... */
2221	spin_lock(&oi->ip_lock);
2222	oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
2223	i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
2224
2225	oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
2226	oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures);
2227	ocfs2_set_inode_flags(inode);
2228
2229	/* fast-symlinks are a special case */
2230	if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
2231		inode->i_blocks = 0;
2232	else
2233		inode->i_blocks = ocfs2_inode_sector_count(inode);
2234
2235	i_uid_write(inode, be32_to_cpu(lvb->lvb_iuid));
2236	i_gid_write(inode, be32_to_cpu(lvb->lvb_igid));
2237	inode->i_mode    = be16_to_cpu(lvb->lvb_imode);
2238	set_nlink(inode, be16_to_cpu(lvb->lvb_inlink));
2239	ocfs2_unpack_timespec(&ts, be64_to_cpu(lvb->lvb_iatime_packed));
2240	inode_set_atime_to_ts(inode, ts);
2241	ocfs2_unpack_timespec(&ts, be64_to_cpu(lvb->lvb_imtime_packed));
2242	inode_set_mtime_to_ts(inode, ts);
2243	ocfs2_unpack_timespec(&ts, be64_to_cpu(lvb->lvb_ictime_packed));
2244	inode_set_ctime_to_ts(inode, ts);
2245	spin_unlock(&oi->ip_lock);
2246	return 0;
2247}
2248
2249static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
2250					      struct ocfs2_lock_res *lockres)
2251{
2252	struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2253
2254	if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)
2255	    && lvb->lvb_version == OCFS2_LVB_VERSION
2256	    && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
2257		return 1;
2258	return 0;
2259}
2260
2261/* Determine whether a lock resource needs to be refreshed, and
2262 * arbitrate who gets to refresh it.
2263 *
2264 *   0 means no refresh needed.
2265 *
2266 *   > 0 means you need to refresh this and you MUST call
2267 *   ocfs2_complete_lock_res_refresh afterwards. */
2268static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
2269{
2270	unsigned long flags;
2271	int status = 0;
2272
2273refresh_check:
2274	spin_lock_irqsave(&lockres->l_lock, flags);
2275	if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
2276		spin_unlock_irqrestore(&lockres->l_lock, flags);
2277		goto bail;
2278	}
2279
2280	if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
2281		spin_unlock_irqrestore(&lockres->l_lock, flags);
2282
2283		ocfs2_wait_on_refreshing_lock(lockres);
2284		goto refresh_check;
2285	}
2286
2287	/* Ok, I'll be the one to refresh this lock. */
2288	lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
2289	spin_unlock_irqrestore(&lockres->l_lock, flags);
2290
2291	status = 1;
2292bail:
2293	mlog(0, "status %d\n", status);
2294	return status;
2295}
2296
2297/* If status is non zero, I'll mark it as not being in refresh
2298 * anymroe, but i won't clear the needs refresh flag. */
2299static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
2300						   int status)
2301{
2302	unsigned long flags;
2303
2304	spin_lock_irqsave(&lockres->l_lock, flags);
2305	lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
2306	if (!status)
2307		lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
2308	spin_unlock_irqrestore(&lockres->l_lock, flags);
2309
2310	wake_up(&lockres->l_event);
2311}
2312
2313/* may or may not return a bh if it went to disk. */
2314static int ocfs2_inode_lock_update(struct inode *inode,
2315				  struct buffer_head **bh)
2316{
2317	int status = 0;
2318	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2319	struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2320	struct ocfs2_dinode *fe;
2321	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2322
2323	if (ocfs2_mount_local(osb))
2324		goto bail;
2325
2326	spin_lock(&oi->ip_lock);
2327	if (oi->ip_flags & OCFS2_INODE_DELETED) {
2328		mlog(0, "Orphaned inode %llu was deleted while we "
2329		     "were waiting on a lock. ip_flags = 0x%x\n",
2330		     (unsigned long long)oi->ip_blkno, oi->ip_flags);
2331		spin_unlock(&oi->ip_lock);
2332		status = -ENOENT;
2333		goto bail;
2334	}
2335	spin_unlock(&oi->ip_lock);
2336
2337	if (!ocfs2_should_refresh_lock_res(lockres))
2338		goto bail;
2339
2340	/* This will discard any caching information we might have had
2341	 * for the inode metadata. */
2342	ocfs2_metadata_cache_purge(INODE_CACHE(inode));
2343
2344	ocfs2_extent_map_trunc(inode, 0);
2345
2346	if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
2347		mlog(0, "Trusting LVB on inode %llu\n",
2348		     (unsigned long long)oi->ip_blkno);
2349		status = ocfs2_refresh_inode_from_lvb(inode);
2350		goto bail_refresh;
2351	} else {
2352		/* Boo, we have to go to disk. */
2353		/* read bh, cast, ocfs2_refresh_inode */
2354		status = ocfs2_read_inode_block(inode, bh);
2355		if (status < 0) {
2356			mlog_errno(status);
2357			goto bail_refresh;
2358		}
2359		fe = (struct ocfs2_dinode *) (*bh)->b_data;
2360		if (inode_wrong_type(inode, le16_to_cpu(fe->i_mode))) {
2361			status = -ESTALE;
2362			goto bail_refresh;
2363		}
2364
2365		/* This is a good chance to make sure we're not
2366		 * locking an invalid object.  ocfs2_read_inode_block()
2367		 * already checked that the inode block is sane.
2368		 *
2369		 * We bug on a stale inode here because we checked
2370		 * above whether it was wiped from disk. The wiping
2371		 * node provides a guarantee that we receive that
2372		 * message and can mark the inode before dropping any
2373		 * locks associated with it. */
2374		mlog_bug_on_msg(inode->i_generation !=
2375				le32_to_cpu(fe->i_generation),
2376				"Invalid dinode %llu disk generation: %u "
2377				"inode->i_generation: %u\n",
2378				(unsigned long long)oi->ip_blkno,
2379				le32_to_cpu(fe->i_generation),
2380				inode->i_generation);
2381		mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
2382				!(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
2383				"Stale dinode %llu dtime: %llu flags: 0x%x\n",
2384				(unsigned long long)oi->ip_blkno,
2385				(unsigned long long)le64_to_cpu(fe->i_dtime),
2386				le32_to_cpu(fe->i_flags));
2387
2388		ocfs2_refresh_inode(inode, fe);
2389		ocfs2_track_lock_refresh(lockres);
2390	}
2391
2392	status = 0;
2393bail_refresh:
2394	ocfs2_complete_lock_res_refresh(lockres, status);
2395bail:
2396	return status;
2397}
2398
2399static int ocfs2_assign_bh(struct inode *inode,
2400			   struct buffer_head **ret_bh,
2401			   struct buffer_head *passed_bh)
2402{
2403	int status;
2404
2405	if (passed_bh) {
2406		/* Ok, the update went to disk for us, use the
2407		 * returned bh. */
2408		*ret_bh = passed_bh;
2409		get_bh(*ret_bh);
2410
2411		return 0;
2412	}
2413
2414	status = ocfs2_read_inode_block(inode, ret_bh);
2415	if (status < 0)
2416		mlog_errno(status);
2417
2418	return status;
2419}
2420
2421/*
2422 * returns < 0 error if the callback will never be called, otherwise
2423 * the result of the lock will be communicated via the callback.
2424 */
2425int ocfs2_inode_lock_full_nested(struct inode *inode,
2426				 struct buffer_head **ret_bh,
2427				 int ex,
2428				 int arg_flags,
2429				 int subclass)
2430{
2431	int status, level, acquired;
2432	u32 dlm_flags;
2433	struct ocfs2_lock_res *lockres = NULL;
2434	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2435	struct buffer_head *local_bh = NULL;
2436
2437	mlog(0, "inode %llu, take %s META lock\n",
2438	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
2439	     ex ? "EXMODE" : "PRMODE");
2440
2441	status = 0;
2442	acquired = 0;
2443	/* We'll allow faking a readonly metadata lock for
2444	 * rodevices. */
2445	if (ocfs2_is_hard_readonly(osb)) {
2446		if (ex)
2447			status = -EROFS;
2448		goto getbh;
2449	}
2450
2451	if ((arg_flags & OCFS2_META_LOCK_GETBH) ||
2452	    ocfs2_mount_local(osb))
2453		goto update;
2454
2455	if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2456		ocfs2_wait_for_recovery(osb);
2457
2458	lockres = &OCFS2_I(inode)->ip_inode_lockres;
2459	level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2460	dlm_flags = 0;
2461	if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
2462		dlm_flags |= DLM_LKF_NOQUEUE;
2463
2464	status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
2465				      arg_flags, subclass, _RET_IP_);
2466	if (status < 0) {
2467		if (status != -EAGAIN)
2468			mlog_errno(status);
2469		goto bail;
2470	}
2471
2472	/* Notify the error cleanup path to drop the cluster lock. */
2473	acquired = 1;
2474
2475	/* We wait twice because a node may have died while we were in
2476	 * the lower dlm layers. The second time though, we've
2477	 * committed to owning this lock so we don't allow signals to
2478	 * abort the operation. */
2479	if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2480		ocfs2_wait_for_recovery(osb);
2481
2482update:
2483	/*
2484	 * We only see this flag if we're being called from
2485	 * ocfs2_read_locked_inode(). It means we're locking an inode
2486	 * which hasn't been populated yet, so clear the refresh flag
2487	 * and let the caller handle it.
