fs/dlm/lock.c

/******************************************************************************
*******************************************************************************
**
**  Copyright (C) 2005-2010 Red Hat, Inc.  All rights reserved.
**
**  This copyrighted material is made available to anyone wishing to use,
**  modify, copy, or redistribute it subject to the terms and conditions
**  of the GNU General Public License v.2.
**
*******************************************************************************
******************************************************************************/

/* Central locking logic has four stages:

   dlm_lock()
   dlm_unlock()

   request_lock(ls, lkb)
   convert_lock(ls, lkb)
   unlock_lock(ls, lkb)
   cancel_lock(ls, lkb)

   _request_lock(r, lkb)
   _convert_lock(r, lkb)
   _unlock_lock(r, lkb)
   _cancel_lock(r, lkb)

   do_request(r, lkb)
   do_convert(r, lkb)
   do_unlock(r, lkb)
   do_cancel(r, lkb)

   Stage 1 (lock, unlock) is mainly about checking input args and
   splitting into one of the four main operations:

       dlm_lock          = request_lock
       dlm_lock+CONVERT  = convert_lock
       dlm_unlock        = unlock_lock
       dlm_unlock+CANCEL = cancel_lock

   Stage 2, xxxx_lock(), just finds and locks the relevant rsb which is
   provided to the next stage.

   Stage 3, _xxxx_lock(), determines if the operation is local or remote.
   When remote, it calls send_xxxx(), when local it calls do_xxxx().

   Stage 4, do_xxxx(), is the guts of the operation.  It manipulates the
   given rsb and lkb and queues callbacks.

   For remote operations, send_xxxx() results in the corresponding do_xxxx()
   function being executed on the remote node.  The connecting send/receive
   calls on local (L) and remote (R) nodes:

   L: send_xxxx()              ->  R: receive_xxxx()
                                   R: do_xxxx()
   L: receive_xxxx_reply()     <-  R: send_xxxx_reply()
*/
#include <linux/types.h>
#include <linux/rbtree.h>
#include <linux/slab.h>
#include "dlm_internal.h"
#include <linux/dlm_device.h>
#include "memory.h"
#include "lowcomms.h"
#include "requestqueue.h"
#include "util.h"
#include "dir.h"
#include "member.h"
#include "lockspace.h"
#include "ast.h"
#include "lock.h"
#include "rcom.h"
#include "recover.h"
#include "lvb_table.h"
#include "user.h"
#include "config.h"

static int send_request(struct dlm_rsb *r, struct dlm_lkb *lkb);
static int send_convert(struct dlm_rsb *r, struct dlm_lkb *lkb);
static int send_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb);
static int send_cancel(struct dlm_rsb *r, struct dlm_lkb *lkb);
static int send_grant(struct dlm_rsb *r, struct dlm_lkb *lkb);
static int send_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int mode);
static int send_lookup(struct dlm_rsb *r, struct dlm_lkb *lkb);
static int send_remove(struct dlm_rsb *r);
static int _request_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
static int _cancel_lock(struct dlm_rsb *r, struct dlm_lkb *lkb);
static void __receive_convert_reply(struct dlm_rsb *r, struct dlm_lkb *lkb,
				    struct dlm_message *ms);
static int receive_extralen(struct dlm_message *ms);
static void do_purge(struct dlm_ls *ls, int nodeid, int pid);
static void del_timeout(struct dlm_lkb *lkb);
static void toss_rsb(struct kref *kref);

/*
 * Lock compatibilty matrix - thanks Steve
 * UN = Unlocked state. Not really a state, used as a flag
 * PD = Padding. Used to make the matrix a nice power of two in size
 * Other states are the same as the VMS DLM.
 * Usage: matrix[grmode+1][rqmode+1]  (although m[rq+1][gr+1] is the same)
 */

static const int __dlm_compat_matrix[8][8] = {
      /* UN NL CR CW PR PW EX PD */
        {1, 1, 1, 1, 1, 1, 1, 0},       /* UN */
        {1, 1, 1, 1, 1, 1, 1, 0},       /* NL */
        {1, 1, 1, 1, 1, 1, 0, 0},       /* CR */
        {1, 1, 1, 1, 0, 0, 0, 0},       /* CW */
        {1, 1, 1, 0, 1, 0, 0, 0},       /* PR */
        {1, 1, 1, 0, 0, 0, 0, 0},       /* PW */
        {1, 1, 0, 0, 0, 0, 0, 0},       /* EX */
        {0, 0, 0, 0, 0, 0, 0, 0}        /* PD */
};

/*
 * This defines the direction of transfer of LVB data.
 * Granted mode is the row; requested mode is the column.
 * Usage: matrix[grmode+1][rqmode+1]
 * 1 = LVB is returned to the caller
 * 0 = LVB is written to the resource
 * -1 = nothing happens to the LVB
 */

const int dlm_lvb_operations[8][8] = {
        /* UN   NL  CR  CW  PR  PW  EX  PD*/
        {  -1,  1,  1,  1,  1,  1,  1, -1 }, /* UN */
        {  -1,  1,  1,  1,  1,  1,  1,  0 }, /* NL */
        {  -1, -1,  1,  1,  1,  1,  1,  0 }, /* CR */
        {  -1, -1, -1,  1,  1,  1,  1,  0 }, /* CW */
        {  -1, -1, -1, -1,  1,  1,  1,  0 }, /* PR */
        {  -1,  0,  0,  0,  0,  0,  1,  0 }, /* PW */
        {  -1,  0,  0,  0,  0,  0,  0,  0 }, /* EX */
        {  -1,  0,  0,  0,  0,  0,  0,  0 }  /* PD */
};

#define modes_compat(gr, rq) \
	__dlm_compat_matrix[(gr)->lkb_grmode + 1][(rq)->lkb_rqmode + 1]

int dlm_modes_compat(int mode1, int mode2)
{
	return __dlm_compat_matrix[mode1 + 1][mode2 + 1];
}

/*
 * Compatibility matrix for conversions with QUECVT set.
 * Granted mode is the row; requested mode is the column.
 * Usage: matrix[grmode+1][rqmode+1]
 */

static const int __quecvt_compat_matrix[8][8] = {
      /* UN NL CR CW PR PW EX PD */
        {0, 0, 0, 0, 0, 0, 0, 0},       /* UN */
        {0, 0, 1, 1, 1, 1, 1, 0},       /* NL */
        {0, 0, 0, 1, 1, 1, 1, 0},       /* CR */
        {0, 0, 0, 0, 1, 1, 1, 0},       /* CW */
        {0, 0, 0, 1, 0, 1, 1, 0},       /* PR */
        {0, 0, 0, 0, 0, 0, 1, 0},       /* PW */
        {0, 0, 0, 0, 0, 0, 0, 0},       /* EX */
        {0, 0, 0, 0, 0, 0, 0, 0}        /* PD */
};

void dlm_print_lkb(struct dlm_lkb *lkb)
{
	printk(KERN_ERR "lkb: nodeid %d id %x remid %x exflags %x flags %x "
	       "sts %d rq %d gr %d wait_type %d wait_nodeid %d seq %llu\n",
	       lkb->lkb_nodeid, lkb->lkb_id, lkb->lkb_remid, lkb->lkb_exflags,
	       lkb->lkb_flags, lkb->lkb_status, lkb->lkb_rqmode,
	       lkb->lkb_grmode, lkb->lkb_wait_type, lkb->lkb_wait_nodeid,
	       (unsigned long long)lkb->lkb_recover_seq);
}

static void dlm_print_rsb(struct dlm_rsb *r)
{
	printk(KERN_ERR "rsb: nodeid %d master %d dir %d flags %lx first %x "
	       "rlc %d name %s\n",
	       r->res_nodeid, r->res_master_nodeid, r->res_dir_nodeid,
	       r->res_flags, r->res_first_lkid, r->res_recover_locks_count,
	       r->res_name);
}

void dlm_dump_rsb(struct dlm_rsb *r)
{
	struct dlm_lkb *lkb;

	dlm_print_rsb(r);

	printk(KERN_ERR "rsb: root_list empty %d recover_list empty %d\n",
	       list_empty(&r->res_root_list), list_empty(&r->res_recover_list));
	printk(KERN_ERR "rsb lookup list\n");
	list_for_each_entry(lkb, &r->res_lookup, lkb_rsb_lookup)
		dlm_print_lkb(lkb);
	printk(KERN_ERR "rsb grant queue:\n");
	list_for_each_entry(lkb, &r->res_grantqueue, lkb_statequeue)
		dlm_print_lkb(lkb);
	printk(KERN_ERR "rsb convert queue:\n");
	list_for_each_entry(lkb, &r->res_convertqueue, lkb_statequeue)
		dlm_print_lkb(lkb);
	printk(KERN_ERR "rsb wait queue:\n");
	list_for_each_entry(lkb, &r->res_waitqueue, lkb_statequeue)
		dlm_print_lkb(lkb);
}

/* Threads cannot use the lockspace while it's being recovered */

static inline void dlm_lock_recovery(struct dlm_ls *ls)
{
	down_read(&ls->ls_in_recovery);
}

void dlm_unlock_recovery(struct dlm_ls *ls)
{
	up_read(&ls->ls_in_recovery);
}

int dlm_lock_recovery_try(struct dlm_ls *ls)
{
	return down_read_trylock(&ls->ls_in_recovery);
}

static inline int can_be_queued(struct dlm_lkb *lkb)
{
	return !(lkb->lkb_exflags & DLM_LKF_NOQUEUE);
}

static inline int force_blocking_asts(struct dlm_lkb *lkb)
{
	return (lkb->lkb_exflags & DLM_LKF_NOQUEUEBAST);
}

static inline int is_demoted(struct dlm_lkb *lkb)
{
	return (lkb->lkb_sbflags & DLM_SBF_DEMOTED);
}

static inline int is_altmode(struct dlm_lkb *lkb)
{
	return (lkb->lkb_sbflags & DLM_SBF_ALTMODE);
}

static inline int is_granted(struct dlm_lkb *lkb)
{
	return (lkb->lkb_status == DLM_LKSTS_GRANTED);
}

static inline int is_remote(struct dlm_rsb *r)
{
	DLM_ASSERT(r->res_nodeid >= 0, dlm_print_rsb(r););
	return !!r->res_nodeid;
}

static inline int is_process_copy(struct dlm_lkb *lkb)
{
	return (lkb->lkb_nodeid && !(lkb->lkb_flags & DLM_IFL_MSTCPY));
}

static inline int is_master_copy(struct dlm_lkb *lkb)
{
	return (lkb->lkb_flags & DLM_IFL_MSTCPY) ? 1 : 0;
}

static inline int middle_conversion(struct dlm_lkb *lkb)
{
	if ((lkb->lkb_grmode==DLM_LOCK_PR && lkb->lkb_rqmode==DLM_LOCK_CW) ||
	    (lkb->lkb_rqmode==DLM_LOCK_PR && lkb->lkb_grmode==DLM_LOCK_CW))
		return 1;
	return 0;
}

static inline int down_conversion(struct dlm_lkb *lkb)
{
	return (!middle_conversion(lkb) && lkb->lkb_rqmode < lkb->lkb_grmode);
}

static inline int is_overlap_unlock(struct dlm_lkb *lkb)
{
	return lkb->lkb_flags & DLM_IFL_OVERLAP_UNLOCK;
}

static inline int is_overlap_cancel(struct dlm_lkb *lkb)
{
	return lkb->lkb_flags & DLM_IFL_OVERLAP_CANCEL;
}

static inline int is_overlap(struct dlm_lkb *lkb)
{
	return (lkb->lkb_flags & (DLM_IFL_OVERLAP_UNLOCK |
				  DLM_IFL_OVERLAP_CANCEL));
}

static void queue_cast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rv)
{
	if (is_master_copy(lkb))
		return;

	del_timeout(lkb);

	DLM_ASSERT(lkb->lkb_lksb, dlm_print_lkb(lkb););

	/* if the operation was a cancel, then return -DLM_ECANCEL, if a
	   timeout caused the cancel then return -ETIMEDOUT */
	if (rv == -DLM_ECANCEL && (lkb->lkb_flags & DLM_IFL_TIMEOUT_CANCEL)) {
		lkb->lkb_flags &= ~DLM_IFL_TIMEOUT_CANCEL;
		rv = -ETIMEDOUT;
	}

	if (rv == -DLM_ECANCEL && (lkb->lkb_flags & DLM_IFL_DEADLOCK_CANCEL)) {
		lkb->lkb_flags &= ~DLM_IFL_DEADLOCK_CANCEL;
		rv = -EDEADLK;
	}

	dlm_add_cb(lkb, DLM_CB_CAST, lkb->lkb_grmode, rv, lkb->lkb_sbflags);
}

static inline void queue_cast_overlap(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	queue_cast(r, lkb,
		   is_overlap_unlock(lkb) ? -DLM_EUNLOCK : -DLM_ECANCEL);
}

static void queue_bast(struct dlm_rsb *r, struct dlm_lkb *lkb, int rqmode)
{
	if (is_master_copy(lkb)) {
		send_bast(r, lkb, rqmode);
	} else {
		dlm_add_cb(lkb, DLM_CB_BAST, rqmode, 0, 0);
	}
}