2488	 */
2489	if (inode->i_state & I_NEW) {
2490		status = 0;
2491		if (lockres)
2492			ocfs2_complete_lock_res_refresh(lockres, 0);
2493		goto bail;
2494	}
2495
2496	/* This is fun. The caller may want a bh back, or it may
2497	 * not. ocfs2_inode_lock_update definitely wants one in, but
2498	 * may or may not read one, depending on what's in the
2499	 * LVB. The result of all of this is that we've *only* gone to
2500	 * disk if we have to, so the complexity is worthwhile. */
2501	status = ocfs2_inode_lock_update(inode, &local_bh);
2502	if (status < 0) {
2503		if (status != -ENOENT)
2504			mlog_errno(status);
2505		goto bail;
2506	}
2507getbh:
2508	if (ret_bh) {
2509		status = ocfs2_assign_bh(inode, ret_bh, local_bh);
2510		if (status < 0) {
2511			mlog_errno(status);
2512			goto bail;
2513		}
2514	}
2515
2516bail:
2517	if (status < 0) {
2518		if (ret_bh && (*ret_bh)) {
2519			brelse(*ret_bh);
2520			*ret_bh = NULL;
2521		}
2522		if (acquired)
2523			ocfs2_inode_unlock(inode, ex);
2524	}
2525
2526	brelse(local_bh);
 
 
2527	return status;
2528}
2529
2530/*
2531 * This is working around a lock inversion between tasks acquiring DLM
2532 * locks while holding a page lock and the downconvert thread which
2533 * blocks dlm lock acquiry while acquiring page locks.
2534 *
2535 * ** These _with_page variantes are only intended to be called from aop
2536 * methods that hold page locks and return a very specific *positive* error
2537 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2538 *
2539 * The DLM is called such that it returns -EAGAIN if it would have
2540 * blocked waiting for the downconvert thread.  In that case we unlock
2541 * our page so the downconvert thread can make progress.  Once we've
2542 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2543 * that called us can bubble that back up into the VFS who will then
2544 * immediately retry the aop call.
2545 */
2546int ocfs2_inode_lock_with_page(struct inode *inode,
2547			      struct buffer_head **ret_bh,
2548			      int ex,
2549			      struct page *page)
2550{
2551	int ret;
2552
2553	ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
2554	if (ret == -EAGAIN) {
2555		unlock_page(page);
2556		/*
2557		 * If we can't get inode lock immediately, we should not return
2558		 * directly here, since this will lead to a softlockup problem.
2559		 * The method is to get a blocking lock and immediately unlock
2560		 * before returning, this can avoid CPU resource waste due to
2561		 * lots of retries, and benefits fairness in getting lock.
2562		 */
2563		if (ocfs2_inode_lock(inode, ret_bh, ex) == 0)
2564			ocfs2_inode_unlock(inode, ex);
2565		ret = AOP_TRUNCATED_PAGE;
2566	}
2567
2568	return ret;
2569}
2570
2571int ocfs2_inode_lock_atime(struct inode *inode,
2572			  struct vfsmount *vfsmnt,
2573			  int *level, int wait)
2574{
2575	int ret;
2576
2577	if (wait)
2578		ret = ocfs2_inode_lock(inode, NULL, 0);
2579	else
2580		ret = ocfs2_try_inode_lock(inode, NULL, 0);
2581
2582	if (ret < 0) {
2583		if (ret != -EAGAIN)
2584			mlog_errno(ret);
2585		return ret;
2586	}
2587
2588	/*
2589	 * If we should update atime, we will get EX lock,
2590	 * otherwise we just get PR lock.
2591	 */
2592	if (ocfs2_should_update_atime(inode, vfsmnt)) {
2593		struct buffer_head *bh = NULL;
2594
2595		ocfs2_inode_unlock(inode, 0);
2596		if (wait)
2597			ret = ocfs2_inode_lock(inode, &bh, 1);
2598		else
2599			ret = ocfs2_try_inode_lock(inode, &bh, 1);
2600
2601		if (ret < 0) {
2602			if (ret != -EAGAIN)
2603				mlog_errno(ret);
2604			return ret;
2605		}
2606		*level = 1;
2607		if (ocfs2_should_update_atime(inode, vfsmnt))
2608			ocfs2_update_inode_atime(inode, bh);
2609		brelse(bh);
 
2610	} else
2611		*level = 0;
2612
2613	return ret;
2614}
2615
2616void ocfs2_inode_unlock(struct inode *inode,
2617		       int ex)
2618{
2619	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2620	struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres;
2621	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2622
2623	mlog(0, "inode %llu drop %s META lock\n",
2624	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
2625	     ex ? "EXMODE" : "PRMODE");
2626
2627	if (!ocfs2_is_hard_readonly(osb) &&
2628	    !ocfs2_mount_local(osb))
2629		ocfs2_cluster_unlock(osb, lockres, level);
2630}
2631
2632/*
2633 * This _tracker variantes are introduced to deal with the recursive cluster
2634 * locking issue. The idea is to keep track of a lock holder on the stack of
2635 * the current process. If there's a lock holder on the stack, we know the
2636 * task context is already protected by cluster locking. Currently, they're
2637 * used in some VFS entry routines.
2638 *
2639 * return < 0 on error, return == 0 if there's no lock holder on the stack
2640 * before this call, return == 1 if this call would be a recursive locking.
2641 * return == -1 if this lock attempt will cause an upgrade which is forbidden.
2642 *
2643 * When taking lock levels into account,we face some different situations.
2644 *
2645 * 1. no lock is held
2646 *    In this case, just lock the inode as requested and return 0
2647 *
2648 * 2. We are holding a lock
2649 *    For this situation, things diverges into several cases
2650 *
2651 *    wanted     holding	     what to do
2652 *    ex		ex	    see 2.1 below
2653 *    ex		pr	    see 2.2 below
2654 *    pr		ex	    see 2.1 below
2655 *    pr		pr	    see 2.1 below
2656 *
2657 *    2.1 lock level that is been held is compatible
2658 *    with the wanted level, so no lock action will be tacken.
2659 *
2660 *    2.2 Otherwise, an upgrade is needed, but it is forbidden.
2661 *
2662 * Reason why upgrade within a process is forbidden is that
2663 * lock upgrade may cause dead lock. The following illustrates
2664 * how it happens.
2665 *
2666 *         thread on node1                             thread on node2
2667 * ocfs2_inode_lock_tracker(ex=0)
2668 *
2669 *                                <======   ocfs2_inode_lock_tracker(ex=1)
2670 *
2671 * ocfs2_inode_lock_tracker(ex=1)
2672 */
2673int ocfs2_inode_lock_tracker(struct inode *inode,
2674			     struct buffer_head **ret_bh,
2675			     int ex,
2676			     struct ocfs2_lock_holder *oh)
2677{
2678	int status = 0;
2679	struct ocfs2_lock_res *lockres;
2680	struct ocfs2_lock_holder *tmp_oh;
2681	struct pid *pid = task_pid(current);
2682
2683
2684	lockres = &OCFS2_I(inode)->ip_inode_lockres;
2685	tmp_oh = ocfs2_pid_holder(lockres, pid);
2686
2687	if (!tmp_oh) {
2688		/*
2689		 * This corresponds to the case 1.
2690		 * We haven't got any lock before.
2691		 */
2692		status = ocfs2_inode_lock_full(inode, ret_bh, ex, 0);
2693		if (status < 0) {
2694			if (status != -ENOENT)
2695				mlog_errno(status);
2696			return status;
2697		}
2698
2699		oh->oh_ex = ex;
2700		ocfs2_add_holder(lockres, oh);
2701		return 0;
2702	}
2703
2704	if (unlikely(ex && !tmp_oh->oh_ex)) {
2705		/*
2706		 * case 2.2 upgrade may cause dead lock, forbid it.
2707		 */
2708		mlog(ML_ERROR, "Recursive locking is not permitted to "
2709		     "upgrade to EX level from PR level.\n");
2710		dump_stack();
2711		return -EINVAL;
2712	}
2713
2714	/*
2715	 *  case 2.1 OCFS2_META_LOCK_GETBH flag make ocfs2_inode_lock_full.
2716	 *  ignore the lock level and just update it.
2717	 */
2718	if (ret_bh) {
2719		status = ocfs2_inode_lock_full(inode, ret_bh, ex,
2720					       OCFS2_META_LOCK_GETBH);
2721		if (status < 0) {
2722			if (status != -ENOENT)
2723				mlog_errno(status);
2724			return status;
2725		}
2726	}
2727	return 1;
2728}
2729
2730void ocfs2_inode_unlock_tracker(struct inode *inode,
2731				int ex,
2732				struct ocfs2_lock_holder *oh,
2733				int had_lock)
2734{
2735	struct ocfs2_lock_res *lockres;
2736
2737	lockres = &OCFS2_I(inode)->ip_inode_lockres;
2738	/* had_lock means that the currect process already takes the cluster
2739	 * lock previously.
2740	 * If had_lock is 1, we have nothing to do here.
2741	 * If had_lock is 0, we will release the lock.
2742	 */
2743	if (!had_lock) {
2744		ocfs2_inode_unlock(inode, oh->oh_ex);
2745		ocfs2_remove_holder(lockres, oh);
2746	}
2747}
2748
2749int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno)
2750{
2751	struct ocfs2_lock_res *lockres;
2752	struct ocfs2_orphan_scan_lvb *lvb;
2753	int status = 0;
2754
2755	if (ocfs2_is_hard_readonly(osb))
2756		return -EROFS;
2757
2758	if (ocfs2_mount_local(osb))
2759		return 0;
2760
2761	lockres = &osb->osb_orphan_scan.os_lockres;
2762	status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2763	if (status < 0)
2764		return status;
2765
2766	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2767	if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2768	    lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION)
2769		*seqno = be32_to_cpu(lvb->lvb_os_seqno);
2770	else
2771		*seqno = osb->osb_orphan_scan.os_seqno + 1;
2772
2773	return status;
2774}
2775
2776void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno)
2777{
2778	struct ocfs2_lock_res *lockres;
2779	struct ocfs2_orphan_scan_lvb *lvb;
2780
2781	if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) {
2782		lockres = &osb->osb_orphan_scan.os_lockres;
2783		lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2784		lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION;
2785		lvb->lvb_os_seqno = cpu_to_be32(seqno);
2786		ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2787	}
2788}
2789
2790int ocfs2_super_lock(struct ocfs2_super *osb,
2791		     int ex)
2792{
2793	int status = 0;
2794	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2795	struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2796
2797	if (ocfs2_is_hard_readonly(osb))
2798		return -EROFS;
2799
2800	if (ocfs2_mount_local(osb))
2801		goto bail;
2802
2803	status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
2804	if (status < 0) {
2805		mlog_errno(status);
2806		goto bail;
2807	}
2808
2809	/* The super block lock path is really in the best position to
2810	 * know when resources covered by the lock need to be
2811	 * refreshed, so we do it here. Of course, making sense of
2812	 * everything is up to the caller :) */
2813	status = ocfs2_should_refresh_lock_res(lockres);
2814	if (status) {
2815		status = ocfs2_refresh_slot_info(osb);
2816
2817		ocfs2_complete_lock_res_refresh(lockres, status);
2818
2819		if (status < 0) {
2820			ocfs2_cluster_unlock(osb, lockres, level);
2821			mlog_errno(status);
2822		}
2823		ocfs2_track_lock_refresh(lockres);
2824	}
2825bail:
2826	return status;
2827}
2828
2829void ocfs2_super_unlock(struct ocfs2_super *osb,
2830			int ex)
2831{
2832	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2833	struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2834
2835	if (!ocfs2_mount_local(osb))
2836		ocfs2_cluster_unlock(osb, lockres, level);
2837}
2838
2839int ocfs2_rename_lock(struct ocfs2_super *osb)
2840{
2841	int status;
2842	struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2843
2844	if (ocfs2_is_hard_readonly(osb))
2845		return -EROFS;
2846
2847	if (ocfs2_mount_local(osb))
2848		return 0;
2849
2850	status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2851	if (status < 0)
2852		mlog_errno(status);
2853
2854	return status;
2855}
2856
2857void ocfs2_rename_unlock(struct ocfs2_super *osb)
2858{
2859	struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2860
2861	if (!ocfs2_mount_local(osb))
2862		ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2863}
2864
2865int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex)
2866{
2867	int status;
2868	struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2869
2870	if (ocfs2_is_hard_readonly(osb))
2871		return -EROFS;
2872
2873	if (ex)
2874		down_write(&osb->nfs_sync_rwlock);
2875	else
2876		down_read(&osb->nfs_sync_rwlock);
2877
2878	if (ocfs2_mount_local(osb))
2879		return 0;
2880
2881	status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE,
2882				    0, 0);
2883	if (status < 0) {
2884		mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status);
2885
2886		if (ex)
2887			up_write(&osb->nfs_sync_rwlock);
2888		else
2889			up_read(&osb->nfs_sync_rwlock);
2890	}
2891
2892	return status;
2893}
2894
2895void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex)
2896{
2897	struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2898
2899	if (!ocfs2_mount_local(osb))
2900		ocfs2_cluster_unlock(osb, lockres,
2901				     ex ? LKM_EXMODE : LKM_PRMODE);
2902	if (ex)
2903		up_write(&osb->nfs_sync_rwlock);
2904	else
2905		up_read(&osb->nfs_sync_rwlock);
2906}
2907
2908int ocfs2_trim_fs_lock(struct ocfs2_super *osb,
2909		       struct ocfs2_trim_fs_info *info, int trylock)
2910{
2911	int status;
2912	struct ocfs2_trim_fs_lvb *lvb;
2913	struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
2914
2915	if (info)
2916		info->tf_valid = 0;
2917
2918	if (ocfs2_is_hard_readonly(osb))
2919		return -EROFS;
2920
2921	if (ocfs2_mount_local(osb))
2922		return 0;
2923
2924	status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX,
2925				    trylock ? DLM_LKF_NOQUEUE : 0, 0);
2926	if (status < 0) {
2927		if (status != -EAGAIN)
2928			mlog_errno(status);
2929		return status;
2930	}
2931
2932	if (info) {
2933		lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2934		if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2935		    lvb->lvb_version == OCFS2_TRIMFS_LVB_VERSION) {
2936			info->tf_valid = 1;
2937			info->tf_success = lvb->lvb_success;
2938			info->tf_nodenum = be32_to_cpu(lvb->lvb_nodenum);
2939			info->tf_start = be64_to_cpu(lvb->lvb_start);
2940			info->tf_len = be64_to_cpu(lvb->lvb_len);
2941			info->tf_minlen = be64_to_cpu(lvb->lvb_minlen);
2942			info->tf_trimlen = be64_to_cpu(lvb->lvb_trimlen);
2943		}
2944	}
2945
2946	return status;
2947}
2948
2949void ocfs2_trim_fs_unlock(struct ocfs2_super *osb,
2950			  struct ocfs2_trim_fs_info *info)
2951{
2952	struct ocfs2_trim_fs_lvb *lvb;
2953	struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
2954
2955	if (ocfs2_mount_local(osb))
2956		return;
2957
2958	if (info) {
2959		lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2960		lvb->lvb_version = OCFS2_TRIMFS_LVB_VERSION;
2961		lvb->lvb_success = info->tf_success;
2962		lvb->lvb_nodenum = cpu_to_be32(info->tf_nodenum);
2963		lvb->lvb_start = cpu_to_be64(info->tf_start);
2964		lvb->lvb_len = cpu_to_be64(info->tf_len);
2965		lvb->lvb_minlen = cpu_to_be64(info->tf_minlen);
2966		lvb->lvb_trimlen = cpu_to_be64(info->tf_trimlen);
2967	}
2968
2969	ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2970}
2971
2972int ocfs2_dentry_lock(struct dentry *dentry, int ex)
2973{
2974	int ret;
2975	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2976	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2977	struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2978
2979	BUG_ON(!dl);
2980
2981	if (ocfs2_is_hard_readonly(osb)) {
2982		if (ex)
2983			return -EROFS;
2984		return 0;
2985	}
2986
2987	if (ocfs2_mount_local(osb))
2988		return 0;
2989
2990	ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2991	if (ret < 0)
2992		mlog_errno(ret);
2993
2994	return ret;
2995}
2996
2997void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2998{
2999	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3000	struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
3001	struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
3002
3003	if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
3004		ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
3005}
3006
3007/* Reference counting of the dlm debug structure. We want this because
3008 * open references on the debug inodes can live on after a mount, so
3009 * we can't rely on the ocfs2_super to always exist. */
3010static void ocfs2_dlm_debug_free(struct kref *kref)
3011{
3012	struct ocfs2_dlm_debug *dlm_debug;
3013
3014	dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
3015
3016	kfree(dlm_debug);
3017}
3018
3019void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
3020{
3021	if (dlm_debug)
3022		kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
3023}
3024
3025static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
3026{
3027	kref_get(&debug->d_refcnt);
3028}
3029
3030struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
3031{
3032	struct ocfs2_dlm_debug *dlm_debug;
3033
3034	dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
3035	if (!dlm_debug) {
3036		mlog_errno(-ENOMEM);
3037		goto out;
3038	}
3039
3040	kref_init(&dlm_debug->d_refcnt);
3041	INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
3042	dlm_debug->d_filter_secs = 0;
3043out:
3044	return dlm_debug;
3045}
3046
3047/* Access to this is arbitrated for us via seq_file->sem. */
3048struct ocfs2_dlm_seq_priv {
3049	struct ocfs2_dlm_debug *p_dlm_debug;
3050	struct ocfs2_lock_res p_iter_res;
3051	struct ocfs2_lock_res p_tmp_res;
3052};
3053
3054static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
3055						 struct ocfs2_dlm_seq_priv *priv)
3056{
3057	struct ocfs2_lock_res *iter, *ret = NULL;
3058	struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
3059
3060	assert_spin_locked(&ocfs2_dlm_tracking_lock);
3061
3062	list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
3063		/* discover the head of the list */
3064		if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
3065			mlog(0, "End of list found, %p\n", ret);
3066			break;
3067		}
3068
3069		/* We track our "dummy" iteration lockres' by a NULL
3070		 * l_ops field. */
3071		if (iter->l_ops != NULL) {
3072			ret = iter;
3073			break;
3074		}
3075	}
3076
3077	return ret;
3078}
3079
3080static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
3081{
3082	struct ocfs2_dlm_seq_priv *priv = m->private;
3083	struct ocfs2_lock_res *iter;
3084
3085	spin_lock(&ocfs2_dlm_tracking_lock);
3086	iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
3087	if (iter) {
3088		/* Since lockres' have the lifetime of their container
3089		 * (which can be inodes, ocfs2_supers, etc) we want to
3090		 * copy this out to a temporary lockres while still
3091		 * under the spinlock. Obviously after this we can't
3092		 * trust any pointers on the copy returned, but that's
3093		 * ok as the information we want isn't typically held
3094		 * in them. */
3095		priv->p_tmp_res = *iter;
3096		iter = &priv->p_tmp_res;
3097	}
3098	spin_unlock(&ocfs2_dlm_tracking_lock);
3099
3100	return iter;
3101}
3102
3103static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
3104{
3105}
3106
3107static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
3108{
3109	struct ocfs2_dlm_seq_priv *priv = m->private;
3110	struct ocfs2_lock_res *iter = v;
3111	struct ocfs2_lock_res *dummy = &priv->p_iter_res;
3112
3113	(*pos)++;
3114	spin_lock(&ocfs2_dlm_tracking_lock);
3115	iter = ocfs2_dlm_next_res(iter, priv);
3116	list_del_init(&dummy->l_debug_list);
3117	if (iter) {
3118		list_add(&dummy->l_debug_list, &iter->l_debug_list);
3119		priv->p_tmp_res = *iter;
3120		iter = &priv->p_tmp_res;
3121	}
3122	spin_unlock(&ocfs2_dlm_tracking_lock);
3123
3124	return iter;
3125}
3126
3127/*
3128 * Version is used by debugfs.ocfs2 to determine the format being used
3129 *
3130 * New in version 2
3131 *	- Lock stats printed
3132 * New in version 3
3133 *	- Max time in lock stats is in usecs (instead of nsecs)
3134 * New in version 4
3135 *	- Add last pr/ex unlock times and first lock wait time in usecs
3136 */
3137#define OCFS2_DLM_DEBUG_STR_VERSION 4
3138static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
3139{
3140	int i;
3141	char *lvb;
3142	struct ocfs2_lock_res *lockres = v;
3143#ifdef CONFIG_OCFS2_FS_STATS
3144	u64 now, last;
3145	struct ocfs2_dlm_debug *dlm_debug =
3146			((struct ocfs2_dlm_seq_priv *)m->private)->p_dlm_debug;
3147#endif
3148
3149	if (!lockres)
3150		return -EINVAL;
3151
3152#ifdef CONFIG_OCFS2_FS_STATS
3153	if (!lockres->l_lock_wait && dlm_debug->d_filter_secs) {
3154		now = ktime_to_us(ktime_get_real());
3155		last = max(lockres->l_lock_prmode.ls_last,
3156			   lockres->l_lock_exmode.ls_last);
3157		/*
3158		 * Use d_filter_secs field to filter lock resources dump,
3159		 * the default d_filter_secs(0) value filters nothing,
3160		 * otherwise, only dump the last N seconds active lock
3161		 * resources.