/*
 * Basic operations on rsb's and lkb's
 */

/* This is only called to add a reference when the code already holds
   a valid reference to the rsb, so there's no need for locking. */

static inline void hold_rsb(struct dlm_rsb *r)
{
	kref_get(&r->res_ref);
}

void dlm_hold_rsb(struct dlm_rsb *r)
{
	hold_rsb(r);
}

/* When all references to the rsb are gone it's transferred to
   the tossed list for later disposal. */

static void put_rsb(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;
	uint32_t bucket = r->res_bucket;

	spin_lock(&ls->ls_rsbtbl[bucket].lock);
	kref_put(&r->res_ref, toss_rsb);
	spin_unlock(&ls->ls_rsbtbl[bucket].lock);
}

void dlm_put_rsb(struct dlm_rsb *r)
{
	put_rsb(r);
}

static int pre_rsb_struct(struct dlm_ls *ls)
{
	struct dlm_rsb *r1, *r2;
	int count = 0;

	spin_lock(&ls->ls_new_rsb_spin);
	if (ls->ls_new_rsb_count > dlm_config.ci_new_rsb_count / 2) {
		spin_unlock(&ls->ls_new_rsb_spin);
		return 0;
	}
	spin_unlock(&ls->ls_new_rsb_spin);

	r1 = dlm_allocate_rsb(ls);
	r2 = dlm_allocate_rsb(ls);

	spin_lock(&ls->ls_new_rsb_spin);
	if (r1) {
		list_add(&r1->res_hashchain, &ls->ls_new_rsb);
		ls->ls_new_rsb_count++;
	}
	if (r2) {
		list_add(&r2->res_hashchain, &ls->ls_new_rsb);
		ls->ls_new_rsb_count++;
	}
	count = ls->ls_new_rsb_count;
	spin_unlock(&ls->ls_new_rsb_spin);

	if (!count)
		return -ENOMEM;
	return 0;
}

/* If ls->ls_new_rsb is empty, return -EAGAIN, so the caller can
   unlock any spinlocks, go back and call pre_rsb_struct again.
   Otherwise, take an rsb off the list and return it. */

static int get_rsb_struct(struct dlm_ls *ls, char *name, int len,
			  struct dlm_rsb **r_ret)
{
	struct dlm_rsb *r;
	int count;

	spin_lock(&ls->ls_new_rsb_spin);
	if (list_empty(&ls->ls_new_rsb)) {
		count = ls->ls_new_rsb_count;
		spin_unlock(&ls->ls_new_rsb_spin);
		log_debug(ls, "find_rsb retry %d %d %s",
			  count, dlm_config.ci_new_rsb_count, name);
		return -EAGAIN;
	}

	r = list_first_entry(&ls->ls_new_rsb, struct dlm_rsb, res_hashchain);
	list_del(&r->res_hashchain);
	/* Convert the empty list_head to a NULL rb_node for tree usage: */
	memset(&r->res_hashnode, 0, sizeof(struct rb_node));
	ls->ls_new_rsb_count--;
	spin_unlock(&ls->ls_new_rsb_spin);

	r->res_ls = ls;
	r->res_length = len;
	memcpy(r->res_name, name, len);
	mutex_init(&r->res_mutex);

	INIT_LIST_HEAD(&r->res_lookup);
	INIT_LIST_HEAD(&r->res_grantqueue);
	INIT_LIST_HEAD(&r->res_convertqueue);
	INIT_LIST_HEAD(&r->res_waitqueue);
	INIT_LIST_HEAD(&r->res_root_list);
	INIT_LIST_HEAD(&r->res_recover_list);

	*r_ret = r;
	return 0;
}

static int rsb_cmp(struct dlm_rsb *r, const char *name, int nlen)
{
	char maxname[DLM_RESNAME_MAXLEN];

	memset(maxname, 0, DLM_RESNAME_MAXLEN);
	memcpy(maxname, name, nlen);
	return memcmp(r->res_name, maxname, DLM_RESNAME_MAXLEN);
}

int dlm_search_rsb_tree(struct rb_root *tree, char *name, int len,
			struct dlm_rsb **r_ret)
{
	struct rb_node *node = tree->rb_node;
	struct dlm_rsb *r;
	int rc;

	while (node) {
		r = rb_entry(node, struct dlm_rsb, res_hashnode);
		rc = rsb_cmp(r, name, len);
		if (rc < 0)
			node = node->rb_left;
		else if (rc > 0)
			node = node->rb_right;
		else
			goto found;
	}
	*r_ret = NULL;
	return -EBADR;

 found:
	*r_ret = r;
	return 0;
}

static int rsb_insert(struct dlm_rsb *rsb, struct rb_root *tree)
{
	struct rb_node **newn = &tree->rb_node;
	struct rb_node *parent = NULL;
	int rc;

	while (*newn) {
		struct dlm_rsb *cur = rb_entry(*newn, struct dlm_rsb,
					       res_hashnode);

		parent = *newn;
		rc = rsb_cmp(cur, rsb->res_name, rsb->res_length);
		if (rc < 0)
			newn = &parent->rb_left;
		else if (rc > 0)
			newn = &parent->rb_right;
		else {
			log_print("rsb_insert match");
			dlm_dump_rsb(rsb);
			dlm_dump_rsb(cur);
			return -EEXIST;
		}
	}

	rb_link_node(&rsb->res_hashnode, parent, newn);
	rb_insert_color(&rsb->res_hashnode, tree);
	return 0;
}

/*
 * Find rsb in rsbtbl and potentially create/add one
 *
 * Delaying the release of rsb's has a similar benefit to applications keeping
 * NL locks on an rsb, but without the guarantee that the cached master value
 * will still be valid when the rsb is reused.  Apps aren't always smart enough
 * to keep NL locks on an rsb that they may lock again shortly; this can lead
 * to excessive master lookups and removals if we don't delay the release.
 *
 * Searching for an rsb means looking through both the normal list and toss
 * list.  When found on the toss list the rsb is moved to the normal list with
 * ref count of 1; when found on normal list the ref count is incremented.
 *
 * rsb's on the keep list are being used locally and refcounted.
 * rsb's on the toss list are not being used locally, and are not refcounted.
 *
 * The toss list rsb's were either
 * - previously used locally but not any more (were on keep list, then
 *   moved to toss list when last refcount dropped)
 * - created and put on toss list as a directory record for a lookup
 *   (we are the dir node for the res, but are not using the res right now,
 *   but some other node is)
 *
 * The purpose of find_rsb() is to return a refcounted rsb for local use.
 * So, if the given rsb is on the toss list, it is moved to the keep list
 * before being returned.
 *
 * toss_rsb() happens when all local usage of the rsb is done, i.e. no
 * more refcounts exist, so the rsb is moved from the keep list to the
 * toss list.
 *
 * rsb's on both keep and toss lists are used for doing a name to master
 * lookups.  rsb's that are in use locally (and being refcounted) are on
 * the keep list, rsb's that are not in use locally (not refcounted) and
 * only exist for name/master lookups are on the toss list.
 *
 * rsb's on the toss list who's dir_nodeid is not local can have stale
 * name/master mappings.  So, remote requests on such rsb's can potentially
 * return with an error, which means the mapping is stale and needs to
 * be updated with a new lookup.  (The idea behind MASTER UNCERTAIN and
 * first_lkid is to keep only a single outstanding request on an rsb
 * while that rsb has a potentially stale master.)
 */

static int find_rsb_dir(struct dlm_ls *ls, char *name, int len,
			uint32_t hash, uint32_t b,
			int dir_nodeid, int from_nodeid,
			unsigned int flags, struct dlm_rsb **r_ret)
{
	struct dlm_rsb *r = NULL;
	int our_nodeid = dlm_our_nodeid();
	int from_local = 0;
	int from_other = 0;
	int from_dir = 0;
	int create = 0;
	int error;

	if (flags & R_RECEIVE_REQUEST) {
		if (from_nodeid == dir_nodeid)
			from_dir = 1;
		else
			from_other = 1;
	} else if (flags & R_REQUEST) {
		from_local = 1;
	}

	/*
	 * flags & R_RECEIVE_RECOVER is from dlm_recover_master_copy, so
	 * from_nodeid has sent us a lock in dlm_recover_locks, believing
	 * we're the new master.  Our local recovery may not have set
	 * res_master_nodeid to our_nodeid yet, so allow either.  Don't
	 * create the rsb; dlm_recover_process_copy() will handle EBADR
	 * by resending.
	 *
	 * If someone sends us a request, we are the dir node, and we do
	 * not find the rsb anywhere, then recreate it.  This happens if
	 * someone sends us a request after we have removed/freed an rsb
	 * from our toss list.  (They sent a request instead of lookup
	 * because they are using an rsb from their toss list.)
	 */

	if (from_local || from_dir ||
	    (from_other && (dir_nodeid == our_nodeid))) {
		create = 1;
	}

 retry:
	if (create) {
		error = pre_rsb_struct(ls);
		if (error < 0)
			goto out;
	}

	spin_lock(&ls->ls_rsbtbl[b].lock);

	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
	if (error)
		goto do_toss;
	
	/*
	 * rsb is active, so we can't check master_nodeid without lock_rsb.
	 */

	kref_get(&r->res_ref);
	error = 0;
	goto out_unlock;


 do_toss:
	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
	if (error)
		goto do_new;

	/*
	 * rsb found inactive (master_nodeid may be out of date unless
	 * we are the dir_nodeid or were the master)  No other thread
	 * is using this rsb because it's on the toss list, so we can
	 * look at or update res_master_nodeid without lock_rsb.
	 */

	if ((r->res_master_nodeid != our_nodeid) && from_other) {
		/* our rsb was not master, and another node (not the dir node)
		   has sent us a request */
		log_debug(ls, "find_rsb toss from_other %d master %d dir %d %s",
			  from_nodeid, r->res_master_nodeid, dir_nodeid,
			  r->res_name);
		error = -ENOTBLK;
		goto out_unlock;
	}

	if ((r->res_master_nodeid != our_nodeid) && from_dir) {
		/* don't think this should ever happen */
		log_error(ls, "find_rsb toss from_dir %d master %d",
			  from_nodeid, r->res_master_nodeid);
		dlm_print_rsb(r);
		/* fix it and go on */
		r->res_master_nodeid = our_nodeid;
		r->res_nodeid = 0;
		rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
		r->res_first_lkid = 0;
	}

	if (from_local && (r->res_master_nodeid != our_nodeid)) {
		/* Because we have held no locks on this rsb,
		   res_master_nodeid could have become stale. */
		rsb_set_flag(r, RSB_MASTER_UNCERTAIN);
		r->res_first_lkid = 0;
	}

	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
	goto out_unlock;


 do_new:
	/*
	 * rsb not found
	 */

	if (error == -EBADR && !create)
		goto out_unlock;

	error = get_rsb_struct(ls, name, len, &r);
	if (error == -EAGAIN) {
		spin_unlock(&ls->ls_rsbtbl[b].lock);
		goto retry;
	}
	if (error)
		goto out_unlock;

	r->res_hash = hash;
	r->res_bucket = b;
	r->res_dir_nodeid = dir_nodeid;
	kref_init(&r->res_ref);

	if (from_dir) {
		/* want to see how often this happens */
		log_debug(ls, "find_rsb new from_dir %d recreate %s",
			  from_nodeid, r->res_name);
		r->res_master_nodeid = our_nodeid;
		r->res_nodeid = 0;
		goto out_add;
	}

	if (from_other && (dir_nodeid != our_nodeid)) {
		/* should never happen */
		log_error(ls, "find_rsb new from_other %d dir %d our %d %s",
			  from_nodeid, dir_nodeid, our_nodeid, r->res_name);
		dlm_free_rsb(r);
		r = NULL;
		error = -ENOTBLK;
		goto out_unlock;
	}

	if (from_other) {
		log_debug(ls, "find_rsb new from_other %d dir %d %s",
			  from_nodeid, dir_nodeid, r->res_name);
	}

	if (dir_nodeid == our_nodeid) {
		/* When we are the dir nodeid, we can set the master
		   node immediately */
		r->res_master_nodeid = our_nodeid;
		r->res_nodeid = 0;
	} else {
		/* set_master will send_lookup to dir_nodeid */
		r->res_master_nodeid = 0;
		r->res_nodeid = -1;
	}

 out_add:
	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 out_unlock:
	spin_unlock(&ls->ls_rsbtbl[b].lock);
 out:
	*r_ret = r;
	return error;
}

/* During recovery, other nodes can send us new MSTCPY locks (from
   dlm_recover_locks) before we've made ourself master (in
   dlm_recover_masters). */

static int find_rsb_nodir(struct dlm_ls *ls, char *name, int len,
			  uint32_t hash, uint32_t b,
			  int dir_nodeid, int from_nodeid,
			  unsigned int flags, struct dlm_rsb **r_ret)
{
	struct dlm_rsb *r = NULL;
	int our_nodeid = dlm_our_nodeid();
	int recover = (flags & R_RECEIVE_RECOVER);
	int error;

 retry:
	error = pre_rsb_struct(ls);
	if (error < 0)
		goto out;

	spin_lock(&ls->ls_rsbtbl[b].lock);