3162		 */
3163		if (div_u64(now - last, 1000000) > dlm_debug->d_filter_secs)
3164			return 0;
3165	}
3166#endif
3167
3168	seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
3169
3170	if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
3171		seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
3172			   lockres->l_name,
3173			   (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
3174	else
3175		seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
3176
3177	seq_printf(m, "%d\t"
3178		   "0x%lx\t"
3179		   "0x%x\t"
3180		   "0x%x\t"
3181		   "%u\t"
3182		   "%u\t"
3183		   "%d\t"
3184		   "%d\t",
3185		   lockres->l_level,
3186		   lockres->l_flags,
3187		   lockres->l_action,
3188		   lockres->l_unlock_action,
3189		   lockres->l_ro_holders,
3190		   lockres->l_ex_holders,
3191		   lockres->l_requested,
3192		   lockres->l_blocking);
3193
3194	/* Dump the raw LVB */
3195	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3196	for(i = 0; i < DLM_LVB_LEN; i++)
3197		seq_printf(m, "0x%x\t", lvb[i]);
3198
3199#ifdef CONFIG_OCFS2_FS_STATS
3200# define lock_num_prmode(_l)		((_l)->l_lock_prmode.ls_gets)
3201# define lock_num_exmode(_l)		((_l)->l_lock_exmode.ls_gets)
3202# define lock_num_prmode_failed(_l)	((_l)->l_lock_prmode.ls_fail)
3203# define lock_num_exmode_failed(_l)	((_l)->l_lock_exmode.ls_fail)
3204# define lock_total_prmode(_l)		((_l)->l_lock_prmode.ls_total)
3205# define lock_total_exmode(_l)		((_l)->l_lock_exmode.ls_total)
3206# define lock_max_prmode(_l)		((_l)->l_lock_prmode.ls_max)
3207# define lock_max_exmode(_l)		((_l)->l_lock_exmode.ls_max)
3208# define lock_refresh(_l)		((_l)->l_lock_refresh)
3209# define lock_last_prmode(_l)		((_l)->l_lock_prmode.ls_last)
3210# define lock_last_exmode(_l)		((_l)->l_lock_exmode.ls_last)
3211# define lock_wait(_l)			((_l)->l_lock_wait)
3212#else
3213# define lock_num_prmode(_l)		(0)
3214# define lock_num_exmode(_l)		(0)
3215# define lock_num_prmode_failed(_l)	(0)
3216# define lock_num_exmode_failed(_l)	(0)
3217# define lock_total_prmode(_l)		(0ULL)
3218# define lock_total_exmode(_l)		(0ULL)
3219# define lock_max_prmode(_l)		(0)
3220# define lock_max_exmode(_l)		(0)
3221# define lock_refresh(_l)		(0)
3222# define lock_last_prmode(_l)		(0ULL)
3223# define lock_last_exmode(_l)		(0ULL)
3224# define lock_wait(_l)			(0ULL)
3225#endif
3226	/* The following seq_print was added in version 2 of this output */
3227	seq_printf(m, "%u\t"
3228		   "%u\t"
3229		   "%u\t"
3230		   "%u\t"
3231		   "%llu\t"
3232		   "%llu\t"
3233		   "%u\t"
3234		   "%u\t"
3235		   "%u\t"
3236		   "%llu\t"
3237		   "%llu\t"
3238		   "%llu\t",
3239		   lock_num_prmode(lockres),
3240		   lock_num_exmode(lockres),
3241		   lock_num_prmode_failed(lockres),
3242		   lock_num_exmode_failed(lockres),
3243		   lock_total_prmode(lockres),
3244		   lock_total_exmode(lockres),
3245		   lock_max_prmode(lockres),
3246		   lock_max_exmode(lockres),
3247		   lock_refresh(lockres),
3248		   lock_last_prmode(lockres),
3249		   lock_last_exmode(lockres),
3250		   lock_wait(lockres));
3251
3252	/* End the line */
3253	seq_printf(m, "\n");
3254	return 0;
3255}
3256
3257static const struct seq_operations ocfs2_dlm_seq_ops = {
3258	.start =	ocfs2_dlm_seq_start,
3259	.stop =		ocfs2_dlm_seq_stop,
3260	.next =		ocfs2_dlm_seq_next,
3261	.show =		ocfs2_dlm_seq_show,
3262};
3263
3264static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
3265{
3266	struct seq_file *seq = file->private_data;
3267	struct ocfs2_dlm_seq_priv *priv = seq->private;
3268	struct ocfs2_lock_res *res = &priv->p_iter_res;
3269
3270	ocfs2_remove_lockres_tracking(res);
3271	ocfs2_put_dlm_debug(priv->p_dlm_debug);
3272	return seq_release_private(inode, file);
3273}
3274
3275static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
3276{
3277	struct ocfs2_dlm_seq_priv *priv;
3278	struct ocfs2_super *osb;
3279
3280	priv = __seq_open_private(file, &ocfs2_dlm_seq_ops, sizeof(*priv));
3281	if (!priv) {
3282		mlog_errno(-ENOMEM);
3283		return -ENOMEM;
3284	}
3285
3286	osb = inode->i_private;
3287	ocfs2_get_dlm_debug(osb->osb_dlm_debug);
3288	priv->p_dlm_debug = osb->osb_dlm_debug;
3289	INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
3290
3291	ocfs2_add_lockres_tracking(&priv->p_iter_res,
3292				   priv->p_dlm_debug);
3293
3294	return 0;
3295}
3296
3297static const struct file_operations ocfs2_dlm_debug_fops = {
3298	.open =		ocfs2_dlm_debug_open,
3299	.release =	ocfs2_dlm_debug_release,
3300	.read =		seq_read,
3301	.llseek =	seq_lseek,
3302};
3303
3304static void ocfs2_dlm_init_debug(struct ocfs2_super *osb)
3305{
 
3306	struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
3307
3308	debugfs_create_file("locking_state", S_IFREG|S_IRUSR,
3309			    osb->osb_debug_root, osb, &ocfs2_dlm_debug_fops);
 
 
 
 
 
 
 
 
 
3310
3311	debugfs_create_u32("locking_filter", 0600, osb->osb_debug_root,
3312			   &dlm_debug->d_filter_secs);
3313	ocfs2_get_dlm_debug(dlm_debug);
 
 
3314}
3315
3316static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
3317{
3318	struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
3319
3320	if (dlm_debug)
 
3321		ocfs2_put_dlm_debug(dlm_debug);
 
3322}
3323
3324int ocfs2_dlm_init(struct ocfs2_super *osb)
3325{
3326	int status = 0;
3327	struct ocfs2_cluster_connection *conn = NULL;
3328
3329	if (ocfs2_mount_local(osb)) {
3330		osb->node_num = 0;
3331		goto local;
3332	}
3333
3334	ocfs2_dlm_init_debug(osb);
 
 
 
 
3335
3336	/* launch downconvert thread */
3337	osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc-%s",
3338			osb->uuid_str);
3339	if (IS_ERR(osb->dc_task)) {
3340		status = PTR_ERR(osb->dc_task);
3341		osb->dc_task = NULL;
3342		mlog_errno(status);
3343		goto bail;
3344	}
3345
3346	/* for now, uuid == domain */
3347	status = ocfs2_cluster_connect(osb->osb_cluster_stack,
3348				       osb->osb_cluster_name,
3349				       strlen(osb->osb_cluster_name),
3350				       osb->uuid_str,
3351				       strlen(osb->uuid_str),
3352				       &lproto, ocfs2_do_node_down, osb,
3353				       &conn);
3354	if (status) {
3355		mlog_errno(status);
3356		goto bail;
3357	}
3358
3359	status = ocfs2_cluster_this_node(conn, &osb->node_num);
3360	if (status < 0) {
3361		mlog_errno(status);
3362		mlog(ML_ERROR,
3363		     "could not find this host's node number\n");
3364		ocfs2_cluster_disconnect(conn, 0);
3365		goto bail;
3366	}
3367
3368local:
3369	ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
3370	ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
3371	ocfs2_nfs_sync_lock_init(osb);
3372	ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb);
3373
3374	osb->cconn = conn;
3375bail:
3376	if (status < 0) {
3377		ocfs2_dlm_shutdown_debug(osb);
3378		if (osb->dc_task)
3379			kthread_stop(osb->dc_task);
3380	}
3381
3382	return status;
3383}
3384
3385void ocfs2_dlm_shutdown(struct ocfs2_super *osb,
3386			int hangup_pending)
3387{
3388	ocfs2_drop_osb_locks(osb);
3389
3390	/*
3391	 * Now that we have dropped all locks and ocfs2_dismount_volume()
3392	 * has disabled recovery, the DLM won't be talking to us.  It's
3393	 * safe to tear things down before disconnecting the cluster.
3394	 */
3395
3396	if (osb->dc_task) {
3397		kthread_stop(osb->dc_task);
3398		osb->dc_task = NULL;
3399	}
3400
3401	ocfs2_lock_res_free(&osb->osb_super_lockres);
3402	ocfs2_lock_res_free(&osb->osb_rename_lockres);
3403	ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres);
3404	ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres);
3405
3406	if (osb->cconn) {
3407		ocfs2_cluster_disconnect(osb->cconn, hangup_pending);
3408		osb->cconn = NULL;
3409
3410		ocfs2_dlm_shutdown_debug(osb);
3411	}
3412}
3413
3414static int ocfs2_drop_lock(struct ocfs2_super *osb,
3415			   struct ocfs2_lock_res *lockres)
3416{
3417	int ret;
3418	unsigned long flags;
3419	u32 lkm_flags = 0;
3420
3421	/* We didn't get anywhere near actually using this lockres. */
3422	if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
3423		goto out;
3424
3425	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3426		lkm_flags |= DLM_LKF_VALBLK;
3427
3428	spin_lock_irqsave(&lockres->l_lock, flags);
3429
3430	mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
3431			"lockres %s, flags 0x%lx\n",
3432			lockres->l_name, lockres->l_flags);
3433
3434	while (lockres->l_flags & OCFS2_LOCK_BUSY) {
3435		mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3436		     "%u, unlock_action = %u\n",
3437		     lockres->l_name, lockres->l_flags, lockres->l_action,
3438		     lockres->l_unlock_action);
3439
3440		spin_unlock_irqrestore(&lockres->l_lock, flags);
3441
3442		/* XXX: Today we just wait on any busy
3443		 * locks... Perhaps we need to cancel converts in the
3444		 * future? */
3445		ocfs2_wait_on_busy_lock(lockres);
3446
3447		spin_lock_irqsave(&lockres->l_lock, flags);
3448	}
3449
3450	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3451		if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
3452		    lockres->l_level == DLM_LOCK_EX &&
3453		    !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3454			lockres->l_ops->set_lvb(lockres);
3455	}
3456
3457	if (lockres->l_flags & OCFS2_LOCK_BUSY)
3458		mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
3459		     lockres->l_name);
3460	if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3461		mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
3462
3463	if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
3464		spin_unlock_irqrestore(&lockres->l_lock, flags);
3465		goto out;
3466	}
3467
3468	lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
3469
3470	/* make sure we never get here while waiting for an ast to
3471	 * fire. */
3472	BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
3473
3474	/* is this necessary? */
3475	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3476	lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
3477	spin_unlock_irqrestore(&lockres->l_lock, flags);
3478
3479	mlog(0, "lock %s\n", lockres->l_name);
3480
3481	ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags);
3482	if (ret) {
3483		ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3484		mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
3485		ocfs2_dlm_dump_lksb(&lockres->l_lksb);
3486		BUG();
3487	}
3488	mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3489	     lockres->l_name);
3490
3491	ocfs2_wait_on_busy_lock(lockres);
3492out:
3493	return 0;
3494}
3495
3496static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3497				       struct ocfs2_lock_res *lockres);
3498
3499/* Mark the lockres as being dropped. It will no longer be
3500 * queued if blocking, but we still may have to wait on it
3501 * being dequeued from the downconvert thread before we can consider
3502 * it safe to drop.
3503 *
3504 * You can *not* attempt to call cluster_lock on this lockres anymore. */
3505void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb,
3506				struct ocfs2_lock_res *lockres)
3507{
3508	int status;
3509	struct ocfs2_mask_waiter mw;
3510	unsigned long flags, flags2;
3511
3512	ocfs2_init_mask_waiter(&mw);
3513
3514	spin_lock_irqsave(&lockres->l_lock, flags);
3515	lockres->l_flags |= OCFS2_LOCK_FREEING;
3516	if (lockres->l_flags & OCFS2_LOCK_QUEUED && current == osb->dc_task) {
3517		/*
3518		 * We know the downconvert is queued but not in progress
3519		 * because we are the downconvert thread and processing
3520		 * different lock. So we can just remove the lock from the
3521		 * queue. This is not only an optimization but also a way
3522		 * to avoid the following deadlock:
3523		 *   ocfs2_dentry_post_unlock()
3524		 *     ocfs2_dentry_lock_put()
3525		 *       ocfs2_drop_dentry_lock()
3526		 *         iput()
3527		 *           ocfs2_evict_inode()
3528		 *             ocfs2_clear_inode()
3529		 *               ocfs2_mark_lockres_freeing()
3530		 *                 ... blocks waiting for OCFS2_LOCK_QUEUED
3531		 *                 since we are the downconvert thread which
3532		 *                 should clear the flag.
3533		 */
3534		spin_unlock_irqrestore(&lockres->l_lock, flags);
3535		spin_lock_irqsave(&osb->dc_task_lock, flags2);
3536		list_del_init(&lockres->l_blocked_list);
3537		osb->blocked_lock_count--;
3538		spin_unlock_irqrestore(&osb->dc_task_lock, flags2);
3539		/*
3540		 * Warn if we recurse into another post_unlock call.  Strictly
3541		 * speaking it isn't a problem but we need to be careful if
3542		 * that happens (stack overflow, deadlocks, ...) so warn if
3543		 * ocfs2 grows a path for which this can happen.
3544		 */
3545		WARN_ON_ONCE(lockres->l_ops->post_unlock);
3546		/* Since the lock is freeing we don't do much in the fn below */
3547		ocfs2_process_blocked_lock(osb, lockres);
3548		return;
3549	}
3550	while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
3551		lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
3552		spin_unlock_irqrestore(&lockres->l_lock, flags);
3553
3554		mlog(0, "Waiting on lockres %s\n", lockres->l_name);
3555
3556		status = ocfs2_wait_for_mask(&mw);
3557		if (status)
3558			mlog_errno(status);
3559
3560		spin_lock_irqsave(&lockres->l_lock, flags);
3561	}
3562	spin_unlock_irqrestore(&lockres->l_lock, flags);
3563}
3564
3565void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
3566			       struct ocfs2_lock_res *lockres)
3567{
3568	int ret;
3569
3570	ocfs2_mark_lockres_freeing(osb, lockres);
3571	ret = ocfs2_drop_lock(osb, lockres);
3572	if (ret)
3573		mlog_errno(ret);
3574}
3575
3576static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
3577{
3578	ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
3579	ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
3580	ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres);
3581	ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres);
3582}
3583
3584int ocfs2_drop_inode_locks(struct inode *inode)
3585{
3586	int status, err;
3587
3588	/* No need to call ocfs2_mark_lockres_freeing here -
3589	 * ocfs2_clear_inode has done it for us. */
3590
3591	err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3592			      &OCFS2_I(inode)->ip_open_lockres);
3593	if (err < 0)
3594		mlog_errno(err);
3595
3596	status = err;
3597
3598	err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3599			      &OCFS2_I(inode)->ip_inode_lockres);
3600	if (err < 0)
3601		mlog_errno(err);
3602	if (err < 0 && !status)
3603		status = err;
3604
3605	err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3606			      &OCFS2_I(inode)->ip_rw_lockres);
3607	if (err < 0)
3608		mlog_errno(err);
3609	if (err < 0 && !status)
3610		status = err;
3611
3612	return status;
3613}
3614
3615static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
3616					      int new_level)
3617{
3618	assert_spin_locked(&lockres->l_lock);
3619
3620	BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
3621
3622	if (lockres->l_level <= new_level) {
3623		mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
3624		     "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
3625		     "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
3626		     new_level, list_empty(&lockres->l_blocked_list),
3627		     list_empty(&lockres->l_mask_waiters), lockres->l_type,
3628		     lockres->l_flags, lockres->l_ro_holders,
3629		     lockres->l_ex_holders, lockres->l_action,
3630		     lockres->l_unlock_action, lockres->l_requested,
3631		     lockres->l_blocking, lockres->l_pending_gen);
3632		BUG();
3633	}
3634
3635	mlog(ML_BASTS, "lockres %s, level %d => %d, blocking %d\n",
3636	     lockres->l_name, lockres->l_level, new_level, lockres->l_blocking);
3637
3638	lockres->l_action = OCFS2_AST_DOWNCONVERT;
3639	lockres->l_requested = new_level;
3640	lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3641	return lockres_set_pending(lockres);
3642}
3643
3644static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
3645				  struct ocfs2_lock_res *lockres,
3646				  int new_level,
3647				  int lvb,
3648				  unsigned int generation)
3649{
3650	int ret;
3651	u32 dlm_flags = DLM_LKF_CONVERT;
3652
3653	mlog(ML_BASTS, "lockres %s, level %d => %d\n", lockres->l_name,
3654	     lockres->l_level, new_level);
3655
3656	/*
3657	 * On DLM_LKF_VALBLK, fsdlm behaves differently with o2cb. It always
3658	 * expects DLM_LKF_VALBLK being set if the LKB has LVB, so that
3659	 * we can recover correctly from node failure. Otherwise, we may get
3660	 * invalid LVB in LKB, but without DLM_SBF_VALNOTVALID being set.