	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
	if (error)
		goto do_toss;

	/*
	 * rsb is active, so we can't check master_nodeid without lock_rsb.
	 */

	kref_get(&r->res_ref);
	goto out_unlock;


 do_toss:
	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
	if (error)
		goto do_new;

	/*
	 * rsb found inactive. No other thread is using this rsb because
	 * it's on the toss list, so we can look at or update
	 * res_master_nodeid without lock_rsb.
	 */

	if (!recover && (r->res_master_nodeid != our_nodeid) && from_nodeid) {
		/* our rsb is not master, and another node has sent us a
		   request; this should never happen */
		log_error(ls, "find_rsb toss from_nodeid %d master %d dir %d",
			  from_nodeid, r->res_master_nodeid, dir_nodeid);
		dlm_print_rsb(r);
		error = -ENOTBLK;
		goto out_unlock;
	}

	if (!recover && (r->res_master_nodeid != our_nodeid) &&
	    (dir_nodeid == our_nodeid)) {
		/* our rsb is not master, and we are dir; may as well fix it;
		   this should never happen */
		log_error(ls, "find_rsb toss our %d master %d dir %d",
			  our_nodeid, r->res_master_nodeid, dir_nodeid);
		dlm_print_rsb(r);
		r->res_master_nodeid = our_nodeid;
		r->res_nodeid = 0;
	}

	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
	goto out_unlock;


 do_new:
	/*
	 * rsb not found
	 */

	error = get_rsb_struct(ls, name, len, &r);
	if (error == -EAGAIN) {
		spin_unlock(&ls->ls_rsbtbl[b].lock);
		goto retry;
	}
	if (error)
		goto out_unlock;

	r->res_hash = hash;
	r->res_bucket = b;
	r->res_dir_nodeid = dir_nodeid;
	r->res_master_nodeid = dir_nodeid;
	r->res_nodeid = (dir_nodeid == our_nodeid) ? 0 : dir_nodeid;
	kref_init(&r->res_ref);

	error = rsb_insert(r, &ls->ls_rsbtbl[b].keep);
 out_unlock:
	spin_unlock(&ls->ls_rsbtbl[b].lock);
 out:
	*r_ret = r;
	return error;
}

static int find_rsb(struct dlm_ls *ls, char *name, int len, int from_nodeid,
		    unsigned int flags, struct dlm_rsb **r_ret)
{
	uint32_t hash, b;
	int dir_nodeid;

	if (len > DLM_RESNAME_MAXLEN)
		return -EINVAL;

	hash = jhash(name, len, 0);
	b = hash & (ls->ls_rsbtbl_size - 1);

	dir_nodeid = dlm_hash2nodeid(ls, hash);

	if (dlm_no_directory(ls))
		return find_rsb_nodir(ls, name, len, hash, b, dir_nodeid,
				      from_nodeid, flags, r_ret);
	else
		return find_rsb_dir(ls, name, len, hash, b, dir_nodeid,
				      from_nodeid, flags, r_ret);
}

/* we have received a request and found that res_master_nodeid != our_nodeid,
   so we need to return an error or make ourself the master */

static int validate_master_nodeid(struct dlm_ls *ls, struct dlm_rsb *r,
				  int from_nodeid)
{
	if (dlm_no_directory(ls)) {
		log_error(ls, "find_rsb keep from_nodeid %d master %d dir %d",
			  from_nodeid, r->res_master_nodeid,
			  r->res_dir_nodeid);
		dlm_print_rsb(r);
		return -ENOTBLK;
	}

	if (from_nodeid != r->res_dir_nodeid) {
		/* our rsb is not master, and another node (not the dir node)
	   	   has sent us a request.  this is much more common when our
	   	   master_nodeid is zero, so limit debug to non-zero.  */

		if (r->res_master_nodeid) {
			log_debug(ls, "validate master from_other %d master %d "
				  "dir %d first %x %s", from_nodeid,
				  r->res_master_nodeid, r->res_dir_nodeid,
				  r->res_first_lkid, r->res_name);
		}
		return -ENOTBLK;
	} else {
		/* our rsb is not master, but the dir nodeid has sent us a
	   	   request; this could happen with master 0 / res_nodeid -1 */

		if (r->res_master_nodeid) {
			log_error(ls, "validate master from_dir %d master %d "
				  "first %x %s",
				  from_nodeid, r->res_master_nodeid,
				  r->res_first_lkid, r->res_name);
		}

		r->res_master_nodeid = dlm_our_nodeid();
		r->res_nodeid = 0;
		return 0;
	}
}

/*
 * We're the dir node for this res and another node wants to know the
 * master nodeid.  During normal operation (non recovery) this is only
 * called from receive_lookup(); master lookups when the local node is
 * the dir node are done by find_rsb().
 *
 * normal operation, we are the dir node for a resource
 * . _request_lock
 * . set_master
 * . send_lookup
 * . receive_lookup
 * . dlm_master_lookup flags 0
 *
 * recover directory, we are rebuilding dir for all resources
 * . dlm_recover_directory
 * . dlm_rcom_names
 *   remote node sends back the rsb names it is master of and we are dir of
 * . dlm_master_lookup RECOVER_DIR (fix_master 0, from_master 1)
 *   we either create new rsb setting remote node as master, or find existing
 *   rsb and set master to be the remote node.
 *
 * recover masters, we are finding the new master for resources
 * . dlm_recover_masters
 * . recover_master
 * . dlm_send_rcom_lookup
 * . receive_rcom_lookup
 * . dlm_master_lookup RECOVER_MASTER (fix_master 1, from_master 0)
 */

int dlm_master_lookup(struct dlm_ls *ls, int from_nodeid, char *name, int len,
		      unsigned int flags, int *r_nodeid, int *result)
{
	struct dlm_rsb *r = NULL;
	uint32_t hash, b;
	int from_master = (flags & DLM_LU_RECOVER_DIR);
	int fix_master = (flags & DLM_LU_RECOVER_MASTER);
	int our_nodeid = dlm_our_nodeid();
	int dir_nodeid, error, toss_list = 0;

	if (len > DLM_RESNAME_MAXLEN)
		return -EINVAL;

	if (from_nodeid == our_nodeid) {
		log_error(ls, "dlm_master_lookup from our_nodeid %d flags %x",
			  our_nodeid, flags);
		return -EINVAL;
	}

	hash = jhash(name, len, 0);
	b = hash & (ls->ls_rsbtbl_size - 1);

	dir_nodeid = dlm_hash2nodeid(ls, hash);
	if (dir_nodeid != our_nodeid) {
		log_error(ls, "dlm_master_lookup from %d dir %d our %d h %x %d",
			  from_nodeid, dir_nodeid, our_nodeid, hash,
			  ls->ls_num_nodes);
		*r_nodeid = -1;
		return -EINVAL;
	}

 retry:
	error = pre_rsb_struct(ls);
	if (error < 0)
		return error;

	spin_lock(&ls->ls_rsbtbl[b].lock);
	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
	if (!error) {
		/* because the rsb is active, we need to lock_rsb before
		   checking/changing re_master_nodeid */

		hold_rsb(r);
		spin_unlock(&ls->ls_rsbtbl[b].lock);
		lock_rsb(r);
		goto found;
	}

	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
	if (error)
		goto not_found;

	/* because the rsb is inactive (on toss list), it's not refcounted
	   and lock_rsb is not used, but is protected by the rsbtbl lock */

	toss_list = 1;
 found:
	if (r->res_dir_nodeid != our_nodeid) {
		/* should not happen, but may as well fix it and carry on */
		log_error(ls, "dlm_master_lookup res_dir %d our %d %s",
			  r->res_dir_nodeid, our_nodeid, r->res_name);
		r->res_dir_nodeid = our_nodeid;
	}

	if (fix_master && dlm_is_removed(ls, r->res_master_nodeid)) {
		/* Recovery uses this function to set a new master when
		   the previous master failed.  Setting NEW_MASTER will
		   force dlm_recover_masters to call recover_master on this
		   rsb even though the res_nodeid is no longer removed. */

		r->res_master_nodeid = from_nodeid;
		r->res_nodeid = from_nodeid;
		rsb_set_flag(r, RSB_NEW_MASTER);

		if (toss_list) {
			/* I don't think we should ever find it on toss list. */
			log_error(ls, "dlm_master_lookup fix_master on toss");
			dlm_dump_rsb(r);
		}
	}

	if (from_master && (r->res_master_nodeid != from_nodeid)) {
		/* this will happen if from_nodeid became master during
		   a previous recovery cycle, and we aborted the previous
		   cycle before recovering this master value */

		log_limit(ls, "dlm_master_lookup from_master %d "
			  "master_nodeid %d res_nodeid %d first %x %s",
			  from_nodeid, r->res_master_nodeid, r->res_nodeid,
			  r->res_first_lkid, r->res_name);

		if (r->res_master_nodeid == our_nodeid) {
			log_error(ls, "from_master %d our_master", from_nodeid);
			dlm_dump_rsb(r);
			dlm_send_rcom_lookup_dump(r, from_nodeid);
			goto out_found;
		}

		r->res_master_nodeid = from_nodeid;
		r->res_nodeid = from_nodeid;
		rsb_set_flag(r, RSB_NEW_MASTER);
	}

	if (!r->res_master_nodeid) {
		/* this will happen if recovery happens while we're looking
		   up the master for this rsb */

		log_debug(ls, "dlm_master_lookup master 0 to %d first %x %s",
			  from_nodeid, r->res_first_lkid, r->res_name);
		r->res_master_nodeid = from_nodeid;
		r->res_nodeid = from_nodeid;
	}

	if (!from_master && !fix_master &&
	    (r->res_master_nodeid == from_nodeid)) {
		/* this can happen when the master sends remove, the dir node
		   finds the rsb on the keep list and ignores the remove,
		   and the former master sends a lookup */

		log_limit(ls, "dlm_master_lookup from master %d flags %x "
			  "first %x %s", from_nodeid, flags,
			  r->res_first_lkid, r->res_name);
	}

 out_found:
	*r_nodeid = r->res_master_nodeid;
	if (result)
		*result = DLM_LU_MATCH;

	if (toss_list) {
		r->res_toss_time = jiffies;
		/* the rsb was inactive (on toss list) */
		spin_unlock(&ls->ls_rsbtbl[b].lock);
	} else {
		/* the rsb was active */
		unlock_rsb(r);
		put_rsb(r);
	}
	return 0;

 not_found:
	error = get_rsb_struct(ls, name, len, &r);
	if (error == -EAGAIN) {
		spin_unlock(&ls->ls_rsbtbl[b].lock);
		goto retry;
	}
	if (error)
		goto out_unlock;

	r->res_hash = hash;
	r->res_bucket = b;
	r->res_dir_nodeid = our_nodeid;
	r->res_master_nodeid = from_nodeid;
	r->res_nodeid = from_nodeid;
	kref_init(&r->res_ref);
	r->res_toss_time = jiffies;

	error = rsb_insert(r, &ls->ls_rsbtbl[b].toss);
	if (error) {
		/* should never happen */
		dlm_free_rsb(r);
		spin_unlock(&ls->ls_rsbtbl[b].lock);
		goto retry;
	}

	if (result)
		*result = DLM_LU_ADD;
	*r_nodeid = from_nodeid;
	error = 0;
 out_unlock:
	spin_unlock(&ls->ls_rsbtbl[b].lock);
	return error;
}

static void dlm_dump_rsb_hash(struct dlm_ls *ls, uint32_t hash)
{
	struct rb_node *n;
	struct dlm_rsb *r;
	int i;

	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
		spin_lock(&ls->ls_rsbtbl[i].lock);
		for (n = rb_first(&ls->ls_rsbtbl[i].keep); n; n = rb_next(n)) {
			r = rb_entry(n, struct dlm_rsb, res_hashnode);
			if (r->res_hash == hash)
				dlm_dump_rsb(r);
		}
		spin_unlock(&ls->ls_rsbtbl[i].lock);
	}
}

void dlm_dump_rsb_name(struct dlm_ls *ls, char *name, int len)
{
	struct dlm_rsb *r = NULL;
	uint32_t hash, b;
	int error;

	hash = jhash(name, len, 0);
	b = hash & (ls->ls_rsbtbl_size - 1);

	spin_lock(&ls->ls_rsbtbl[b].lock);
	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].keep, name, len, &r);
	if (!error)
		goto out_dump;

	error = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
	if (error)
		goto out;
 out_dump:
	dlm_dump_rsb(r);
 out:
	spin_unlock(&ls->ls_rsbtbl[b].lock);
}

static void toss_rsb(struct kref *kref)
{
	struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);
	struct dlm_ls *ls = r->res_ls;

	DLM_ASSERT(list_empty(&r->res_root_list), dlm_print_rsb(r););
	kref_init(&r->res_ref);
	rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[r->res_bucket].keep);
	rsb_insert(r, &ls->ls_rsbtbl[r->res_bucket].toss);
	r->res_toss_time = jiffies;
	ls->ls_rsbtbl[r->res_bucket].flags |= DLM_RTF_SHRINK;
	if (r->res_lvbptr) {
		dlm_free_lvb(r->res_lvbptr);
		r->res_lvbptr = NULL;
	}
}

/* See comment for unhold_lkb */

static void unhold_rsb(struct dlm_rsb *r)
{
	int rv;
	rv = kref_put(&r->res_ref, toss_rsb);
	DLM_ASSERT(!rv, dlm_dump_rsb(r););
}

static void kill_rsb(struct kref *kref)
{
	struct dlm_rsb *r = container_of(kref, struct dlm_rsb, res_ref);

	/* All work is done after the return from kref_put() so we
	   can release the write_lock before the remove and free. */

	DLM_ASSERT(list_empty(&r->res_lookup), dlm_dump_rsb(r););
	DLM_ASSERT(list_empty(&r->res_grantqueue), dlm_dump_rsb(r););
	DLM_ASSERT(list_empty(&r->res_convertqueue), dlm_dump_rsb(r););
	DLM_ASSERT(list_empty(&r->res_waitqueue), dlm_dump_rsb(r););
	DLM_ASSERT(list_empty(&r->res_root_list), dlm_dump_rsb(r););
	DLM_ASSERT(list_empty(&r->res_recover_list), dlm_dump_rsb(r););
}

/* Attaching/detaching lkb's from rsb's is for rsb reference counting.
   The rsb must exist as long as any lkb's for it do. */

static void attach_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	hold_rsb(r);
	lkb->lkb_resource = r;
}

static void detach_lkb(struct dlm_lkb *lkb)
{
	if (lkb->lkb_resource) {
		put_rsb(lkb->lkb_resource);
		lkb->lkb_resource = NULL;
	}
}

static int create_lkb(struct dlm_ls *ls, struct dlm_lkb **lkb_ret)
{
	struct dlm_lkb *lkb;
	int rv;

	lkb = dlm_allocate_lkb(ls);
	if (!lkb)
		return -ENOMEM;

	lkb->lkb_nodeid = -1;
	lkb->lkb_grmode = DLM_LOCK_IV;
	kref_init(&lkb->lkb_ref);
	INIT_LIST_HEAD(&lkb->lkb_ownqueue);
	INIT_LIST_HEAD(&lkb->lkb_rsb_lookup);
	INIT_LIST_HEAD(&lkb->lkb_time_list);
	INIT_LIST_HEAD(&lkb->lkb_cb_list);
	mutex_init(&lkb->lkb_cb_mutex);
	INIT_WORK(&lkb->lkb_cb_work, dlm_callback_work);

	idr_preload(GFP_NOFS);
	spin_lock(&ls->ls_lkbidr_spin);
	rv = idr_alloc(&ls->ls_lkbidr, lkb, 1, 0, GFP_NOWAIT);
	if (rv >= 0)
		lkb->lkb_id = rv;
	spin_unlock(&ls->ls_lkbidr_spin);
	idr_preload_end();

	if (rv < 0) {
		log_error(ls, "create_lkb idr error %d", rv);
		return rv;
	}

	*lkb_ret = lkb;
	return 0;
}

static int find_lkb(struct dlm_ls *ls, uint32_t lkid, struct dlm_lkb **lkb_ret)
{
	struct dlm_lkb *lkb;

	spin_lock(&ls->ls_lkbidr_spin);
	lkb = idr_find(&ls->ls_lkbidr, lkid);
	if (lkb)
		kref_get(&lkb->lkb_ref);
	spin_unlock(&ls->ls_lkbidr_spin);

	*lkb_ret = lkb;
	return lkb ? 0 : -ENOENT;
}

static void kill_lkb(struct kref *kref)
{
	struct dlm_lkb *lkb = container_of(kref, struct dlm_lkb, lkb_ref);

	/* All work is done after the return from kref_put() so we
	   can release the write_lock before the detach_lkb */

	DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););
}

/* __put_lkb() is used when an lkb may not have an rsb attached to
   it so we need to provide the lockspace explicitly */

static int __put_lkb(struct dlm_ls *ls, struct dlm_lkb *lkb)
{
	uint32_t lkid = lkb->lkb_id;

	spin_lock(&ls->ls_lkbidr_spin);
	if (kref_put(&lkb->lkb_ref, kill_lkb)) {
		idr_remove(&ls->ls_lkbidr, lkid);
		spin_unlock(&ls->ls_lkbidr_spin);

		detach_lkb(lkb);

		/* for local/process lkbs, lvbptr points to caller's lksb */
		if (lkb->lkb_lvbptr && is_master_copy(lkb))
			dlm_free_lvb(lkb->lkb_lvbptr);
		dlm_free_lkb(lkb);
		return 1;
	} else {
		spin_unlock(&ls->ls_lkbidr_spin);
		return 0;
	}
}

int dlm_put_lkb(struct dlm_lkb *lkb)
{
	struct dlm_ls *ls;

	DLM_ASSERT(lkb->lkb_resource, dlm_print_lkb(lkb););
	DLM_ASSERT(lkb->lkb_resource->res_ls, dlm_print_lkb(lkb););

	ls = lkb->lkb_resource->res_ls;
	return __put_lkb(ls, lkb);
}

/* This is only called to add a reference when the code already holds
   a valid reference to the lkb, so there's no need for locking. */

static inline void hold_lkb(struct dlm_lkb *lkb)
{
	kref_get(&lkb->lkb_ref);
}

/* This is called when we need to remove a reference and are certain
   it's not the last ref.  e.g. del_lkb is always called between a
   find_lkb/put_lkb and is always the inverse of a previous add_lkb.
   put_lkb would work fine, but would involve unnecessary locking */

static inline void unhold_lkb(struct dlm_lkb *lkb)
{
	int rv;
	rv = kref_put(&lkb->lkb_ref, kill_lkb);
	DLM_ASSERT(!rv, dlm_print_lkb(lkb););
}

static void lkb_add_ordered(struct list_head *new, struct list_head *head,
			    int mode)
{
	struct dlm_lkb *lkb = NULL;

	list_for_each_entry(lkb, head, lkb_statequeue)
		if (lkb->lkb_rqmode < mode)
			break;

	__list_add(new, lkb->lkb_statequeue.prev, &lkb->lkb_statequeue);
}

/* add/remove lkb to rsb's grant/convert/wait queue */

static void add_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int status)
{
	kref_get(&lkb->lkb_ref);

	DLM_ASSERT(!lkb->lkb_status, dlm_print_lkb(lkb););

	lkb->lkb_timestamp = ktime_get();

	lkb->lkb_status = status;

	switch (status) {
	case DLM_LKSTS_WAITING:
		if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
			list_add(&lkb->lkb_statequeue, &r->res_waitqueue);
		else
			list_add_tail(&lkb->lkb_statequeue, &r->res_waitqueue);
		break;
	case DLM_LKSTS_GRANTED:
		/* convention says granted locks kept in order of grmode */
		lkb_add_ordered(&lkb->lkb_statequeue, &r->res_grantqueue,
				lkb->lkb_grmode);
		break;
	case DLM_LKSTS_CONVERT:
		if (lkb->lkb_exflags & DLM_LKF_HEADQUE)
			list_add(&lkb->lkb_statequeue, &r->res_convertqueue);
		else
			list_add_tail(&lkb->lkb_statequeue,
				      &r->res_convertqueue);
		break;
	default:
		DLM_ASSERT(0, dlm_print_lkb(lkb); printk("sts=%d\n", status););
	}
}

static void del_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	lkb->lkb_status = 0;
	list_del(&lkb->lkb_statequeue);
	unhold_lkb(lkb);
}

static void move_lkb(struct dlm_rsb *r, struct dlm_lkb *lkb, int sts)
{
	hold_lkb(lkb);
	del_lkb(r, lkb);
	add_lkb(r, lkb, sts);
	unhold_lkb(lkb);
}

static int msg_reply_type(int mstype)
{
	switch (mstype) {
	case DLM_MSG_REQUEST:
		return DLM_MSG_REQUEST_REPLY;
	case DLM_MSG_CONVERT:
		return DLM_MSG_CONVERT_REPLY;
	case DLM_MSG_UNLOCK:
		return DLM_MSG_UNLOCK_REPLY;
	case DLM_MSG_CANCEL:
		return DLM_MSG_CANCEL_REPLY;
	case DLM_MSG_LOOKUP:
		return DLM_MSG_LOOKUP_REPLY;
	}
	return -1;
}

static int nodeid_warned(int nodeid, int num_nodes, int *warned)
{
	int i;

	for (i = 0; i < num_nodes; i++) {
		if (!warned[i]) {
			warned[i] = nodeid;
			return 0;
		}
		if (warned[i] == nodeid)
			return 1;
	}
	return 0;
}

void dlm_scan_waiters(struct dlm_ls *ls)
{
	struct dlm_lkb *lkb;
	ktime_t zero = ktime_set(0, 0);
	s64 us;
	s64 debug_maxus = 0;
	u32 debug_scanned = 0;
	u32 debug_expired = 0;
	int num_nodes = 0;
	int *warned = NULL;

	if (!dlm_config.ci_waitwarn_us)
		return;

	mutex_lock(&ls->ls_waiters_mutex);

	list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
		if (ktime_equal(lkb->lkb_wait_time, zero))
			continue;

		debug_scanned++;

		us = ktime_to_us(ktime_sub(ktime_get(), lkb->lkb_wait_time));

		if (us < dlm_config.ci_waitwarn_us)
			continue;

		lkb->lkb_wait_time = zero;

		debug_expired++;
		if (us > debug_maxus)
			debug_maxus = us;

		if (!num_nodes) {
			num_nodes = ls->ls_num_nodes;
			warned = kzalloc(num_nodes * sizeof(int), GFP_KERNEL);
		}
		if (!warned)
			continue;
		if (nodeid_warned(lkb->lkb_wait_nodeid, num_nodes, warned))
			continue;

		log_error(ls, "waitwarn %x %lld %d us check connection to "
			  "node %d", lkb->lkb_id, (long long)us,
			  dlm_config.ci_waitwarn_us, lkb->lkb_wait_nodeid);
	}
	mutex_unlock(&ls->ls_waiters_mutex);
	kfree(warned);

	if (debug_expired)
		log_debug(ls, "scan_waiters %u warn %u over %d us max %lld us",
			  debug_scanned, debug_expired,
			  dlm_config.ci_waitwarn_us, (long long)debug_maxus);
}

/* add/remove lkb from global waiters list of lkb's waiting for
   a reply from a remote node */

static int add_to_waiters(struct dlm_lkb *lkb, int mstype, int to_nodeid)
{
	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
	int error = 0;

	mutex_lock(&ls->ls_waiters_mutex);

	if (is_overlap_unlock(lkb) ||
	    (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL))) {
		error = -EINVAL;
		goto out;
	}

	if (lkb->lkb_wait_type || is_overlap_cancel(lkb)) {
		switch (mstype) {
		case DLM_MSG_UNLOCK:
			lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
			break;
		case DLM_MSG_CANCEL:
			lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
			break;
		default:
			error = -EBUSY;
			goto out;
		}
		lkb->lkb_wait_count++;
		hold_lkb(lkb);

		log_debug(ls, "addwait %x cur %d overlap %d count %d f %x",
			  lkb->lkb_id, lkb->lkb_wait_type, mstype,
			  lkb->lkb_wait_count, lkb->lkb_flags);
		goto out;
	}

	DLM_ASSERT(!lkb->lkb_wait_count,
		   dlm_print_lkb(lkb);
		   printk("wait_count %d\n", lkb->lkb_wait_count););

	lkb->lkb_wait_count++;
	lkb->lkb_wait_type = mstype;
	lkb->lkb_wait_time = ktime_get();
	lkb->lkb_wait_nodeid = to_nodeid; /* for debugging */
	hold_lkb(lkb);
	list_add(&lkb->lkb_wait_reply, &ls->ls_waiters);
 out:
	if (error)
		log_error(ls, "addwait error %x %d flags %x %d %d %s",
			  lkb->lkb_id, error, lkb->lkb_flags, mstype,
			  lkb->lkb_wait_type, lkb->lkb_resource->res_name);
	mutex_unlock(&ls->ls_waiters_mutex);
	return error;
}

/* We clear the RESEND flag because we might be taking an lkb off the waiters
   list as part of process_requestqueue (e.g. a lookup that has an optimized
   request reply on the requestqueue) between dlm_recover_waiters_pre() which
   set RESEND and dlm_recover_waiters_post() */

static int _remove_from_waiters(struct dlm_lkb *lkb, int mstype,
				struct dlm_message *ms)
{
	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
	int overlap_done = 0;

	if (is_overlap_unlock(lkb) && (mstype == DLM_MSG_UNLOCK_REPLY)) {
		log_debug(ls, "remwait %x unlock_reply overlap", lkb->lkb_id);
		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_UNLOCK;
		overlap_done = 1;
		goto out_del;
	}

	if (is_overlap_cancel(lkb) && (mstype == DLM_MSG_CANCEL_REPLY)) {
		log_debug(ls, "remwait %x cancel_reply overlap", lkb->lkb_id);
		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
		overlap_done = 1;
		goto out_del;
	}

	/* Cancel state was preemptively cleared by a successful convert,
	   see next comment, nothing to do. */

	if ((mstype == DLM_MSG_CANCEL_REPLY) &&
	    (lkb->lkb_wait_type != DLM_MSG_CANCEL)) {
		log_debug(ls, "remwait %x cancel_reply wait_type %d",
			  lkb->lkb_id, lkb->lkb_wait_type);
		return -1;
	}