3661	 */
3662	if (ocfs2_userspace_stack(osb) &&
3663	    lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3664		lvb = 1;
3665
3666	if (lvb)
3667		dlm_flags |= DLM_LKF_VALBLK;
3668
3669	ret = ocfs2_dlm_lock(osb->cconn,
3670			     new_level,
3671			     &lockres->l_lksb,
3672			     dlm_flags,
3673			     lockres->l_name,
3674			     OCFS2_LOCK_ID_MAX_LEN - 1);
3675	lockres_clear_pending(lockres, generation, osb);
3676	if (ret) {
3677		ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
3678		ocfs2_recover_from_dlm_error(lockres, 1);
3679		goto bail;
3680	}
3681
3682	ret = 0;
3683bail:
3684	return ret;
3685}
3686
3687/* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3688static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
3689				        struct ocfs2_lock_res *lockres)
3690{
3691	assert_spin_locked(&lockres->l_lock);
3692
3693	if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
3694		/* If we're already trying to cancel a lock conversion
3695		 * then just drop the spinlock and allow the caller to
3696		 * requeue this lock. */
3697		mlog(ML_BASTS, "lockres %s, skip convert\n", lockres->l_name);
3698		return 0;
3699	}
3700
3701	/* were we in a convert when we got the bast fire? */
3702	BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
3703	       lockres->l_action != OCFS2_AST_DOWNCONVERT);
3704	/* set things up for the unlockast to know to just
3705	 * clear out the ast_action and unset busy, etc. */
3706	lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
3707
3708	mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
3709			"lock %s, invalid flags: 0x%lx\n",
3710			lockres->l_name, lockres->l_flags);
3711
3712	mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3713
3714	return 1;
3715}
3716
3717static int ocfs2_cancel_convert(struct ocfs2_super *osb,
3718				struct ocfs2_lock_res *lockres)
3719{
3720	int ret;
3721
3722	ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb,
3723			       DLM_LKF_CANCEL);
3724	if (ret) {
3725		ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3726		ocfs2_recover_from_dlm_error(lockres, 0);
3727	}
3728
3729	mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3730
3731	return ret;
3732}
3733
3734static int ocfs2_unblock_lock(struct ocfs2_super *osb,
3735			      struct ocfs2_lock_res *lockres,
3736			      struct ocfs2_unblock_ctl *ctl)
3737{
3738	unsigned long flags;
3739	int blocking;
3740	int new_level;
3741	int level;
3742	int ret = 0;
3743	int set_lvb = 0;
3744	unsigned int gen;
3745
3746	spin_lock_irqsave(&lockres->l_lock, flags);
3747
3748recheck:
3749	/*
3750	 * Is it still blocking? If not, we have no more work to do.
3751	 */
3752	if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) {
3753		BUG_ON(lockres->l_blocking != DLM_LOCK_NL);
3754		spin_unlock_irqrestore(&lockres->l_lock, flags);
3755		ret = 0;
3756		goto leave;
3757	}
3758
3759	if (lockres->l_flags & OCFS2_LOCK_BUSY) {
3760		/* XXX
3761		 * This is a *big* race.  The OCFS2_LOCK_PENDING flag
3762		 * exists entirely for one reason - another thread has set
3763		 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3764		 *
3765		 * If we do ocfs2_cancel_convert() before the other thread
3766		 * calls dlm_lock(), our cancel will do nothing.  We will
3767		 * get no ast, and we will have no way of knowing the
3768		 * cancel failed.  Meanwhile, the other thread will call
3769		 * into dlm_lock() and wait...forever.
3770		 *
3771		 * Why forever?  Because another node has asked for the
3772		 * lock first; that's why we're here in unblock_lock().
3773		 *
3774		 * The solution is OCFS2_LOCK_PENDING.  When PENDING is
3775		 * set, we just requeue the unblock.  Only when the other
3776		 * thread has called dlm_lock() and cleared PENDING will
3777		 * we then cancel their request.
3778		 *
3779		 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3780		 * at the same time they set OCFS2_DLM_BUSY.  They must
3781		 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3782		 */
3783		if (lockres->l_flags & OCFS2_LOCK_PENDING) {
3784			mlog(ML_BASTS, "lockres %s, ReQ: Pending\n",
3785			     lockres->l_name);
3786			goto leave_requeue;
3787		}
3788
3789		ctl->requeue = 1;
3790		ret = ocfs2_prepare_cancel_convert(osb, lockres);
3791		spin_unlock_irqrestore(&lockres->l_lock, flags);
3792		if (ret) {
3793			ret = ocfs2_cancel_convert(osb, lockres);
3794			if (ret < 0)
3795				mlog_errno(ret);
3796		}
3797		goto leave;
3798	}
3799
3800	/*
3801	 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3802	 * set when the ast is received for an upconvert just before the
3803	 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3804	 * on the heels of the ast, we want to delay the downconvert just
3805	 * enough to allow the up requestor to do its task. Because this
3806	 * lock is in the blocked queue, the lock will be downconverted
3807	 * as soon as the requestor is done with the lock.
3808	 */
3809	if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
3810		goto leave_requeue;
3811
3812	/*
3813	 * How can we block and yet be at NL?  We were trying to upconvert
3814	 * from NL and got canceled.  The code comes back here, and now
3815	 * we notice and clear BLOCKING.
3816	 */
3817	if (lockres->l_level == DLM_LOCK_NL) {
3818		BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders);
3819		mlog(ML_BASTS, "lockres %s, Aborting dc\n", lockres->l_name);
3820		lockres->l_blocking = DLM_LOCK_NL;
3821		lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
3822		spin_unlock_irqrestore(&lockres->l_lock, flags);
3823		goto leave;
3824	}
3825
3826	/* if we're blocking an exclusive and we have *any* holders,
3827	 * then requeue. */
3828	if ((lockres->l_blocking == DLM_LOCK_EX)
3829	    && (lockres->l_ex_holders || lockres->l_ro_holders)) {
3830		mlog(ML_BASTS, "lockres %s, ReQ: EX/PR Holders %u,%u\n",
3831		     lockres->l_name, lockres->l_ex_holders,
3832		     lockres->l_ro_holders);
3833		goto leave_requeue;
3834	}
3835
3836	/* If it's a PR we're blocking, then only
3837	 * requeue if we've got any EX holders */
3838	if (lockres->l_blocking == DLM_LOCK_PR &&
3839	    lockres->l_ex_holders) {
3840		mlog(ML_BASTS, "lockres %s, ReQ: EX Holders %u\n",
3841		     lockres->l_name, lockres->l_ex_holders);
3842		goto leave_requeue;
3843	}
3844
3845	/*
3846	 * Can we get a lock in this state if the holder counts are
3847	 * zero? The meta data unblock code used to check this.
3848	 */
3849	if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
3850	    && (lockres->l_flags & OCFS2_LOCK_REFRESHING)) {
3851		mlog(ML_BASTS, "lockres %s, ReQ: Lock Refreshing\n",
3852		     lockres->l_name);
3853		goto leave_requeue;
3854	}
3855
3856	new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
3857
3858	if (lockres->l_ops->check_downconvert
3859	    && !lockres->l_ops->check_downconvert(lockres, new_level)) {
3860		mlog(ML_BASTS, "lockres %s, ReQ: Checkpointing\n",
3861		     lockres->l_name);
3862		goto leave_requeue;
3863	}
3864
3865	/* If we get here, then we know that there are no more
3866	 * incompatible holders (and anyone asking for an incompatible
3867	 * lock is blocked). We can now downconvert the lock */
3868	if (!lockres->l_ops->downconvert_worker)
3869		goto downconvert;
3870
3871	/* Some lockres types want to do a bit of work before
3872	 * downconverting a lock. Allow that here. The worker function
3873	 * may sleep, so we save off a copy of what we're blocking as
3874	 * it may change while we're not holding the spin lock. */
3875	blocking = lockres->l_blocking;
3876	level = lockres->l_level;
3877	spin_unlock_irqrestore(&lockres->l_lock, flags);
3878
3879	ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
3880
3881	if (ctl->unblock_action == UNBLOCK_STOP_POST) {
3882		mlog(ML_BASTS, "lockres %s, UNBLOCK_STOP_POST\n",
3883		     lockres->l_name);
3884		goto leave;
3885	}
3886
3887	spin_lock_irqsave(&lockres->l_lock, flags);
3888	if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) {
3889		/* If this changed underneath us, then we can't drop
3890		 * it just yet. */
3891		mlog(ML_BASTS, "lockres %s, block=%d:%d, level=%d:%d, "
3892		     "Recheck\n", lockres->l_name, blocking,
3893		     lockres->l_blocking, level, lockres->l_level);
3894		goto recheck;
3895	}
3896
3897downconvert:
3898	ctl->requeue = 0;
3899
3900	if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3901		if (lockres->l_level == DLM_LOCK_EX)
3902			set_lvb = 1;
3903
3904		/*
3905		 * We only set the lvb if the lock has been fully
3906		 * refreshed - otherwise we risk setting stale
3907		 * data. Otherwise, there's no need to actually clear
3908		 * out the lvb here as it's value is still valid.
3909		 */
3910		if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3911			lockres->l_ops->set_lvb(lockres);
3912	}
3913
3914	gen = ocfs2_prepare_downconvert(lockres, new_level);
3915	spin_unlock_irqrestore(&lockres->l_lock, flags);
3916	ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
3917				     gen);
3918	/* The dlm lock convert is being cancelled in background,
3919	 * ocfs2_cancel_convert() is asynchronous in fs/dlm,
3920	 * requeue it, try again later.
3921	 */
3922	if (ret == -EBUSY) {
3923		ctl->requeue = 1;
3924		mlog(ML_BASTS, "lockres %s, ReQ: Downconvert busy\n",
3925		     lockres->l_name);
3926		ret = 0;
3927		msleep(20);
3928	}
3929
3930leave:
3931	if (ret)
3932		mlog_errno(ret);
3933	return ret;
3934
3935leave_requeue:
3936	spin_unlock_irqrestore(&lockres->l_lock, flags);
3937	ctl->requeue = 1;
3938
3939	return 0;
3940}
3941
3942static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
3943				     int blocking)
3944{
3945	struct inode *inode;
3946	struct address_space *mapping;
3947	struct ocfs2_inode_info *oi;
3948
3949       	inode = ocfs2_lock_res_inode(lockres);
3950	mapping = inode->i_mapping;
3951
3952	if (S_ISDIR(inode->i_mode)) {
3953		oi = OCFS2_I(inode);
3954		oi->ip_dir_lock_gen++;
3955		mlog(0, "generation: %u\n", oi->ip_dir_lock_gen);
3956		goto out_forget;
3957	}
3958
3959	if (!S_ISREG(inode->i_mode))
3960		goto out;
3961
3962	/*
3963	 * We need this before the filemap_fdatawrite() so that it can
3964	 * transfer the dirty bit from the PTE to the
3965	 * page. Unfortunately this means that even for EX->PR
3966	 * downconverts, we'll lose our mappings and have to build
3967	 * them up again.
3968	 */
3969	unmap_mapping_range(mapping, 0, 0, 0);
3970
3971	if (filemap_fdatawrite(mapping)) {
3972		mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
3973		     (unsigned long long)OCFS2_I(inode)->ip_blkno);
3974	}
3975	sync_mapping_buffers(mapping);
3976	if (blocking == DLM_LOCK_EX) {
3977		truncate_inode_pages(mapping, 0);
3978	} else {
3979		/* We only need to wait on the I/O if we're not also
3980		 * truncating pages because truncate_inode_pages waits
3981		 * for us above. We don't truncate pages if we're
3982		 * blocking anything < EXMODE because we want to keep
3983		 * them around in that case. */
3984		filemap_fdatawait(mapping);
3985	}
3986
3987out_forget:
3988	forget_all_cached_acls(inode);
3989
3990out:
3991	return UNBLOCK_CONTINUE;
3992}
3993
3994static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci,
3995				 struct ocfs2_lock_res *lockres,
3996				 int new_level)
3997{
3998	int checkpointed = ocfs2_ci_fully_checkpointed(ci);
3999
4000	BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR);
4001	BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed);
4002
4003	if (checkpointed)
4004		return 1;
4005
4006	ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci)));
4007	return 0;
4008}
4009
4010static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
4011					int new_level)
4012{
4013	struct inode *inode = ocfs2_lock_res_inode(lockres);
4014
4015	return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level);
4016}
4017
4018static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
4019{
4020	struct inode *inode = ocfs2_lock_res_inode(lockres);
4021
4022	__ocfs2_stuff_meta_lvb(inode);
4023}
4024
4025/*
4026 * Does the final reference drop on our dentry lock. Right now this
4027 * happens in the downconvert thread, but we could choose to simplify the
4028 * dlmglue API and push these off to the ocfs2_wq in the future.
4029 */
4030static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
4031				     struct ocfs2_lock_res *lockres)
4032{
4033	struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
4034	ocfs2_dentry_lock_put(osb, dl);
4035}
4036
4037/*
4038 * d_delete() matching dentries before the lock downconvert.
4039 *
4040 * At this point, any process waiting to destroy the
4041 * dentry_lock due to last ref count is stopped by the
4042 * OCFS2_LOCK_QUEUED flag.
4043 *
4044 * We have two potential problems
4045 *
4046 * 1) If we do the last reference drop on our dentry_lock (via dput)
4047 *    we'll wind up in ocfs2_release_dentry_lock(), waiting on
4048 *    the downconvert to finish. Instead we take an elevated
4049 *    reference and push the drop until after we've completed our
4050 *    unblock processing.
4051 *
4052 * 2) There might be another process with a final reference,
4053 *    waiting on us to finish processing. If this is the case, we
4054 *    detect it and exit out - there's no more dentries anyway.
4055 */
4056static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
4057				       int blocking)
4058{
4059	struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
4060	struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
4061	struct dentry *dentry;
4062	unsigned long flags;
4063	int extra_ref = 0;
4064
4065	/*
4066	 * This node is blocking another node from getting a read
4067	 * lock. This happens when we've renamed within a
4068	 * directory. We've forced the other nodes to d_delete(), but
4069	 * we never actually dropped our lock because it's still
4070	 * valid. The downconvert code will retain a PR for this node,
4071	 * so there's no further work to do.
4072	 */
4073	if (blocking == DLM_LOCK_PR)
4074		return UNBLOCK_CONTINUE;
4075
4076	/*
4077	 * Mark this inode as potentially orphaned. The code in
4078	 * ocfs2_delete_inode() will figure out whether it actually
4079	 * needs to be freed or not.
4080	 */
4081	spin_lock(&oi->ip_lock);
4082	oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
4083	spin_unlock(&oi->ip_lock);
4084
4085	/*
4086	 * Yuck. We need to make sure however that the check of
4087	 * OCFS2_LOCK_FREEING and the extra reference are atomic with
4088	 * respect to a reference decrement or the setting of that
4089	 * flag.
4090	 */
4091	spin_lock_irqsave(&lockres->l_lock, flags);
4092	spin_lock(&dentry_attach_lock);
4093	if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
4094	    && dl->dl_count) {
4095		dl->dl_count++;
4096		extra_ref = 1;
4097	}
4098	spin_unlock(&dentry_attach_lock);
4099	spin_unlock_irqrestore(&lockres->l_lock, flags);
4100
4101	mlog(0, "extra_ref = %d\n", extra_ref);
4102
4103	/*
4104	 * We have a process waiting on us in ocfs2_dentry_iput(),
4105	 * which means we can't have any more outstanding
4106	 * aliases. There's no need to do any more work.
4107	 */
4108	if (!extra_ref)
4109		return UNBLOCK_CONTINUE;
4110
4111	spin_lock(&dentry_attach_lock);
4112	while (1) {
4113		dentry = ocfs2_find_local_alias(dl->dl_inode,
4114						dl->dl_parent_blkno, 1);
4115		if (!dentry)
4116			break;
4117		spin_unlock(&dentry_attach_lock);
4118
4119		if (S_ISDIR(dl->dl_inode->i_mode))
4120			shrink_dcache_parent(dentry);
4121
4122		mlog(0, "d_delete(%pd);\n", dentry);
4123
4124		/*
4125		 * The following dcache calls may do an
4126		 * iput(). Normally we don't want that from the
4127		 * downconverting thread, but in this case it's ok
4128		 * because the requesting node already has an
4129		 * exclusive lock on the inode, so it can't be queued
4130		 * for a downconvert.
4131		 */
4132		d_delete(dentry);
4133		dput(dentry);
4134
4135		spin_lock(&dentry_attach_lock);
4136	}
4137	spin_unlock(&dentry_attach_lock);
4138
4139	/*
4140	 * If we are the last holder of this dentry lock, there is no
4141	 * reason to downconvert so skip straight to the unlock.
4142	 */
4143	if (dl->dl_count == 1)
4144		return UNBLOCK_STOP_POST;
4145
4146	return UNBLOCK_CONTINUE_POST;
4147}
4148
4149static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
4150					    int new_level)
4151{
4152	struct ocfs2_refcount_tree *tree =
4153				ocfs2_lock_res_refcount_tree(lockres);
4154
4155	return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level);
4156}
4157
4158static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
4159					 int blocking)
4160{
4161	struct ocfs2_refcount_tree *tree =
4162				ocfs2_lock_res_refcount_tree(lockres);
4163
4164	ocfs2_metadata_cache_purge(&tree->rf_ci);
4165
4166	return UNBLOCK_CONTINUE;
4167}
4168
4169static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres)
4170{
4171	struct ocfs2_qinfo_lvb *lvb;
4172	struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres);
4173	struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
4174					    oinfo->dqi_gi.dqi_type);
4175
4176	lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
4177	lvb->lvb_version = OCFS2_QINFO_LVB_VERSION;
4178	lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace);
4179	lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace);
4180	lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms);
4181	lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks);
4182	lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk);
4183	lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry);
4184}
4185
4186void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex)
4187{
4188	struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
4189	struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
4190	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4191
4192	if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
4193		ocfs2_cluster_unlock(osb, lockres, level);
4194}
4195
4196static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo)
4197{
4198	struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
4199					    oinfo->dqi_gi.dqi_type);
4200	struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
4201	struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
4202	struct buffer_head *bh = NULL;
4203	struct ocfs2_global_disk_dqinfo *gdinfo;
4204	int status = 0;
4205
4206	if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
4207	    lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) {
4208		info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace);
4209		info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace);
4210		oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms);
4211		oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks);
4212		oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk);
4213		oinfo->dqi_gi.dqi_free_entry =
4214					be32_to_cpu(lvb->lvb_free_entry);
4215	} else {
4216		status = ocfs2_read_quota_phys_block(oinfo->dqi_gqinode,
4217						     oinfo->dqi_giblk, &bh);
4218		if (status) {
4219			mlog_errno(status);
4220			goto bail;
4221		}
4222		gdinfo = (struct ocfs2_global_disk_dqinfo *)
4223					(bh->b_data + OCFS2_GLOBAL_INFO_OFF);
4224		info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace);
4225		info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace);
4226		oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms);
4227		oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks);
4228		oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk);
4229		oinfo->dqi_gi.dqi_free_entry =
4230					le32_to_cpu(gdinfo->dqi_free_entry);
4231		brelse(bh);
4232		ocfs2_track_lock_refresh(lockres);
4233	}
4234
4235bail:
4236	return status;
4237}
4238
4239/* Lock quota info, this function expects at least shared lock on the quota file
4240 * so that we can safely refresh quota info from disk. */
4241int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex)
4242{
4243	struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
4244	struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
4245	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4246	int status = 0;
4247
4248	/* On RO devices, locking really isn't needed... */
4249	if (ocfs2_is_hard_readonly(osb)) {
4250		if (ex)
4251			status = -EROFS;
4252		goto bail;
4253	}
4254	if (ocfs2_mount_local(osb))
4255		goto bail;
4256
4257	status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
4258	if (status < 0) {
4259		mlog_errno(status);
4260		goto bail;
4261	}
4262	if (!ocfs2_should_refresh_lock_res(lockres))
4263		goto bail;
4264	/* OK, we have the lock but we need to refresh the quota info */
4265	status = ocfs2_refresh_qinfo(oinfo);
4266	if (status)
4267		ocfs2_qinfo_unlock(oinfo, ex);
4268	ocfs2_complete_lock_res_refresh(lockres, status);
4269bail:
4270	return status;
4271}
4272
4273int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex)
4274{
4275	int status;
4276	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4277	struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
4278	struct ocfs2_super *osb = lockres->l_priv;
4279
4280
4281	if (ocfs2_is_hard_readonly(osb))
4282		return -EROFS;
4283
4284	if (ocfs2_mount_local(osb))
4285		return 0;
4286
4287	status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
4288	if (status < 0)
4289		mlog_errno(status);
4290
4291	return status;
4292}
4293
4294void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex)
4295{
4296	int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4297	struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
4298	struct ocfs2_super *osb = lockres->l_priv;
4299
4300	if (!ocfs2_mount_local(osb))
4301		ocfs2_cluster_unlock(osb, lockres, level);
4302}
4303
4304static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
4305				       struct ocfs2_lock_res *lockres)
4306{
4307	int status;
4308	struct ocfs2_unblock_ctl ctl = {0, 0,};
4309	unsigned long flags;
4310
4311	/* Our reference to the lockres in this function can be
4312	 * considered valid until we remove the OCFS2_LOCK_QUEUED
4313	 * flag. */
4314
4315	BUG_ON(!lockres);
4316	BUG_ON(!lockres->l_ops);
4317
4318	mlog(ML_BASTS, "lockres %s blocked\n", lockres->l_name);
4319
4320	/* Detect whether a lock has been marked as going away while
4321	 * the downconvert thread was processing other things. A lock can
4322	 * still be marked with OCFS2_LOCK_FREEING after this check,
4323	 * but short circuiting here will still save us some
4324	 * performance. */
4325	spin_lock_irqsave(&lockres->l_lock, flags);
4326	if (lockres->l_flags & OCFS2_LOCK_FREEING)
4327		goto unqueue;
4328	spin_unlock_irqrestore(&lockres->l_lock, flags);
4329
4330	status = ocfs2_unblock_lock(osb, lockres, &ctl);
4331	if (status < 0)
4332		mlog_errno(status);
4333
4334	spin_lock_irqsave(&lockres->l_lock, flags);
4335unqueue:
4336	if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
4337		lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
4338	} else
4339		ocfs2_schedule_blocked_lock(osb, lockres);
4340
4341	mlog(ML_BASTS, "lockres %s, requeue = %s.\n", lockres->l_name,
4342	     ctl.requeue ? "yes" : "no");
4343	spin_unlock_irqrestore(&lockres->l_lock, flags);
4344
4345	if (ctl.unblock_action != UNBLOCK_CONTINUE
4346	    && lockres->l_ops->post_unlock)
4347		lockres->l_ops->post_unlock(osb, lockres);
4348}
4349
4350static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
4351					struct ocfs2_lock_res *lockres)
4352{
4353	unsigned long flags;
4354
4355	assert_spin_locked(&lockres->l_lock);
4356
4357	if (lockres->l_flags & OCFS2_LOCK_FREEING) {
4358		/* Do not schedule a lock for downconvert when it's on
4359		 * the way to destruction - any nodes wanting access
4360		 * to the resource will get it soon. */
4361		mlog(ML_BASTS, "lockres %s won't be scheduled: flags 0x%lx\n",
4362		     lockres->l_name, lockres->l_flags);
4363		return;
4364	}
4365
4366	lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
4367
4368	spin_lock_irqsave(&osb->dc_task_lock, flags);
4369	if (list_empty(&lockres->l_blocked_list)) {
4370		list_add_tail(&lockres->l_blocked_list,
4371			      &osb->blocked_lock_list);
4372		osb->blocked_lock_count++;
4373	}
4374	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4375}
4376
4377static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb)
4378{
4379	unsigned long processed;
4380	unsigned long flags;
4381	struct ocfs2_lock_res *lockres;
4382
4383	spin_lock_irqsave(&osb->dc_task_lock, flags);
4384	/* grab this early so we know to try again if a state change and
4385	 * wake happens part-way through our work  */
4386	osb->dc_work_sequence = osb->dc_wake_sequence;
4387
4388	processed = osb->blocked_lock_count;
4389	/*
4390	 * blocked lock processing in this loop might call iput which can
4391	 * remove items off osb->blocked_lock_list. Downconvert up to
4392	 * 'processed' number of locks, but stop short if we had some
4393	 * removed in ocfs2_mark_lockres_freeing when downconverting.
4394	 */
4395	while (processed && !list_empty(&osb->blocked_lock_list)) {
4396		lockres = list_entry(osb->blocked_lock_list.next,
4397				     struct ocfs2_lock_res, l_blocked_list);
4398		list_del_init(&lockres->l_blocked_list);
4399		osb->blocked_lock_count--;
4400		spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4401
4402		BUG_ON(!processed);
4403		processed--;
4404
4405		ocfs2_process_blocked_lock(osb, lockres);
4406
4407		spin_lock_irqsave(&osb->dc_task_lock, flags);
4408	}
4409	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4410}
4411
4412static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb)
4413{
4414	int empty = 0;
4415	unsigned long flags;
4416
4417	spin_lock_irqsave(&osb->dc_task_lock, flags);
4418	if (list_empty(&osb->blocked_lock_list))
4419		empty = 1;
4420
4421	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4422	return empty;
4423}
4424
4425static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb)
4426{
4427	int should_wake = 0;
4428	unsigned long flags;
4429
4430	spin_lock_irqsave(&osb->dc_task_lock, flags);
4431	if (osb->dc_work_sequence != osb->dc_wake_sequence)
4432		should_wake = 1;
4433	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4434
4435	return should_wake;
4436}
4437
4438static int ocfs2_downconvert_thread(void *arg)
4439{
 
4440	struct ocfs2_super *osb = arg;
4441
4442	/* only quit once we've been asked to stop and there is no more
4443	 * work available */
4444	while (!(kthread_should_stop() &&
4445		ocfs2_downconvert_thread_lists_empty(osb))) {
4446
4447		wait_event_interruptible(osb->dc_event,
4448					 ocfs2_downconvert_thread_should_wake(osb) ||
4449					 kthread_should_stop());
4450
4451		mlog(0, "downconvert_thread: awoken\n");
4452
4453		ocfs2_downconvert_thread_do_work(osb);
4454	}
4455
4456	osb->dc_task = NULL;
4457	return 0;
4458}
4459
4460void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb)
4461{
4462	unsigned long flags;
4463
4464	spin_lock_irqsave(&osb->dc_task_lock, flags);
4465	/* make sure the voting thread gets a swipe at whatever changes
4466	 * the caller may have made to the voting state */
4467	osb->dc_wake_sequence++;
4468	spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4469	wake_up(&osb->dc_event);
4470}