	/* Remove for the convert reply, and premptively remove for the
	   cancel reply.  A convert has been granted while there's still
	   an outstanding cancel on it (the cancel is moot and the result
	   in the cancel reply should be 0).  We preempt the cancel reply
	   because the app gets the convert result and then can follow up
	   with another op, like convert.  This subsequent op would see the
	   lingering state of the cancel and fail with -EBUSY. */

	if ((mstype == DLM_MSG_CONVERT_REPLY) &&
	    (lkb->lkb_wait_type == DLM_MSG_CONVERT) &&
	    is_overlap_cancel(lkb) && ms && !ms->m_result) {
		log_debug(ls, "remwait %x convert_reply zap overlap_cancel",
			  lkb->lkb_id);
		lkb->lkb_wait_type = 0;
		lkb->lkb_flags &= ~DLM_IFL_OVERLAP_CANCEL;
		lkb->lkb_wait_count--;
		goto out_del;
	}

	/* N.B. type of reply may not always correspond to type of original
	   msg due to lookup->request optimization, verify others? */

	if (lkb->lkb_wait_type) {
		lkb->lkb_wait_type = 0;
		goto out_del;
	}

	log_error(ls, "remwait error %x remote %d %x msg %d flags %x no wait",
		  lkb->lkb_id, ms ? ms->m_header.h_nodeid : 0, lkb->lkb_remid,
		  mstype, lkb->lkb_flags);
	return -1;

 out_del:
	/* the force-unlock/cancel has completed and we haven't recvd a reply
	   to the op that was in progress prior to the unlock/cancel; we
	   give up on any reply to the earlier op.  FIXME: not sure when/how
	   this would happen */

	if (overlap_done && lkb->lkb_wait_type) {
		log_error(ls, "remwait error %x reply %d wait_type %d overlap",
			  lkb->lkb_id, mstype, lkb->lkb_wait_type);
		lkb->lkb_wait_count--;
		lkb->lkb_wait_type = 0;
	}

	DLM_ASSERT(lkb->lkb_wait_count, dlm_print_lkb(lkb););

	lkb->lkb_flags &= ~DLM_IFL_RESEND;
	lkb->lkb_wait_count--;
	if (!lkb->lkb_wait_count)
		list_del_init(&lkb->lkb_wait_reply);
	unhold_lkb(lkb);
	return 0;
}

static int remove_from_waiters(struct dlm_lkb *lkb, int mstype)
{
	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
	int error;

	mutex_lock(&ls->ls_waiters_mutex);
	error = _remove_from_waiters(lkb, mstype, NULL);
	mutex_unlock(&ls->ls_waiters_mutex);
	return error;
}

/* Handles situations where we might be processing a "fake" or "stub" reply in
   which we can't try to take waiters_mutex again. */

static int remove_from_waiters_ms(struct dlm_lkb *lkb, struct dlm_message *ms)
{
	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
	int error;

	if (ms->m_flags != DLM_IFL_STUB_MS)
		mutex_lock(&ls->ls_waiters_mutex);
	error = _remove_from_waiters(lkb, ms->m_type, ms);
	if (ms->m_flags != DLM_IFL_STUB_MS)
		mutex_unlock(&ls->ls_waiters_mutex);
	return error;
}

/* If there's an rsb for the same resource being removed, ensure
   that the remove message is sent before the new lookup message.
   It should be rare to need a delay here, but if not, then it may
   be worthwhile to add a proper wait mechanism rather than a delay. */

static void wait_pending_remove(struct dlm_rsb *r)
{
	struct dlm_ls *ls = r->res_ls;
 restart:
	spin_lock(&ls->ls_remove_spin);
	if (ls->ls_remove_len &&
	    !rsb_cmp(r, ls->ls_remove_name, ls->ls_remove_len)) {
		log_debug(ls, "delay lookup for remove dir %d %s",
		  	  r->res_dir_nodeid, r->res_name);
		spin_unlock(&ls->ls_remove_spin);
		msleep(1);
		goto restart;
	}
	spin_unlock(&ls->ls_remove_spin);
}

/*
 * ls_remove_spin protects ls_remove_name and ls_remove_len which are
 * read by other threads in wait_pending_remove.  ls_remove_names
 * and ls_remove_lens are only used by the scan thread, so they do
 * not need protection.
 */

static void shrink_bucket(struct dlm_ls *ls, int b)
{
	struct rb_node *n, *next;
	struct dlm_rsb *r;
	char *name;
	int our_nodeid = dlm_our_nodeid();
	int remote_count = 0;
	int need_shrink = 0;
	int i, len, rv;

	memset(&ls->ls_remove_lens, 0, sizeof(int) * DLM_REMOVE_NAMES_MAX);

	spin_lock(&ls->ls_rsbtbl[b].lock);

	if (!(ls->ls_rsbtbl[b].flags & DLM_RTF_SHRINK)) {
		spin_unlock(&ls->ls_rsbtbl[b].lock);
		return;
	}

	for (n = rb_first(&ls->ls_rsbtbl[b].toss); n; n = next) {
		next = rb_next(n);
		r = rb_entry(n, struct dlm_rsb, res_hashnode);

		/* If we're the directory record for this rsb, and
		   we're not the master of it, then we need to wait
		   for the master node to send us a dir remove for
		   before removing the dir record. */

		if (!dlm_no_directory(ls) &&
		    (r->res_master_nodeid != our_nodeid) &&
		    (dlm_dir_nodeid(r) == our_nodeid)) {
			continue;
		}

		need_shrink = 1;

		if (!time_after_eq(jiffies, r->res_toss_time +
				   dlm_config.ci_toss_secs * HZ)) {
			continue;
		}

		if (!dlm_no_directory(ls) &&
		    (r->res_master_nodeid == our_nodeid) &&
		    (dlm_dir_nodeid(r) != our_nodeid)) {

			/* We're the master of this rsb but we're not
			   the directory record, so we need to tell the
			   dir node to remove the dir record. */

			ls->ls_remove_lens[remote_count] = r->res_length;
			memcpy(ls->ls_remove_names[remote_count], r->res_name,
			       DLM_RESNAME_MAXLEN);
			remote_count++;

			if (remote_count >= DLM_REMOVE_NAMES_MAX)
				break;
			continue;
		}

		if (!kref_put(&r->res_ref, kill_rsb)) {
			log_error(ls, "tossed rsb in use %s", r->res_name);
			continue;
		}

		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);
		dlm_free_rsb(r);
	}

	if (need_shrink)
		ls->ls_rsbtbl[b].flags |= DLM_RTF_SHRINK;
	else
		ls->ls_rsbtbl[b].flags &= ~DLM_RTF_SHRINK;
	spin_unlock(&ls->ls_rsbtbl[b].lock);

	/*
	 * While searching for rsb's to free, we found some that require
	 * remote removal.  We leave them in place and find them again here
	 * so there is a very small gap between removing them from the toss
	 * list and sending the removal.  Keeping this gap small is
	 * important to keep us (the master node) from being out of sync
	 * with the remote dir node for very long.
	 *
	 * From the time the rsb is removed from toss until just after
	 * send_remove, the rsb name is saved in ls_remove_name.  A new
	 * lookup checks this to ensure that a new lookup message for the
	 * same resource name is not sent just before the remove message.
	 */

	for (i = 0; i < remote_count; i++) {
		name = ls->ls_remove_names[i];
		len = ls->ls_remove_lens[i];

		spin_lock(&ls->ls_rsbtbl[b].lock);
		rv = dlm_search_rsb_tree(&ls->ls_rsbtbl[b].toss, name, len, &r);
		if (rv) {
			spin_unlock(&ls->ls_rsbtbl[b].lock);
			log_debug(ls, "remove_name not toss %s", name);
			continue;
		}

		if (r->res_master_nodeid != our_nodeid) {
			spin_unlock(&ls->ls_rsbtbl[b].lock);
			log_debug(ls, "remove_name master %d dir %d our %d %s",
				  r->res_master_nodeid, r->res_dir_nodeid,
				  our_nodeid, name);
			continue;
		}

		if (r->res_dir_nodeid == our_nodeid) {
			/* should never happen */
			spin_unlock(&ls->ls_rsbtbl[b].lock);
			log_error(ls, "remove_name dir %d master %d our %d %s",
				  r->res_dir_nodeid, r->res_master_nodeid,
				  our_nodeid, name);
			continue;
		}

		if (!time_after_eq(jiffies, r->res_toss_time +
				   dlm_config.ci_toss_secs * HZ)) {
			spin_unlock(&ls->ls_rsbtbl[b].lock);
			log_debug(ls, "remove_name toss_time %lu now %lu %s",
				  r->res_toss_time, jiffies, name);
			continue;
		}

		if (!kref_put(&r->res_ref, kill_rsb)) {
			spin_unlock(&ls->ls_rsbtbl[b].lock);
			log_error(ls, "remove_name in use %s", name);
			continue;
		}

		rb_erase(&r->res_hashnode, &ls->ls_rsbtbl[b].toss);

		/* block lookup of same name until we've sent remove */
		spin_lock(&ls->ls_remove_spin);
		ls->ls_remove_len = len;
		memcpy(ls->ls_remove_name, name, DLM_RESNAME_MAXLEN);
		spin_unlock(&ls->ls_remove_spin);
		spin_unlock(&ls->ls_rsbtbl[b].lock);

		send_remove(r);

		/* allow lookup of name again */
		spin_lock(&ls->ls_remove_spin);
		ls->ls_remove_len = 0;
		memset(ls->ls_remove_name, 0, DLM_RESNAME_MAXLEN);
		spin_unlock(&ls->ls_remove_spin);

		dlm_free_rsb(r);
	}
}

void dlm_scan_rsbs(struct dlm_ls *ls)
{
	int i;

	for (i = 0; i < ls->ls_rsbtbl_size; i++) {
		shrink_bucket(ls, i);
		if (dlm_locking_stopped(ls))
			break;
		cond_resched();
	}
}

static void add_timeout(struct dlm_lkb *lkb)
{
	struct dlm_ls *ls = lkb->lkb_resource->res_ls;

	if (is_master_copy(lkb))
		return;

	if (test_bit(LSFL_TIMEWARN, &ls->ls_flags) &&
	    !(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
		lkb->lkb_flags |= DLM_IFL_WATCH_TIMEWARN;
		goto add_it;
	}
	if (lkb->lkb_exflags & DLM_LKF_TIMEOUT)
		goto add_it;
	return;

 add_it:
	DLM_ASSERT(list_empty(&lkb->lkb_time_list), dlm_print_lkb(lkb););
	mutex_lock(&ls->ls_timeout_mutex);
	hold_lkb(lkb);
	list_add_tail(&lkb->lkb_time_list, &ls->ls_timeout);
	mutex_unlock(&ls->ls_timeout_mutex);
}

static void del_timeout(struct dlm_lkb *lkb)
{
	struct dlm_ls *ls = lkb->lkb_resource->res_ls;

	mutex_lock(&ls->ls_timeout_mutex);
	if (!list_empty(&lkb->lkb_time_list)) {
		list_del_init(&lkb->lkb_time_list);
		unhold_lkb(lkb);
	}
	mutex_unlock(&ls->ls_timeout_mutex);
}

/* FIXME: is it safe to look at lkb_exflags, lkb_flags, lkb_timestamp, and
   lkb_lksb_timeout without lock_rsb?  Note: we can't lock timeout_mutex
   and then lock rsb because of lock ordering in add_timeout.  We may need
   to specify some special timeout-related bits in the lkb that are just to
   be accessed under the timeout_mutex. */

void dlm_scan_timeout(struct dlm_ls *ls)
{
	struct dlm_rsb *r;
	struct dlm_lkb *lkb;
	int do_cancel, do_warn;
	s64 wait_us;

	for (;;) {
		if (dlm_locking_stopped(ls))
			break;

		do_cancel = 0;
		do_warn = 0;
		mutex_lock(&ls->ls_timeout_mutex);
		list_for_each_entry(lkb, &ls->ls_timeout, lkb_time_list) {

			wait_us = ktime_to_us(ktime_sub(ktime_get(),
					      		lkb->lkb_timestamp));

			if ((lkb->lkb_exflags & DLM_LKF_TIMEOUT) &&
			    wait_us >= (lkb->lkb_timeout_cs * 10000))
				do_cancel = 1;

			if ((lkb->lkb_flags & DLM_IFL_WATCH_TIMEWARN) &&
			    wait_us >= dlm_config.ci_timewarn_cs * 10000)
				do_warn = 1;

			if (!do_cancel && !do_warn)
				continue;
			hold_lkb(lkb);
			break;
		}
		mutex_unlock(&ls->ls_timeout_mutex);

		if (!do_cancel && !do_warn)
			break;

		r = lkb->lkb_resource;
		hold_rsb(r);
		lock_rsb(r);

		if (do_warn) {
			/* clear flag so we only warn once */
			lkb->lkb_flags &= ~DLM_IFL_WATCH_TIMEWARN;
			if (!(lkb->lkb_exflags & DLM_LKF_TIMEOUT))
				del_timeout(lkb);
			dlm_timeout_warn(lkb);
		}

		if (do_cancel) {
			log_debug(ls, "timeout cancel %x node %d %s",
				  lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
			lkb->lkb_flags &= ~DLM_IFL_WATCH_TIMEWARN;
			lkb->lkb_flags |= DLM_IFL_TIMEOUT_CANCEL;
			del_timeout(lkb);
			_cancel_lock(r, lkb);
		}

		unlock_rsb(r);
		unhold_rsb(r);
		dlm_put_lkb(lkb);
	}
}

/* This is only called by dlm_recoverd, and we rely on dlm_ls_stop() stopping
   dlm_recoverd before checking/setting ls_recover_begin. */

void dlm_adjust_timeouts(struct dlm_ls *ls)
{
	struct dlm_lkb *lkb;
	u64 adj_us = jiffies_to_usecs(jiffies - ls->ls_recover_begin);

	ls->ls_recover_begin = 0;
	mutex_lock(&ls->ls_timeout_mutex);
	list_for_each_entry(lkb, &ls->ls_timeout, lkb_time_list)
		lkb->lkb_timestamp = ktime_add_us(lkb->lkb_timestamp, adj_us);
	mutex_unlock(&ls->ls_timeout_mutex);

	if (!dlm_config.ci_waitwarn_us)
		return;

	mutex_lock(&ls->ls_waiters_mutex);
	list_for_each_entry(lkb, &ls->ls_waiters, lkb_wait_reply) {
		if (ktime_to_us(lkb->lkb_wait_time))
			lkb->lkb_wait_time = ktime_get();
	}
	mutex_unlock(&ls->ls_waiters_mutex);
}

/* lkb is master or local copy */

static void set_lvb_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	int b, len = r->res_ls->ls_lvblen;

	/* b=1 lvb returned to caller
	   b=0 lvb written to rsb or invalidated
	   b=-1 do nothing */

	b =  dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];

	if (b == 1) {
		if (!lkb->lkb_lvbptr)
			return;

		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
			return;

		if (!r->res_lvbptr)
			return;

		memcpy(lkb->lkb_lvbptr, r->res_lvbptr, len);
		lkb->lkb_lvbseq = r->res_lvbseq;

	} else if (b == 0) {
		if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
			rsb_set_flag(r, RSB_VALNOTVALID);
			return;
		}

		if (!lkb->lkb_lvbptr)
			return;

		if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
			return;

		if (!r->res_lvbptr)
			r->res_lvbptr = dlm_allocate_lvb(r->res_ls);

		if (!r->res_lvbptr)
			return;

		memcpy(r->res_lvbptr, lkb->lkb_lvbptr, len);
		r->res_lvbseq++;
		lkb->lkb_lvbseq = r->res_lvbseq;
		rsb_clear_flag(r, RSB_VALNOTVALID);
	}

	if (rsb_flag(r, RSB_VALNOTVALID))
		lkb->lkb_sbflags |= DLM_SBF_VALNOTVALID;
}

static void set_lvb_unlock(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	if (lkb->lkb_grmode < DLM_LOCK_PW)
		return;

	if (lkb->lkb_exflags & DLM_LKF_IVVALBLK) {
		rsb_set_flag(r, RSB_VALNOTVALID);
		return;
	}

	if (!lkb->lkb_lvbptr)
		return;

	if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
		return;

	if (!r->res_lvbptr)
		r->res_lvbptr = dlm_allocate_lvb(r->res_ls);

	if (!r->res_lvbptr)
		return;

	memcpy(r->res_lvbptr, lkb->lkb_lvbptr, r->res_ls->ls_lvblen);
	r->res_lvbseq++;
	rsb_clear_flag(r, RSB_VALNOTVALID);
}

/* lkb is process copy (pc) */

static void set_lvb_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
			    struct dlm_message *ms)
{
	int b;

	if (!lkb->lkb_lvbptr)
		return;

	if (!(lkb->lkb_exflags & DLM_LKF_VALBLK))
		return;

	b = dlm_lvb_operations[lkb->lkb_grmode + 1][lkb->lkb_rqmode + 1];
	if (b == 1) {
		int len = receive_extralen(ms);
		if (len > r->res_ls->ls_lvblen)
			len = r->res_ls->ls_lvblen;
		memcpy(lkb->lkb_lvbptr, ms->m_extra, len);
		lkb->lkb_lvbseq = ms->m_lvbseq;
	}
}

/* Manipulate lkb's on rsb's convert/granted/waiting queues
   remove_lock -- used for unlock, removes lkb from granted
   revert_lock -- used for cancel, moves lkb from convert to granted
   grant_lock  -- used for request and convert, adds lkb to granted or
                  moves lkb from convert or waiting to granted

   Each of these is used for master or local copy lkb's.  There is
   also a _pc() variation used to make the corresponding change on
   a process copy (pc) lkb. */

static void _remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	del_lkb(r, lkb);
	lkb->lkb_grmode = DLM_LOCK_IV;
	/* this unhold undoes the original ref from create_lkb()
	   so this leads to the lkb being freed */
	unhold_lkb(lkb);
}

static void remove_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	set_lvb_unlock(r, lkb);
	_remove_lock(r, lkb);
}

static void remove_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	_remove_lock(r, lkb);
}

/* returns: 0 did nothing
	    1 moved lock to granted
	   -1 removed lock */

static int revert_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	int rv = 0;

	lkb->lkb_rqmode = DLM_LOCK_IV;

	switch (lkb->lkb_status) {
	case DLM_LKSTS_GRANTED:
		break;
	case DLM_LKSTS_CONVERT:
		move_lkb(r, lkb, DLM_LKSTS_GRANTED);
		rv = 1;
		break;
	case DLM_LKSTS_WAITING:
		del_lkb(r, lkb);
		lkb->lkb_grmode = DLM_LOCK_IV;
		/* this unhold undoes the original ref from create_lkb()
		   so this leads to the lkb being freed */
		unhold_lkb(lkb);
		rv = -1;
		break;
	default:
		log_print("invalid status for revert %d", lkb->lkb_status);
	}
	return rv;
}

static int revert_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	return revert_lock(r, lkb);
}

static void _grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	if (lkb->lkb_grmode != lkb->lkb_rqmode) {
		lkb->lkb_grmode = lkb->lkb_rqmode;
		if (lkb->lkb_status)
			move_lkb(r, lkb, DLM_LKSTS_GRANTED);
		else
			add_lkb(r, lkb, DLM_LKSTS_GRANTED);
	}

	lkb->lkb_rqmode = DLM_LOCK_IV;
	lkb->lkb_highbast = 0;
}

static void grant_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	set_lvb_lock(r, lkb);
	_grant_lock(r, lkb);
}

static void grant_lock_pc(struct dlm_rsb *r, struct dlm_lkb *lkb,
			  struct dlm_message *ms)
{
	set_lvb_lock_pc(r, lkb, ms);
	_grant_lock(r, lkb);
}

/* called by grant_pending_locks() which means an async grant message must
   be sent to the requesting node in addition to granting the lock if the
   lkb belongs to a remote node. */

static void grant_lock_pending(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	grant_lock(r, lkb);
	if (is_master_copy(lkb))
		send_grant(r, lkb);
	else
		queue_cast(r, lkb, 0);
}

/* The special CONVDEADLK, ALTPR and ALTCW flags allow the master to
   change the granted/requested modes.  We're munging things accordingly in
   the process copy.
   CONVDEADLK: our grmode may have been forced down to NL to resolve a
   conversion deadlock
   ALTPR/ALTCW: our rqmode may have been changed to PR or CW to become
   compatible with other granted locks */

static void munge_demoted(struct dlm_lkb *lkb)
{
	if (lkb->lkb_rqmode == DLM_LOCK_IV || lkb->lkb_grmode == DLM_LOCK_IV) {
		log_print("munge_demoted %x invalid modes gr %d rq %d",
			  lkb->lkb_id, lkb->lkb_grmode, lkb->lkb_rqmode);
		return;
	}

	lkb->lkb_grmode = DLM_LOCK_NL;
}

static void munge_altmode(struct dlm_lkb *lkb, struct dlm_message *ms)
{
	if (ms->m_type != DLM_MSG_REQUEST_REPLY &&
	    ms->m_type != DLM_MSG_GRANT) {
		log_print("munge_altmode %x invalid reply type %d",
			  lkb->lkb_id, ms->m_type);
		return;
	}

	if (lkb->lkb_exflags & DLM_LKF_ALTPR)
		lkb->lkb_rqmode = DLM_LOCK_PR;
	else if (lkb->lkb_exflags & DLM_LKF_ALTCW)
		lkb->lkb_rqmode = DLM_LOCK_CW;
	else {
		log_print("munge_altmode invalid exflags %x", lkb->lkb_exflags);
		dlm_print_lkb(lkb);
	}
}

static inline int first_in_list(struct dlm_lkb *lkb, struct list_head *head)
{
	struct dlm_lkb *first = list_entry(head->next, struct dlm_lkb,
					   lkb_statequeue);
	if (lkb->lkb_id == first->lkb_id)
		return 1;

	return 0;
}

/* Check if the given lkb conflicts with another lkb on the queue. */

static int queue_conflict(struct list_head *head, struct dlm_lkb *lkb)
{
	struct dlm_lkb *this;

	list_for_each_entry(this, head, lkb_statequeue) {
		if (this == lkb)
			continue;
		if (!modes_compat(this, lkb))
			return 1;
	}
	return 0;
}

/*
 * "A conversion deadlock arises with a pair of lock requests in the converting
 * queue for one resource.  The granted mode of each lock blocks the requested
 * mode of the other lock."
 *
 * Part 2: if the granted mode of lkb is preventing an earlier lkb in the
 * convert queue from being granted, then deadlk/demote lkb.
 *
 * Example:
 * Granted Queue: empty
 * Convert Queue: NL->EX (first lock)
 *                PR->EX (second lock)
 *
 * The first lock can't be granted because of the granted mode of the second
 * lock and the second lock can't be granted because it's not first in the
 * list.  We either cancel lkb's conversion (PR->EX) and return EDEADLK, or we
 * demote the granted mode of lkb (from PR to NL) if it has the CONVDEADLK
 * flag set and return DEMOTED in the lksb flags.
 *
 * Originally, this function detected conv-deadlk in a more limited scope:
 * - if !modes_compat(lkb1, lkb2) && !modes_compat(lkb2, lkb1), or
 * - if lkb1 was the first entry in the queue (not just earlier), and was
 *   blocked by the granted mode of lkb2, and there was nothing on the
 *   granted queue preventing lkb1 from being granted immediately, i.e.
 *   lkb2 was the only thing preventing lkb1 from being granted.
 *
 * That second condition meant we'd only say there was conv-deadlk if
 * resolving it (by demotion) would lead to the first lock on the convert
 * queue being granted right away.  It allowed conversion deadlocks to exist
 * between locks on the convert queue while they couldn't be granted anyway.
 *
 * Now, we detect and take action on conversion deadlocks immediately when
 * they're created, even if they may not be immediately consequential.  If
 * lkb1 exists anywhere in the convert queue and lkb2 comes in with a granted
 * mode that would prevent lkb1's conversion from being granted, we do a
 * deadlk/demote on lkb2 right away and don't let it onto the convert queue.
 * I think this means that the lkb_is_ahead condition below should always
 * be zero, i.e. there will never be conv-deadlk between two locks that are
 * both already on the convert queue.
 */

static int conversion_deadlock_detect(struct dlm_rsb *r, struct dlm_lkb *lkb2)
{
	struct dlm_lkb *lkb1;
	int lkb_is_ahead = 0;

	list_for_each_entry(lkb1, &r->res_convertqueue, lkb_statequeue) {
		if (lkb1 == lkb2) {
			lkb_is_ahead = 1;
			continue;
		}

		if (!lkb_is_ahead) {
			if (!modes_compat(lkb2, lkb1))
				return 1;
		} else {
			if (!modes_compat(lkb2, lkb1) &&
			    !modes_compat(lkb1, lkb2))
				return 1;
		}
	}
	return 0;
}

/*
 * Return 1 if the lock can be granted, 0 otherwise.
 * Also detect and resolve conversion deadlocks.
 *
 * lkb is the lock to be granted
 *
 * now is 1 if the function is being called in the context of the
 * immediate request, it is 0 if called later, after the lock has been
 * queued.
 *
 * recover is 1 if dlm_recover_grant() is trying to grant conversions
 * after recovery.
 *
 * References are from chapter 6 of "VAXcluster Principles" by Roy Davis
 */

static int _can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
			   int recover)
{
	int8_t conv = (lkb->lkb_grmode != DLM_LOCK_IV);

	/*
	 * 6-10: Version 5.4 introduced an option to address the phenomenon of
	 * a new request for a NL mode lock being blocked.
	 *
	 * 6-11: If the optional EXPEDITE flag is used with the new NL mode
	 * request, then it would be granted.  In essence, the use of this flag
	 * tells the Lock Manager to expedite theis request by not considering
	 * what may be in the CONVERTING or WAITING queues...  As of this
	 * writing, the EXPEDITE flag can be used only with new requests for NL
	 * mode locks.  This flag is not valid for conversion requests.
	 *
	 * A shortcut.  Earlier checks return an error if EXPEDITE is used in a
	 * conversion or used with a non-NL requested mode.  We also know an
	 * EXPEDITE request is always granted immediately, so now must always
	 * be 1.  The full condition to grant an expedite request: (now &&
	 * !conv && lkb->rqmode == DLM_LOCK_NL && (flags & EXPEDITE)) can
	 * therefore be shortened to just checking the flag.
	 */

	if (lkb->lkb_exflags & DLM_LKF_EXPEDITE)
		return 1;

	/*
	 * A shortcut. Without this, !queue_conflict(grantqueue, lkb) would be
	 * added to the remaining conditions.
	 */

	if (queue_conflict(&r->res_grantqueue, lkb))
		return 0;

	/*
	 * 6-3: By default, a conversion request is immediately granted if the
	 * requested mode is compatible with the modes of all other granted
	 * locks
	 */

	if (queue_conflict(&r->res_convertqueue, lkb))
		return 0;

	/*
	 * The RECOVER_GRANT flag means dlm_recover_grant() is granting
	 * locks for a recovered rsb, on which lkb's have been rebuilt.
	 * The lkb's may have been rebuilt on the queues in a different
	 * order than they were in on the previous master.  So, granting
	 * queued conversions in order after recovery doesn't make sense
	 * since the order hasn't been preserved anyway.  The new order
	 * could also have created a new "in place" conversion deadlock.
	 * (e.g. old, failed master held granted EX, with PR->EX, NL->EX.
	 * After recovery, there would be no granted locks, and possibly
	 * NL->EX, PR->EX, an in-place conversion deadlock.)  So, after
	 * recovery, grant conversions without considering order.
	 */

	if (conv && recover)
		return 1;

	/*
	 * 6-5: But the default algorithm for deciding whether to grant or
	 * queue conversion requests does not by itself guarantee that such
	 * requests are serviced on a "first come first serve" basis.  This, in
	 * turn, can lead to a phenomenon known as "indefinate postponement".
	 *
	 * 6-7: This issue is dealt with by using the optional QUECVT flag with
	 * the system service employed to request a lock conversion.  This flag
	 * forces certain conversion requests to be queued, even if they are
	 * compatible with the granted modes of other locks on the same
	 * resource.  Thus, the use of this flag results in conversion requests
	 * being ordered on a "first come first servce" basis.
	 *
	 * DCT: This condition is all about new conversions being able to occur
	 * "in place" while the lock remains on the granted queue (assuming
	 * nothing else conflicts.)  IOW if QUECVT isn't set, a conversion
	 * doesn't _have_ to go onto the convert queue where it's processed in
	 * order.  The "now" variable is necessary to distinguish converts
	 * being received and processed for the first time now, because once a
	 * convert is moved to the conversion queue the condition below applies
	 * requiring fifo granting.
	 */

	if (now && conv && !(lkb->lkb_exflags & DLM_LKF_QUECVT))
		return 1;

	/*
	 * Even if the convert is compat with all granted locks,
	 * QUECVT forces it behind other locks on the convert queue.
	 */

	if (now && conv && (lkb->lkb_exflags & DLM_LKF_QUECVT)) {
		if (list_empty(&r->res_convertqueue))
			return 1;
		else
			return 0;
	}

	/*
	 * The NOORDER flag is set to avoid the standard vms rules on grant
	 * order.
	 */

	if (lkb->lkb_exflags & DLM_LKF_NOORDER)
		return 1;

	/*
	 * 6-3: Once in that queue [CONVERTING], a conversion request cannot be
	 * granted until all other conversion requests ahead of it are granted
	 * and/or canceled.
	 */

	if (!now && conv && first_in_list(lkb, &r->res_convertqueue))
		return 1;

	/*
	 * 6-4: By default, a new request is immediately granted only if all
	 * three of the following conditions are satisfied when the request is
	 * issued:
	 * - The queue of ungranted conversion requests for the resource is
	 *   empty.
	 * - The queue of ungranted new requests for the resource is empty.
	 * - The mode of the new request is compatible with the most
	 *   restrictive mode of all granted locks on the resource.
	 */

	if (now && !conv && list_empty(&r->res_convertqueue) &&
	    list_empty(&r->res_waitqueue))
		return 1;

	/*
	 * 6-4: Once a lock request is in the queue of ungranted new requests,
	 * it cannot be granted until the queue of ungranted conversion
	 * requests is empty, all ungranted new requests ahead of it are
	 * granted and/or canceled, and it is compatible with the granted mode
	 * of the most restrictive lock granted on the resource.
	 */

	if (!now && !conv && list_empty(&r->res_convertqueue) &&
	    first_in_list(lkb, &r->res_waitqueue))
		return 1;

	return 0;
}

static int can_be_granted(struct dlm_rsb *r, struct dlm_lkb *lkb, int now,
			  int recover, int *err)
{
	int rv;
	int8_t alt = 0, rqmode = lkb->lkb_rqmode;
	int8_t is_convert = (lkb->lkb_grmode != DLM_LOCK_IV);

	if (err)
		*err = 0;

	rv = _can_be_granted(r, lkb, now, recover);
	if (rv)
		goto out;

	/*
	 * The CONVDEADLK flag is non-standard and tells the dlm to resolve
	 * conversion deadlocks by demoting grmode to NL, otherwise the dlm
	 * cancels one of the locks.
	 */

	if (is_convert && can_be_queued(lkb) &&
	    conversion_deadlock_detect(r, lkb)) {
		if (lkb->lkb_exflags & DLM_LKF_CONVDEADLK) {
			lkb->lkb_grmode = DLM_LOCK_NL;
			lkb->lkb_sbflags |= DLM_SBF_DEMOTED;
		} else if (!(lkb->lkb_exflags & DLM_LKF_NODLCKWT)) {
			if (err)
				*err = -EDEADLK;
			else {
				log_print("can_be_granted deadlock %x now %d",
					  lkb->lkb_id, now);
				dlm_dump_rsb(r);
			}
		}
		goto out;
	}

	/*
	 * The ALTPR and ALTCW flags are non-standard and tell the dlm to try
	 * to grant a request in a mode other than the normal rqmode.  It's a
	 * simple way to provide a big optimization to applications that can
	 * use them.
	 */

	if (rqmode != DLM_LOCK_PR && (lkb->lkb_exflags & DLM_LKF_ALTPR))
		alt = DLM_LOCK_PR;
	else if (rqmode != DLM_LOCK_CW && (lkb->lkb_exflags & DLM_LKF_ALTCW))
		alt = DLM_LOCK_CW;

	if (alt) {
		lkb->lkb_rqmode = alt;
		rv = _can_be_granted(r, lkb, now, 0);
		if (rv)
			lkb->lkb_sbflags |= DLM_SBF_ALTMODE;
		else
			lkb->lkb_rqmode = rqmode;
	}
 out:
	return rv;
}

/* FIXME: I don't think that can_be_granted() can/will demote or find deadlock
   for locks pending on the convert list.  Once verified (watch for these
   log_prints), we should be able to just call _can_be_granted() and not
   bother with the demote/deadlk cases here (and there's no easy way to deal
   with a deadlk here, we'd have to generate something like grant_lock with
   the deadlk error.) */

/* Returns the highest requested mode of all blocked conversions; sets
   cw if there's a blocked conversion to DLM_LOCK_CW. */

static int grant_pending_convert(struct dlm_rsb *r, int high, int *cw,
				 unsigned int *count)
{
	struct dlm_lkb *lkb, *s;
	int recover = rsb_flag(r, RSB_RECOVER_GRANT);
	int hi, demoted, quit, grant_restart, demote_restart;
	int deadlk;

	quit = 0;
 restart:
	grant_restart = 0;
	demote_restart = 0;
	hi = DLM_LOCK_IV;

	list_for_each_entry_safe(lkb, s, &r->res_convertqueue, lkb_statequeue) {
		demoted = is_demoted(lkb);
		deadlk = 0;

		if (can_be_granted(r, lkb, 0, recover, &deadlk)) {
			grant_lock_pending(r, lkb);
			grant_restart = 1;
			if (count)
				(*count)++;
			continue;
		}

		if (!demoted && is_demoted(lkb)) {
			log_print("WARN: pending demoted %x node %d %s",
				  lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
			demote_restart = 1;
			continue;
		}

		if (deadlk) {
			log_print("WARN: pending deadlock %x node %d %s",
				  lkb->lkb_id, lkb->lkb_nodeid, r->res_name);
			dlm_dump_rsb(r);
			continue;
		}

		hi = max_t(int, lkb->lkb_rqmode, hi);

		if (cw && lkb->lkb_rqmode == DLM_LOCK_CW)
			*cw = 1;
	}

	if (grant_restart)
		goto restart;
	if (demote_restart && !quit) {
		quit = 1;
		goto restart;
	}

	return max_t(int, high, hi);
}

static int grant_pending_wait(struct dlm_rsb *r, int high, int *cw,
			      unsigned int *count)
{
	struct dlm_lkb *lkb, *s;

	list_for_each_entry_safe(lkb, s, &r->res_waitqueue, lkb_statequeue) {
		if (can_be_granted(r, lkb, 0, 0, NULL)) {
			grant_lock_pending(r, lkb);
			if (count)
				(*count)++;
		} else {
			high = max_t(int, lkb->lkb_rqmode, high);
			if (lkb->lkb_rqmode == DLM_LOCK_CW)
				*cw = 1;
		}
	}

	return high;
}

/* cw of 1 means there's a lock with a rqmode of DLM_LOCK_CW that's blocked
   on either the convert or waiting queue.
   high is the largest rqmode of all locks blocked on the convert or
   waiting queue. */

static int lock_requires_bast(struct dlm_lkb *gr, int high, int cw)
{
	if (gr->lkb_grmode == DLM_LOCK_PR && cw) {
		if (gr->lkb_highbast < DLM_LOCK_EX)
			return 1;
		return 0;
	}

	if (gr->lkb_highbast < high &&
	    !__dlm_compat_matrix[gr->lkb_grmode+1][high+1])
		return 1;
	return 0;
}

static void grant_pending_locks(struct dlm_rsb *r, unsigned int *count)
{
	struct dlm_lkb *lkb, *s;
	int high = DLM_LOCK_IV;
	int cw = 0;

	if (!is_master(r)) {
		log_print("grant_pending_locks r nodeid %d", r->res_nodeid);
		dlm_dump_rsb(r);
		return;
	}

	high = grant_pending_convert(r, high, &cw, count);
	high = grant_pending_wait(r, high, &cw, count);

	if (high == DLM_LOCK_IV)
		return;

	/*
	 * If there are locks left on the wait/convert queue then send blocking
	 * ASTs to granted locks based on the largest requested mode (high)
	 * found above.
	 */

	list_for_each_entry_safe(lkb, s, &r->res_grantqueue, lkb_statequeue) {
		if (lkb->lkb_bastfn && lock_requires_bast(lkb, high, cw)) {
			if (cw && high == DLM_LOCK_PR &&
			    lkb->lkb_grmode == DLM_LOCK_PR)
				queue_bast(r, lkb, DLM_LOCK_CW);
			else
				queue_bast(r, lkb, high);
			lkb->lkb_highbast = high;
		}
	}
}

static int modes_require_bast(struct dlm_lkb *gr, struct dlm_lkb *rq)
{
	if ((gr->lkb_grmode == DLM_LOCK_PR && rq->lkb_rqmode == DLM_LOCK_CW) ||
	    (gr->lkb_grmode == DLM_LOCK_CW && rq->lkb_rqmode == DLM_LOCK_PR)) {
		if (gr->lkb_highbast < DLM_LOCK_EX)
			return 1;
		return 0;
	}

	if (gr->lkb_highbast < rq->lkb_rqmode && !modes_compat(gr, rq))
		return 1;
	return 0;
}

static void send_bast_queue(struct dlm_rsb *r, struct list_head *head,
			    struct dlm_lkb *lkb)
{
	struct dlm_lkb *gr;

	list_for_each_entry(gr, head, lkb_statequeue) {
		/* skip self when sending basts to convertqueue */
		if (gr == lkb)
			continue;
		if (gr->lkb_bastfn && modes_require_bast(gr, lkb)) {
			queue_bast(r, gr, lkb->lkb_rqmode);
			gr->lkb_highbast = lkb->lkb_rqmode;
		}
	}
}

static void send_blocking_asts(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	send_bast_queue(r, &r->res_grantqueue, lkb);
}

static void send_blocking_asts_all(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	send_bast_queue(r, &r->res_grantqueue, lkb);
	send_bast_queue(r, &r->res_convertqueue, lkb);
}

/* set_master(r, lkb) -- set the master nodeid of a resource

   The purpose of this function is to set the nodeid field in the given
   lkb using the nodeid field in the given rsb.  If the rsb's nodeid is
   known, it can just be copied to the lkb and the function will return
   0.  If the rsb's nodeid is _not_ known, it needs to be looked up
   before it can be copied to the lkb.

   When the rsb nodeid is being looked up remotely, the initial lkb
   causing the lookup is kept on the ls_waiters list waiting for the
   lookup reply.  Other lkb's waiting for the same rsb lookup are kept
   on the rsb's res_lookup list until the master is verified.

   Return values:
   0: nodeid is set in rsb/lkb and the caller should go ahead and use it
   1: the rsb master is not available and the lkb has been placed on
      a wait queue
*/

static int set_master(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	int our_nodeid = dlm_our_nodeid();

	if (rsb_flag(r, RSB_MASTER_UNCERTAIN)) {
		rsb_clear_flag(r, RSB_MASTER_UNCERTAIN);
		r->res_first_lkid = lkb->lkb_id;
		lkb->lkb_nodeid = r->res_nodeid;
		return 0;
	}

	if (r->res_first_lkid && r->res_first_lkid != lkb->lkb_id) {
		list_add_tail(&lkb->lkb_rsb_lookup, &r->res_lookup);
		return 1;
	}

	if (r->res_master_nodeid == our_nodeid) {
		lkb->lkb_nodeid = 0;
		return 0;
	}

	if (r->res_master_nodeid) {
		lkb->lkb_nodeid = r->res_master_nodeid;
		return 0;
	}

	if (dlm_dir_nodeid(r) == our_nodeid) {
		/* This is a somewhat unusual case; find_rsb will usually
		   have set res_master_nodeid when dir nodeid is local, but
		   there are cases where we become the dir node after we've
		   past find_rsb and go through _request_lock again.
		   confirm_master() or process_lookup_list() needs to be
		   called after this. */
		log_debug(r->res_ls, "set_master %x self master %d dir %d %s",
			  lkb->lkb_id, r->res_master_nodeid, r->res_dir_nodeid,
			  r->res_name);
		r->res_master_nodeid = our_nodeid;
		r->res_nodeid = 0;
		lkb->lkb_nodeid = 0;
		return 0;
	}

	wait_pending_remove(r);

	r->res_first_lkid = lkb->lkb_id;
	send_lookup(r, lkb);
	return 1;
}

static void process_lookup_list(struct dlm_rsb *r)
{
	struct dlm_lkb *lkb, *safe;

	list_for_each_entry_safe(lkb, safe, &r->res_lookup, lkb_rsb_lookup) {
		list_del_init(&lkb->lkb_rsb_lookup);
		_request_lock(r, lkb);
		schedule();
	}
}

/* confirm_master -- confirm (or deny) an rsb's master nodeid */

static void confirm_master(struct dlm_rsb *r, int error)
{
	struct dlm_lkb *lkb;

	if (!r->res_first_lkid)
		return;

	switch (error) {
	case 0:
	case -EINPROGRESS:
		r->res_first_lkid = 0;
		process_lookup_list(r);
		break;

	case -EAGAIN:
	case -EBADR:
	case -ENOTBLK:
		/* the remote request failed and won't be retried (it was
		   a NOQUEUE, or has been canceled/unlocked); make a waiting
		   lkb the first_lkid */

		r->res_first_lkid = 0;

		if (!list_empty(&r->res_lookup)) {
			lkb = list_entry(r->res_lookup.next, struct dlm_lkb,
					 lkb_rsb_lookup);
			list_del_init(&lkb->lkb_rsb_lookup);
			r->res_first_lkid = lkb->lkb_id;
			_request_lock(r, lkb);
		}
		break;

	default:
		log_error(r->res_ls, "confirm_master unknown error %d", error);
	}
}

static int set_lock_args(int mode, struct dlm_lksb *lksb, uint32_t flags,
			 int namelen, unsigned long timeout_cs,
			 void (*ast) (void *astparam),
			 void *astparam,
			 void (*bast) (void *astparam, int mode),
			 struct dlm_args *args)
{
	int rv = -EINVAL;

	/* check for invalid arg usage */

	if (mode < 0 || mode > DLM_LOCK_EX)
		goto out;

	if (!(flags & DLM_LKF_CONVERT) && (namelen > DLM_RESNAME_MAXLEN))
		goto out;

	if (flags & DLM_LKF_CANCEL)
		goto out;

	if (flags & DLM_LKF_QUECVT && !(flags & DLM_LKF_CONVERT))
		goto out;

	if (flags & DLM_LKF_CONVDEADLK && !(flags & DLM_LKF_CONVERT))
		goto out;

	if (flags & DLM_LKF_CONVDEADLK && flags & DLM_LKF_NOQUEUE)
		goto out;

	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_CONVERT)
		goto out;

	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_QUECVT)
		goto out;

	if (flags & DLM_LKF_EXPEDITE && flags & DLM_LKF_NOQUEUE)
		goto out;

	if (flags & DLM_LKF_EXPEDITE && mode != DLM_LOCK_NL)
		goto out;

	if (!ast || !lksb)
		goto out;

	if (flags & DLM_LKF_VALBLK && !lksb->sb_lvbptr)
		goto out;

	if (flags & DLM_LKF_CONVERT && !lksb->sb_lkid)
		goto out;

	/* these args will be copied to the lkb in validate_lock_args,
	   it cannot be done now because when converting locks, fields in
	   an active lkb cannot be modified before locking the rsb */

	args->flags = flags;
	args->astfn = ast;
	args->astparam = astparam;
	args->bastfn = bast;
	args->timeout = timeout_cs;
	args->mode = mode;
	args->lksb = lksb;
	rv = 0;
 out:
	return rv;
}

static int set_unlock_args(uint32_t flags, void *astarg, struct dlm_args *args)
{
	if (flags & ~(DLM_LKF_CANCEL | DLM_LKF_VALBLK | DLM_LKF_IVVALBLK |
 		      DLM_LKF_FORCEUNLOCK))
		return -EINVAL;

	if (flags & DLM_LKF_CANCEL && flags & DLM_LKF_FORCEUNLOCK)
		return -EINVAL;

	args->flags = flags;
	args->astparam = astarg;
	return 0;
}

static int validate_lock_args(struct dlm_ls *ls, struct dlm_lkb *lkb,
			      struct dlm_args *args)
{
	int rv = -EINVAL;

	if (args->flags & DLM_LKF_CONVERT) {
		if (lkb->lkb_flags & DLM_IFL_MSTCPY)
			goto out;

		if (args->flags & DLM_LKF_QUECVT &&
		    !__quecvt_compat_matrix[lkb->lkb_grmode+1][args->mode+1])
			goto out;

		rv = -EBUSY;
		if (lkb->lkb_status != DLM_LKSTS_GRANTED)
			goto out;

		if (lkb->lkb_wait_type)
			goto out;

		if (is_overlap(lkb))
			goto out;
	}

	lkb->lkb_exflags = args->flags;
	lkb->lkb_sbflags = 0;
	lkb->lkb_astfn = args->astfn;
	lkb->lkb_astparam = args->astparam;
	lkb->lkb_bastfn = args->bastfn;
	lkb->lkb_rqmode = args->mode;
	lkb->lkb_lksb = args->lksb;
	lkb->lkb_lvbptr = args->lksb->sb_lvbptr;
	lkb->lkb_ownpid = (int) current->pid;
	lkb->lkb_timeout_cs = args->timeout;
	rv = 0;
 out:
	if (rv)
		log_debug(ls, "validate_lock_args %d %x %x %x %d %d %s",
			  rv, lkb->lkb_id, lkb->lkb_flags, args->flags,
			  lkb->lkb_status, lkb->lkb_wait_type,
			  lkb->lkb_resource->res_name);
	return rv;
}

/* when dlm_unlock() sees -EBUSY with CANCEL/FORCEUNLOCK it returns 0
   for success */

/* note: it's valid for lkb_nodeid/res_nodeid to be -1 when we get here
   because there may be a lookup in progress and it's valid to do
   cancel/unlockf on it */

static int validate_unlock_args(struct dlm_lkb *lkb, struct dlm_args *args)
{
	struct dlm_ls *ls = lkb->lkb_resource->res_ls;
	int rv = -EINVAL;

	if (lkb->lkb_flags & DLM_IFL_MSTCPY) {
		log_error(ls, "unlock on MSTCPY %x", lkb->lkb_id);
		dlm_print_lkb(lkb);
		goto out;
	}

	/* an lkb may still exist even though the lock is EOL'ed due to a
	   cancel, unlock or failed noqueue request; an app can't use these
	   locks; return same error as if the lkid had not been found at all */

	if (lkb->lkb_flags & DLM_IFL_ENDOFLIFE) {
		log_debug(ls, "unlock on ENDOFLIFE %x", lkb->lkb_id);
		rv = -ENOENT;
		goto out;
	}

	/* an lkb may be waiting for an rsb lookup to complete where the
	   lookup was initiated by another lock */

	if (!list_empty(&lkb->lkb_rsb_lookup)) {
		if (args->flags & (DLM_LKF_CANCEL | DLM_LKF_FORCEUNLOCK)) {
			log_debug(ls, "unlock on rsb_lookup %x", lkb->lkb_id);
			list_del_init(&lkb->lkb_rsb_lookup);
			queue_cast(lkb->lkb_resource, lkb,
				   args->flags & DLM_LKF_CANCEL ?
				   -DLM_ECANCEL : -DLM_EUNLOCK);
			unhold_lkb(lkb); /* undoes create_lkb() */
		}
		/* caller changes -EBUSY to 0 for CANCEL and FORCEUNLOCK */
		rv = -EBUSY;
		goto out;
	}

	/* cancel not allowed with another cancel/unlock in progress */

	if (args->flags & DLM_LKF_CANCEL) {
		if (lkb->lkb_exflags & DLM_LKF_CANCEL)
			goto out;

		if (is_overlap(lkb))
			goto out;

		/* don't let scand try to do a cancel */
		del_timeout(lkb);

		if (lkb->lkb_flags & DLM_IFL_RESEND) {
			lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
			rv = -EBUSY;
			goto out;
		}

		/* there's nothing to cancel */
		if (lkb->lkb_status == DLM_LKSTS_GRANTED &&
		    !lkb->lkb_wait_type) {
			rv = -EBUSY;
			goto out;
		}

		switch (lkb->lkb_wait_type) {
		case DLM_MSG_LOOKUP:
		case DLM_MSG_REQUEST:
			lkb->lkb_flags |= DLM_IFL_OVERLAP_CANCEL;
			rv = -EBUSY;
			goto out;
		case DLM_MSG_UNLOCK:
		case DLM_MSG_CANCEL:
			goto out;
		}
		/* add_to_waiters() will set OVERLAP_CANCEL */
		goto out_ok;
	}

	/* do we need to allow a force-unlock if there's a normal unlock
	   already in progress?  in what conditions could the normal unlock
	   fail such that we'd want to send a force-unlock to be sure? */

	if (args->flags & DLM_LKF_FORCEUNLOCK) {
		if (lkb->lkb_exflags & DLM_LKF_FORCEUNLOCK)
			goto out;

		if (is_overlap_unlock(lkb))
			goto out;

		/* don't let scand try to do a cancel */
		del_timeout(lkb);

		if (lkb->lkb_flags & DLM_IFL_RESEND) {
			lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
			rv = -EBUSY;
			goto out;
		}

		switch (lkb->lkb_wait_type) {
		case DLM_MSG_LOOKUP:
		case DLM_MSG_REQUEST:
			lkb->lkb_flags |= DLM_IFL_OVERLAP_UNLOCK;
			rv = -EBUSY;
			goto out;
		case DLM_MSG_UNLOCK:
			goto out;
		}
		/* add_to_waiters() will set OVERLAP_UNLOCK */
		goto out_ok;
	}

	/* normal unlock not allowed if there's any op in progress */
	rv = -EBUSY;
	if (lkb->lkb_wait_type || lkb->lkb_wait_count)
		goto out;

 out_ok:
	/* an overlapping op shouldn't blow away exflags from other op */
	lkb->lkb_exflags |= args->flags;
	lkb->lkb_sbflags = 0;
	lkb->lkb_astparam = args->astparam;
	rv = 0;
 out:
	if (rv)
		log_debug(ls, "validate_unlock_args %d %x %x %x %x %d %s", rv,
			  lkb->lkb_id, lkb->lkb_flags, lkb->lkb_exflags,
			  args->flags, lkb->lkb_wait_type,
			  lkb->lkb_resource->res_name);
	return rv;
}

/*
 * Four stage 4 varieties:
 * do_request(), do_convert(), do_unlock(), do_cancel()
 * These are called on the master node for the given lock and
 * from the central locking logic.
 */

static int do_request(struct dlm_rsb *r, struct dlm_lkb *lkb)
{
	int error = 0;

	if (can_be_granted(r, lkb, 1, 0, NULL)) {
		grant_lock(r, lkb);
		queue_cast(r, lkb, 0);
		goto out;
	}

	if (can_be_queued(lkb)) {
		error = -EINPROGRESS;
		add_lkb(r, lkb, DLM_LKSTS_WAITING);
		add_timeout(lkb);
		goto out;
	}

	error = -EAGAIN;
	queue_cast(r, lkb, -EAGAIN);
 out:
	return error;
}

static void do_request_effects(struct dlm_rsb *r, struct dlm_lkb *lkb,
			       int error)
{
	switch (error) {
	case -EAGAIN:
		if (force_blocking_asts(lkb))
			send_blocking_asts_all(r, lkb);
		break;
	case -EINPROGRESS: