[linux-block.git] / fs / ocfs2 / dlm / dlmthread.c

/* -*- mode: c; c-basic-offset: 8; -*-
 * vim: noexpandtab sw=8 ts=8 sts=0:
 *
 * dlmthread.c
 *
 * standalone DLM module
 *
 * Copyright (C) 2004 Oracle.  All rights reserved.
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public
 * License along with this program; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 021110-1307, USA.
 *
 */


#include <linux/module.h>
#include <linux/fs.h>
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/utsname.h>
#include <linux/init.h>
#include <linux/sysctl.h>
#include <linux/random.h>
#include <linux/blkdev.h>
#include <linux/socket.h>
#include <linux/inet.h>
#include <linux/timer.h>
#include <linux/kthread.h>
#include <linux/delay.h>


#include "cluster/heartbeat.h"
#include "cluster/nodemanager.h"
#include "cluster/tcp.h"

#include "dlmapi.h"
#include "dlmcommon.h"
#include "dlmdomain.h"

#define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_THREAD)
#include "cluster/masklog.h"

static int dlm_thread(void *data);

static void dlm_flush_asts(struct dlm_ctxt *dlm);

#define dlm_lock_is_remote(dlm, lock)     ((lock)->ml.node != (dlm)->node_num)

/* will exit holding res->spinlock, but may drop in function */
/* waits until flags are cleared on res->state */
void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
{
	DECLARE_WAITQUEUE(wait, current);

	assert_spin_locked(&res->spinlock);

	add_wait_queue(&res->wq, &wait);
repeat:
	set_current_state(TASK_UNINTERRUPTIBLE);
	if (res->state & flags) {
		spin_unlock(&res->spinlock);
		schedule();
		spin_lock(&res->spinlock);
		goto repeat;
	}
	remove_wait_queue(&res->wq, &wait);
	current->state = TASK_RUNNING;
}


int __dlm_lockres_unused(struct dlm_lock_resource *res)
{
	if (list_empty(&res->granted) &&
	    list_empty(&res->converting) &&
	    list_empty(&res->blocked) &&
	    list_empty(&res->dirty))
		return 1;
	return 0;
}


/* Call whenever you may have added or deleted something from one of
 * the lockres queue's. This will figure out whether it belongs on the
 * unused list or not and does the appropriate thing. */
void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
			      struct dlm_lock_resource *res)
{
	mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);

	assert_spin_locked(&dlm->spinlock);
	assert_spin_locked(&res->spinlock);

	if (__dlm_lockres_unused(res)){
		if (list_empty(&res->purge)) {
			mlog(0, "putting lockres %.*s from purge list\n",
			     res->lockname.len, res->lockname.name);

			res->last_used = jiffies;
			list_add_tail(&res->purge, &dlm->purge_list);
			dlm->purge_count++;
		}
	} else if (!list_empty(&res->purge)) {
		mlog(0, "removing lockres %.*s from purge list\n",
		     res->lockname.len, res->lockname.name);

		list_del_init(&res->purge);
		dlm->purge_count--;
	}
}

void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
			    struct dlm_lock_resource *res)
{
	mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
	spin_lock(&dlm->spinlock);
	spin_lock(&res->spinlock);

	__dlm_lockres_calc_usage(dlm, res);

	spin_unlock(&res->spinlock);
	spin_unlock(&dlm->spinlock);
}

/* TODO: Eventual API: Called with the dlm spinlock held, may drop it
 * to do migration, but will re-acquire before exit. */
void dlm_purge_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *lockres)
{
	int master;
	int ret;

	spin_lock(&lockres->spinlock);
	master = lockres->owner == dlm->node_num;
	spin_unlock(&lockres->spinlock);

	mlog(0, "purging lockres %.*s, master = %d\n", lockres->lockname.len,
	     lockres->lockname.name, master);

	/* Non master is the easy case -- no migration required, just
	 * quit. */
	if (!master)
		goto finish;

	/* Wheee! Migrate lockres here! */
	spin_unlock(&dlm->spinlock);
again:

	ret = dlm_migrate_lockres(dlm, lockres, O2NM_MAX_NODES);
	if (ret == -ENOTEMPTY) {
		mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
		     lockres->lockname.len, lockres->lockname.name);

		BUG();
	} else if (ret < 0) {
		mlog(ML_NOTICE, "lockres %.*s: migrate failed, retrying\n",
		     lockres->lockname.len, lockres->lockname.name);
		msleep(100);
		goto again;
	}

	spin_lock(&dlm->spinlock);

finish:
	if (!list_empty(&lockres->purge)) {
		list_del_init(&lockres->purge);
		dlm->purge_count--;
	}
	__dlm_unhash_lockres(lockres);
}

static void dlm_run_purge_list(struct dlm_ctxt *dlm,
			       int purge_now)
{
	unsigned int run_max, unused;
	unsigned long purge_jiffies;
	struct dlm_lock_resource *lockres;

	spin_lock(&dlm->spinlock);
	run_max = dlm->purge_count;

	while(run_max && !list_empty(&dlm->purge_list)) {
		run_max--;

		lockres = list_entry(dlm->purge_list.next,
				     struct dlm_lock_resource, purge);

		/* Status of the lockres *might* change so double
		 * check. If the lockres is unused, holding the dlm
		 * spinlock will prevent people from getting and more
		 * refs on it -- there's no need to keep the lockres
		 * spinlock. */
		spin_lock(&lockres->spinlock);
		unused = __dlm_lockres_unused(lockres);
		spin_unlock(&lockres->spinlock);

		if (!unused)
			continue;

		purge_jiffies = lockres->last_used +
			msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);

		/* Make sure that we want to be processing this guy at
		 * this time. */
		if (!purge_now && time_after(purge_jiffies, jiffies)) {
			/* Since resources are added to the purge list
			 * in tail order, we can stop at the first
			 * unpurgable resource -- anyone added after
			 * him will have a greater last_used value */
			break;
		}

		list_del_init(&lockres->purge);
		dlm->purge_count--;

		/* This may drop and reacquire the dlm spinlock if it
		 * has to do migration. */
		mlog(0, "calling dlm_purge_lockres!\n");
		dlm_purge_lockres(dlm, lockres);
		mlog(0, "DONE calling dlm_purge_lockres!\n");

		/* Avoid adding any scheduling latencies */
		cond_resched_lock(&dlm->spinlock);
	}

	spin_unlock(&dlm->spinlock);
}

static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
			      struct dlm_lock_resource *res)
{
	struct dlm_lock *lock, *target;
	struct list_head *iter;
	struct list_head *head;
	int can_grant = 1;

	//mlog(0, "res->lockname.len=%d\n", res->lockname.len);
	//mlog(0, "res->lockname.name=%p\n", res->lockname.name);
	//mlog(0, "shuffle res %.*s\n", res->lockname.len,
	//	  res->lockname.name);

	/* because this function is called with the lockres
	 * spinlock, and because we know that it is not migrating/
	 * recovering/in-progress, it is fine to reserve asts and
	 * basts right before queueing them all throughout */
	assert_spin_locked(&res->spinlock);
	BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING|
			      DLM_LOCK_RES_RECOVERING|
			      DLM_LOCK_RES_IN_PROGRESS)));

converting:
	if (list_empty(&res->converting))
		goto blocked;
	mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
	     res->lockname.name);

	target = list_entry(res->converting.next, struct dlm_lock, list);
	if (target->ml.convert_type == LKM_IVMODE) {
		mlog(ML_ERROR, "%.*s: converting a lock with no "
		     "convert_type!\n", res->lockname.len, res->lockname.name);
		BUG();
	}
	head = &res->granted;
	list_for_each(iter, head) {
		lock = list_entry(iter, struct dlm_lock, list);
		if (lock==target)
			continue;
		if (!dlm_lock_compatible(lock->ml.type,
					 target->ml.convert_type)) {
			can_grant = 0;
			/* queue the BAST if not already */
			if (lock->ml.highest_blocked == LKM_IVMODE) {
				__dlm_lockres_reserve_ast(res);
				dlm_queue_bast(dlm, lock);
			}
			/* update the highest_blocked if needed */
			if (lock->ml.highest_blocked < target->ml.convert_type)
				lock->ml.highest_blocked =
					target->ml.convert_type;
		}
	}
	head = &res->converting;
	list_for_each(iter, head) {
		lock = list_entry(iter, struct dlm_lock, list);
		if (lock==target)
			continue;
		if (!dlm_lock_compatible(lock->ml.type,
					 target->ml.convert_type)) {
			can_grant = 0;
			if (lock->ml.highest_blocked == LKM_IVMODE) {
				__dlm_lockres_reserve_ast(res);
				dlm_queue_bast(dlm, lock);
			}
			if (lock->ml.highest_blocked < target->ml.convert_type)
				lock->ml.highest_blocked =
					target->ml.convert_type;
		}
	}

	/* we can convert the lock */
	if (can_grant) {
		spin_lock(&target->spinlock);
		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);

		mlog(0, "calling ast for converting lock: %.*s, have: %d, "
		     "granting: %d, node: %u\n", res->lockname.len,
		     res->lockname.name, target->ml.type,
		     target->ml.convert_type, target->ml.node);

		target->ml.type = target->ml.convert_type;
		target->ml.convert_type = LKM_IVMODE;
		list_move_tail(&target->list, &res->granted);

		BUG_ON(!target->lksb);
		target->lksb->status = DLM_NORMAL;

		spin_unlock(&target->spinlock);

		__dlm_lockres_reserve_ast(res);
		dlm_queue_ast(dlm, target);
		/* go back and check for more */
		goto converting;
	}

blocked:
	if (list_empty(&res->blocked))
		goto leave;
	target = list_entry(res->blocked.next, struct dlm_lock, list);

	head = &res->granted;
	list_for_each(iter, head) {
		lock = list_entry(iter, struct dlm_lock, list);
		if (lock==target)
			continue;
		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
			can_grant = 0;
			if (lock->ml.highest_blocked == LKM_IVMODE) {
				__dlm_lockres_reserve_ast(res);
				dlm_queue_bast(dlm, lock);
			}
			if (lock->ml.highest_blocked < target->ml.type)
				lock->ml.highest_blocked = target->ml.type;
		}
	}

	head = &res->converting;
	list_for_each(iter, head) {
		lock = list_entry(iter, struct dlm_lock, list);
		if (lock==target)
			continue;
		if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
			can_grant = 0;
			if (lock->ml.highest_blocked == LKM_IVMODE) {
				__dlm_lockres_reserve_ast(res);
				dlm_queue_bast(dlm, lock);
			}
			if (lock->ml.highest_blocked < target->ml.type)
				lock->ml.highest_blocked = target->ml.type;
		}
	}

	/* we can grant the blocked lock (only
	 * possible if converting list empty) */
	if (can_grant) {
		spin_lock(&target->spinlock);
		BUG_ON(target->ml.highest_blocked != LKM_IVMODE);

		mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
		     "node: %u\n", res->lockname.len, res->lockname.name,
		     target->ml.type, target->ml.node);

		// target->ml.type is already correct
		list_move_tail(&target->list, &res->granted);

		BUG_ON(!target->lksb);
		target->lksb->status = DLM_NORMAL;

		spin_unlock(&target->spinlock);

		__dlm_lockres_reserve_ast(res);
		dlm_queue_ast(dlm, target);
		/* go back and check for more */
		goto converting;
	}

leave:
	return;
}

/* must have NO locks when calling this with res !=NULL * */
void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
{
	mlog_entry("dlm=%p, res=%p\n", dlm, res);
	if (res) {
		spin_lock(&dlm->spinlock);
		spin_lock(&res->spinlock);
		__dlm_dirty_lockres(dlm, res);
		spin_unlock(&res->spinlock);
		spin_unlock(&dlm->spinlock);
	}
	wake_up(&dlm->dlm_thread_wq);
}

void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
{
	mlog_entry("dlm=%p, res=%p\n", dlm, res);

	assert_spin_locked(&dlm->spinlock);
	assert_spin_locked(&res->spinlock);

	/* don't shuffle secondary queues */
	if ((res->owner == dlm->node_num) &&
	    !(res->state & DLM_LOCK_RES_DIRTY)) {
		list_add_tail(&res->dirty, &dlm->dirty_list);
		res->state |= DLM_LOCK_RES_DIRTY;
	}
}


/* Launch the NM thread for the mounted volume */
int dlm_launch_thread(struct dlm_ctxt *dlm)
{
	mlog(0, "starting dlm thread...\n");

	dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
	if (IS_ERR(dlm->dlm_thread_task)) {
		mlog_errno(PTR_ERR(dlm->dlm_thread_task));
		dlm->dlm_thread_task = NULL;
		return -EINVAL;
	}

	return 0;
}

void dlm_complete_thread(struct dlm_ctxt *dlm)
{
	if (dlm->dlm_thread_task) {
		mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
		kthread_stop(dlm->dlm_thread_task);
		dlm->dlm_thread_task = NULL;
	}
}

static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
{
	int empty;

	spin_lock(&dlm->spinlock);
	empty = list_empty(&dlm->dirty_list);
	spin_unlock(&dlm->spinlock);

	return empty;
}

static void dlm_flush_asts(struct dlm_ctxt *dlm)
{
	int ret;
	struct dlm_lock *lock;
	struct dlm_lock_resource *res;
	u8 hi;

	spin_lock(&dlm->ast_lock);
	while (!list_empty(&dlm->pending_asts)) {
		lock = list_entry(dlm->pending_asts.next,
				  struct dlm_lock, ast_list);
		/* get an extra ref on lock */
		dlm_lock_get(lock);
		res = lock->lockres;
		mlog(0, "delivering an ast for this lockres\n");

		BUG_ON(!lock->ast_pending);

		/* remove from list (including ref) */
		list_del_init(&lock->ast_list);
		dlm_lock_put(lock);
		spin_unlock(&dlm->ast_lock);

		if (lock->ml.node != dlm->node_num) {
			ret = dlm_do_remote_ast(dlm, res, lock);
			if (ret < 0)
				mlog_errno(ret);
		} else
			dlm_do_local_ast(dlm, res, lock);

		spin_lock(&dlm->ast_lock);

		/* possible that another ast was queued while
		 * we were delivering the last one */
		if (!list_empty(&lock->ast_list)) {
			mlog(0, "aha another ast got queued while "
			     "we were finishing the last one.  will "
			     "keep the ast_pending flag set.\n");
		} else
			lock->ast_pending = 0;

		/* drop the extra ref.
		 * this may drop it completely. */
		dlm_lock_put(lock);
		dlm_lockres_release_ast(dlm, res);
	}

	while (!list_empty(&dlm->pending_basts)) {
		lock = list_entry(dlm->pending_basts.next,
				  struct dlm_lock, bast_list);
		/* get an extra ref on lock */
		dlm_lock_get(lock);
		res = lock->lockres;

		BUG_ON(!lock->bast_pending);

		/* get the highest blocked lock, and reset */
		spin_lock(&lock->spinlock);
		BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
		hi = lock->ml.highest_blocked;
		lock->ml.highest_blocked = LKM_IVMODE;
		spin_unlock(&lock->spinlock);

		/* remove from list (including ref) */
		list_del_init(&lock->bast_list);
		dlm_lock_put(lock);
		spin_unlock(&dlm->ast_lock);

		mlog(0, "delivering a bast for this lockres "
		     "(blocked = %d\n", hi);

		if (lock->ml.node != dlm->node_num) {
			ret = dlm_send_proxy_bast(dlm, res, lock, hi);
			if (ret < 0)
				mlog_errno(ret);
		} else
			dlm_do_local_bast(dlm, res, lock, hi);

		spin_lock(&dlm->ast_lock);

		/* possible that another bast was queued while
		 * we were delivering the last one */
		if (!list_empty(&lock->bast_list)) {
			mlog(0, "aha another bast got queued while "
			     "we were finishing the last one.  will "
			     "keep the bast_pending flag set.\n");
		} else
			lock->bast_pending = 0;

		/* drop the extra ref.
		 * this may drop it completely. */
		dlm_lock_put(lock);
		dlm_lockres_release_ast(dlm, res);
	}
	wake_up(&dlm->ast_wq);
	spin_unlock(&dlm->ast_lock);
}


#define DLM_THREAD_TIMEOUT_MS (4 * 1000)
#define DLM_THREAD_MAX_DIRTY  100
#define DLM_THREAD_MAX_ASTS   10

static int dlm_thread(void *data)
{
	struct dlm_lock_resource *res;
	struct dlm_ctxt *dlm = data;
	unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);

	mlog(0, "dlm thread running for %s...\n", dlm->name);

	while (!kthread_should_stop()) {
		int n = DLM_THREAD_MAX_DIRTY;

		/* dlm_shutting_down is very point-in-time, but that
		 * doesn't matter as we'll just loop back around if we
		 * get false on the leading edge of a state
		 * transition. */
		dlm_run_purge_list(dlm, dlm_shutting_down(dlm));

		/* We really don't want to hold dlm->spinlock while
		 * calling dlm_shuffle_lists on each lockres that
		 * needs to have its queues adjusted and AST/BASTs
		 * run.  So let's pull each entry off the dirty_list
		 * and drop dlm->spinlock ASAP.  Once off the list,
		 * res->spinlock needs to be taken again to protect
		 * the queues while calling dlm_shuffle_lists.  */
		spin_lock(&dlm->spinlock);
		while (!list_empty(&dlm->dirty_list)) {
			int delay = 0;
			res = list_entry(dlm->dirty_list.next,
					 struct dlm_lock_resource, dirty);

			/* peel a lockres off, remove it from the list,
			 * unset the dirty flag and drop the dlm lock */
			BUG_ON(!res);
			dlm_lockres_get(res);

			spin_lock(&res->spinlock);
			res->state &= ~DLM_LOCK_RES_DIRTY;
			list_del_init(&res->dirty);
			spin_unlock(&res->spinlock);
			spin_unlock(&dlm->spinlock);

		 	/* lockres can be re-dirtied/re-added to the
			 * dirty_list in this gap, but that is ok */

			spin_lock(&res->spinlock);
			if (res->owner != dlm->node_num) {
				__dlm_print_one_lock_resource(res);
				mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
				     res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
				     res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
				     res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
				     res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
			}
			BUG_ON(res->owner != dlm->node_num);

			/* it is now ok to move lockreses in these states
			 * to the dirty list, assuming that they will only be
			 * dirty for a short while. */
			if (res->state & (DLM_LOCK_RES_IN_PROGRESS |
					  DLM_LOCK_RES_MIGRATING |
					  DLM_LOCK_RES_RECOVERING)) {
				/* move it to the tail and keep going */
				spin_unlock(&res->spinlock);
				mlog(0, "delaying list shuffling for in-"
				     "progress lockres %.*s, state=%d\n",
				     res->lockname.len, res->lockname.name,
				     res->state);
				delay = 1;
				goto in_progress;
			}

			/* at this point the lockres is not migrating/
			 * recovering/in-progress.  we have the lockres
			 * spinlock and do NOT have the dlm lock.
			 * safe to reserve/queue asts and run the lists. */

			mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
			     "res=%.*s\n", dlm->name,
			     res->lockname.len, res->lockname.name);

			/* called while holding lockres lock */
			dlm_shuffle_lists(dlm, res);
			spin_unlock(&res->spinlock);

			dlm_lockres_calc_usage(dlm, res);

in_progress:

			spin_lock(&dlm->spinlock);
			/* if the lock was in-progress, stick
			 * it on the back of the list */
			if (delay) {
				spin_lock(&res->spinlock);
				list_add_tail(&res->dirty, &dlm->dirty_list);
				res->state |= DLM_LOCK_RES_DIRTY;
				spin_unlock(&res->spinlock);
			}
			dlm_lockres_put(res);

			/* unlikely, but we may need to give time to
			 * other tasks */
			if (!--n) {
				mlog(0, "throttling dlm_thread\n");
				break;
			}
		}

		spin_unlock(&dlm->spinlock);
		dlm_flush_asts(dlm);

		/* yield and continue right away if there is more work to do */
		if (!n) {
			yield();
			continue;
		}

		wait_event_interruptible_timeout(dlm->dlm_thread_wq,
						 !dlm_dirty_list_empty(dlm) ||
						 kthread_should_stop(),
						 timeout);
	}

	mlog(0, "quitting DLM thread\n");
	return 0;
}
Commit	Line	Data
6714d8e8 KH	1	/* -- mode: c; c-basic-offset: 8; --
	2	* vim: noexpandtab sw=8 ts=8 sts=0:
	3	*
	4	* dlmthread.c
	5	*
	6	* standalone DLM module
	7	*
	8	* Copyright (C) 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
	27
	28	#include <linux/module.h>
	29	#include <linux/fs.h>
	30	#include <linux/types.h>
	31	#include <linux/slab.h>
	32	#include <linux/highmem.h>
	33	#include <linux/utsname.h>
	34	#include <linux/init.h>
	35	#include <linux/sysctl.h>
	36	#include <linux/random.h>
	37	#include <linux/blkdev.h>
	38	#include <linux/socket.h>
	39	#include <linux/inet.h>
	40	#include <linux/timer.h>
	41	#include <linux/kthread.h>
8d79d088	42	#include <linux/delay.h>
6714d8e8 KH	43
	44
	45	#include "cluster/heartbeat.h"
	46	#include "cluster/nodemanager.h"
	47	#include "cluster/tcp.h"
	48
	49	#include "dlmapi.h"
	50	#include "dlmcommon.h"
	51	#include "dlmdomain.h"
	52
	53	#define MLOG_MASK_PREFIX (ML_DLM\|ML_DLM_THREAD)
	54	#include "cluster/masklog.h"
	55
6714d8e8 KH	56	static int dlm_thread(void *data);
	57
	58	static void dlm_flush_asts(struct dlm_ctxt *dlm);
	59
	60	#define dlm_lock_is_remote(dlm, lock) ((lock)->ml.node != (dlm)->node_num)
	61
	62	/* will exit holding res->spinlock, but may drop in function */
	63	/* waits until flags are cleared on res->state */
	64	void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags)
	65	{
	66	DECLARE_WAITQUEUE(wait, current);
	67
	68	assert_spin_locked(&res->spinlock);
	69
	70	add_wait_queue(&res->wq, &wait);
	71	repeat:
	72	set_current_state(TASK_UNINTERRUPTIBLE);
	73	if (res->state & flags) {
	74	spin_unlock(&res->spinlock);
	75	schedule();
	76	spin_lock(&res->spinlock);
	77	goto repeat;
	78	}
	79	remove_wait_queue(&res->wq, &wait);
	80	current->state = TASK_RUNNING;
	81	}
	82
	83
69d72b06	84	int __dlm_lockres_unused(struct dlm_lock_resource *res)
6714d8e8 KH	85	{
	86	if (list_empty(&res->granted) &&
	87	list_empty(&res->converting) &&
	88	list_empty(&res->blocked) &&
	89	list_empty(&res->dirty))
	90	return 1;
	91	return 0;
	92	}
	93
	94
	95	/* Call whenever you may have added or deleted something from one of
	96	* the lockres queue's. This will figure out whether it belongs on the
	97	* unused list or not and does the appropriate thing. */
	98	void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
	99	struct dlm_lock_resource *res)
	100	{
	101	mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
	102
	103	assert_spin_locked(&dlm->spinlock);
	104	assert_spin_locked(&res->spinlock);
	105
	106	if (__dlm_lockres_unused(res)){
	107	if (list_empty(&res->purge)) {
	108	mlog(0, "putting lockres %.*s from purge list\n",
	109	res->lockname.len, res->lockname.name);
	110
	111	res->last_used = jiffies;
	112	list_add_tail(&res->purge, &dlm->purge_list);
	113	dlm->purge_count++;
	114	}
	115	} else if (!list_empty(&res->purge)) {
	116	mlog(0, "removing lockres %.*s from purge list\n",
	117	res->lockname.len, res->lockname.name);
	118
	119	list_del_init(&res->purge);
	120	dlm->purge_count--;
	121	}
	122	}
	123
	124	void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
	125	struct dlm_lock_resource *res)
	126	{
	127	mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
	128	spin_lock(&dlm->spinlock);
	129	spin_lock(&res->spinlock);
	130
	131	__dlm_lockres_calc_usage(dlm, res);
	132
	133	spin_unlock(&res->spinlock);
	134	spin_unlock(&dlm->spinlock);
	135	}
	136
	137	/* TODO: Eventual API: Called with the dlm spinlock held, may drop it
	138	* to do migration, but will re-acquire before exit. */
	139	void dlm_purge_lockres(struct dlm_ctxt dlm, struct dlm_lock_resource lockres)
	140	{
	141	int master;
	142	int ret;
	143
	144	spin_lock(&lockres->spinlock);
	145	master = lockres->owner == dlm->node_num;
	146	spin_unlock(&lockres->spinlock);
	147
	148	mlog(0, "purging lockres %.*s, master = %d\n", lockres->lockname.len,
149	lockres->lockname.name, master);
150
151	/* Non master is the easy case -- no migration required, just
152	* quit. */
153	if (!master)
154	goto finish;
155
156	/* Wheee! Migrate lockres here! */
157	spin_unlock(&dlm->spinlock);
158	again:
159
160	ret = dlm_migrate_lockres(dlm, lockres, O2NM_MAX_NODES);
161	if (ret == -ENOTEMPTY) {
162	mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
163	lockres->lockname.len, lockres->lockname.name);
164
165	BUG();
166	} else if (ret < 0) {
167	mlog(ML_NOTICE, "lockres %.*s: migrate failed, retrying\n",
168	lockres->lockname.len, lockres->lockname.name);
8d79d088	169	msleep(100);
6714d8e8 KH	170	goto again;
	171	}
	172
	173	spin_lock(&dlm->spinlock);
	174
	175	finish:
	176	if (!list_empty(&lockres->purge)) {
	177	list_del_init(&lockres->purge);
	178	dlm->purge_count--;
	179	}
	180	__dlm_unhash_lockres(lockres);
	181	}
	182
	183	static void dlm_run_purge_list(struct dlm_ctxt *dlm,
	184	int purge_now)
	185	{
	186	unsigned int run_max, unused;
	187	unsigned long purge_jiffies;
	188	struct dlm_lock_resource *lockres;
	189
	190	spin_lock(&dlm->spinlock);
	191	run_max = dlm->purge_count;
	192
	193	while(run_max && !list_empty(&dlm->purge_list)) {
	194	run_max--;
	195
	196	lockres = list_entry(dlm->purge_list.next,
	197	struct dlm_lock_resource, purge);
	198
	199	/* Status of the lockres might change so double
	200	* check. If the lockres is unused, holding the dlm
	201	* spinlock will prevent people from getting and more
	202	* refs on it -- there's no need to keep the lockres
	203	* spinlock. */
	204	spin_lock(&lockres->spinlock);
	205	unused = __dlm_lockres_unused(lockres);
	206	spin_unlock(&lockres->spinlock);
	207
	208	if (!unused)
	209	continue;
	210
	211	purge_jiffies = lockres->last_used +
	212	msecs_to_jiffies(DLM_PURGE_INTERVAL_MS);
	213
	214	/* Make sure that we want to be processing this guy at
	215	* this time. */
	216	if (!purge_now && time_after(purge_jiffies, jiffies)) {
	217	/* Since resources are added to the purge list
	218	* in tail order, we can stop at the first
	219	* unpurgable resource -- anyone added after
	220	* him will have a greater last_used value */
	221	break;
	222	}
	223
	224	list_del_init(&lockres->purge);
	225	dlm->purge_count--;
	226
	227	/* This may drop and reacquire the dlm spinlock if it
	228	* has to do migration. */
	229	mlog(0, "calling dlm_purge_lockres!\n");
	230	dlm_purge_lockres(dlm, lockres);
	231	mlog(0, "DONE calling dlm_purge_lockres!\n");
	232
	233	/* Avoid adding any scheduling latencies */
234	cond_resched_lock(&dlm->spinlock);
235	}
236
237	spin_unlock(&dlm->spinlock);
238	}
239
240	static void dlm_shuffle_lists(struct dlm_ctxt *dlm,
241	struct dlm_lock_resource *res)
242	{
243	struct dlm_lock lock, target;
244	struct list_head *iter;
245	struct list_head *head;
246	int can_grant = 1;
247
248	//mlog(0, "res->lockname.len=%d\n", res->lockname.len);
249	//mlog(0, "res->lockname.name=%p\n", res->lockname.name);
250	//mlog(0, "shuffle res %.*s\n", res->lockname.len,
251	// res->lockname.name);
252
253	/* because this function is called with the lockres
254	* spinlock, and because we know that it is not migrating/
255	* recovering/in-progress, it is fine to reserve asts and
256	* basts right before queueing them all throughout */
257	assert_spin_locked(&res->spinlock);
258	BUG_ON((res->state & (DLM_LOCK_RES_MIGRATING\|
259	DLM_LOCK_RES_RECOVERING\|
260	DLM_LOCK_RES_IN_PROGRESS)));
261
262	converting:
263	if (list_empty(&res->converting))
264	goto blocked;
265	mlog(0, "res %.*s has locks on a convert queue\n", res->lockname.len,
266	res->lockname.name);
267
268	target = list_entry(res->converting.next, struct dlm_lock, list);
269	if (target->ml.convert_type == LKM_IVMODE) {
270	mlog(ML_ERROR, "%.*s: converting a lock with no "
271	"convert_type!\n", res->lockname.len, res->lockname.name);
272	BUG();
273	}
274	head = &res->granted;
275	list_for_each(iter, head) {
276	lock = list_entry(iter, struct dlm_lock, list);
277	if (lock==target)
278	continue;
279	if (!dlm_lock_compatible(lock->ml.type,
280	target->ml.convert_type)) {
281	can_grant = 0;
282	/* queue the BAST if not already */
283	if (lock->ml.highest_blocked == LKM_IVMODE) {
284	__dlm_lockres_reserve_ast(res);
285	dlm_queue_bast(dlm, lock);
286	}
287	/* update the highest_blocked if needed */
288	if (lock->ml.highest_blocked < target->ml.convert_type)
289	lock->ml.highest_blocked =
290	target->ml.convert_type;
291	}
292	}
293	head = &res->converting;
294	list_for_each(iter, head) {
295	lock = list_entry(iter, struct dlm_lock, list);
296	if (lock==target)
297	continue;
298	if (!dlm_lock_compatible(lock->ml.type,
299	target->ml.convert_type)) {
300	can_grant = 0;
301	if (lock->ml.highest_blocked == LKM_IVMODE) {
302	__dlm_lockres_reserve_ast(res);
303	dlm_queue_bast(dlm, lock);
304	}
305	if (lock->ml.highest_blocked < target->ml.convert_type)
306	lock->ml.highest_blocked =
307	target->ml.convert_type;
308	}
309	}
310
311	/* we can convert the lock */
312	if (can_grant) {
313	spin_lock(&target->spinlock);
314	BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
315
316	mlog(0, "calling ast for converting lock: %.*s, have: %d, "
317	"granting: %d, node: %u\n", res->lockname.len,
318	res->lockname.name, target->ml.type,
319	target->ml.convert_type, target->ml.node);
320
321	target->ml.type = target->ml.convert_type;
322	target->ml.convert_type = LKM_IVMODE;
f116629d	323	list_move_tail(&target->list, &res->granted);
6714d8e8 KH	324
	325	BUG_ON(!target->lksb);
	326	target->lksb->status = DLM_NORMAL;
	327
	328	spin_unlock(&target->spinlock);
	329
	330	__dlm_lockres_reserve_ast(res);
	331	dlm_queue_ast(dlm, target);
	332	/* go back and check for more */
	333	goto converting;
	334	}
	335
	336	blocked:
	337	if (list_empty(&res->blocked))
	338	goto leave;
	339	target = list_entry(res->blocked.next, struct dlm_lock, list);
	340
	341	head = &res->granted;
	342	list_for_each(iter, head) {
	343	lock = list_entry(iter, struct dlm_lock, list);
	344	if (lock==target)
	345	continue;
	346	if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
	347	can_grant = 0;
	348	if (lock->ml.highest_blocked == LKM_IVMODE) {
	349	__dlm_lockres_reserve_ast(res);
	350	dlm_queue_bast(dlm, lock);
	351	}
	352	if (lock->ml.highest_blocked < target->ml.type)
	353	lock->ml.highest_blocked = target->ml.type;
	354	}
	355	}
	356
	357	head = &res->converting;
	358	list_for_each(iter, head) {
	359	lock = list_entry(iter, struct dlm_lock, list);
	360	if (lock==target)
	361	continue;
	362	if (!dlm_lock_compatible(lock->ml.type, target->ml.type)) {
	363	can_grant = 0;
	364	if (lock->ml.highest_blocked == LKM_IVMODE) {
	365	__dlm_lockres_reserve_ast(res);
	366	dlm_queue_bast(dlm, lock);
	367	}
	368	if (lock->ml.highest_blocked < target->ml.type)
	369	lock->ml.highest_blocked = target->ml.type;
	370	}
	371	}
	372
	373	/* we can grant the blocked lock (only
	374	* possible if converting list empty) */
	375	if (can_grant) {
	376	spin_lock(&target->spinlock);
	377	BUG_ON(target->ml.highest_blocked != LKM_IVMODE);
	378
	379	mlog(0, "calling ast for blocked lock: %.*s, granting: %d, "
	380	"node: %u\n", res->lockname.len, res->lockname.name,
	381	target->ml.type, target->ml.node);
	382
	383	// target->ml.type is already correct
f116629d	384	list_move_tail(&target->list, &res->granted);
6714d8e8 KH	385
	386	BUG_ON(!target->lksb);
	387	target->lksb->status = DLM_NORMAL;
	388
	389	spin_unlock(&target->spinlock);
	390
	391	__dlm_lockres_reserve_ast(res);
	392	dlm_queue_ast(dlm, target);
	393	/* go back and check for more */
	394	goto converting;
	395	}
	396
	397	leave:
	398	return;
	399	}
	400
	401	/* must have NO locks when calling this with res !=NULL * */
	402	void dlm_kick_thread(struct dlm_ctxt dlm, struct dlm_lock_resource res)
	403	{
	404	mlog_entry("dlm=%p, res=%p\n", dlm, res);
	405	if (res) {
	406	spin_lock(&dlm->spinlock);
	407	spin_lock(&res->spinlock);
	408	__dlm_dirty_lockres(dlm, res);
	409	spin_unlock(&res->spinlock);
	410	spin_unlock(&dlm->spinlock);
	411	}
	412	wake_up(&dlm->dlm_thread_wq);
	413	}
	414
	415	void __dlm_dirty_lockres(struct dlm_ctxt dlm, struct dlm_lock_resource res)
	416	{
	417	mlog_entry("dlm=%p, res=%p\n", dlm, res);
	418
	419	assert_spin_locked(&dlm->spinlock);
	420	assert_spin_locked(&res->spinlock);
	421
	422	/* don't shuffle secondary queues */
	423	if ((res->owner == dlm->node_num) &&
	424	!(res->state & DLM_LOCK_RES_DIRTY)) {
	425	list_add_tail(&res->dirty, &dlm->dirty_list);
	426	res->state \|= DLM_LOCK_RES_DIRTY;
	427	}
	428	}
	429
	430
	431	/* Launch the NM thread for the mounted volume */
	432	int dlm_launch_thread(struct dlm_ctxt *dlm)
	433	{
	434	mlog(0, "starting dlm thread...\n");
	435
	436	dlm->dlm_thread_task = kthread_run(dlm_thread, dlm, "dlm_thread");
	437	if (IS_ERR(dlm->dlm_thread_task)) {
	438	mlog_errno(PTR_ERR(dlm->dlm_thread_task));
	439	dlm->dlm_thread_task = NULL;
	440	return -EINVAL;
	441	}
	442
	443	return 0;
	444	}
	445
	446	void dlm_complete_thread(struct dlm_ctxt *dlm)
	447	{
	448	if (dlm->dlm_thread_task) {
449	mlog(ML_KTHREAD, "waiting for dlm thread to exit\n");
450	kthread_stop(dlm->dlm_thread_task);
451	dlm->dlm_thread_task = NULL;
452	}
453	}
454
455	static int dlm_dirty_list_empty(struct dlm_ctxt *dlm)
456	{
457	int empty;
458
459	spin_lock(&dlm->spinlock);
460	empty = list_empty(&dlm->dirty_list);
461	spin_unlock(&dlm->spinlock);
462
463	return empty;
464	}
465
466	static void dlm_flush_asts(struct dlm_ctxt *dlm)
467	{
468	int ret;
469	struct dlm_lock *lock;
470	struct dlm_lock_resource *res;
471	u8 hi;
472
473	spin_lock(&dlm->ast_lock);
474	while (!list_empty(&dlm->pending_asts)) {
475	lock = list_entry(dlm->pending_asts.next,
476	struct dlm_lock, ast_list);
477	/* get an extra ref on lock */
478	dlm_lock_get(lock);
479	res = lock->lockres;
480	mlog(0, "delivering an ast for this lockres\n");
481
482	BUG_ON(!lock->ast_pending);
483
484	/* remove from list (including ref) */
485	list_del_init(&lock->ast_list);
486	dlm_lock_put(lock);
487	spin_unlock(&dlm->ast_lock);
488
489	if (lock->ml.node != dlm->node_num) {
490	ret = dlm_do_remote_ast(dlm, res, lock);
491	if (ret < 0)
492	mlog_errno(ret);
493	} else
494	dlm_do_local_ast(dlm, res, lock);
495
496	spin_lock(&dlm->ast_lock);
497
498	/* possible that another ast was queued while
499	* we were delivering the last one */
500	if (!list_empty(&lock->ast_list)) {
501	mlog(0, "aha another ast got queued while "
502	"we were finishing the last one. will "
503	"keep the ast_pending flag set.\n");
504	} else
505	lock->ast_pending = 0;
506
507	/* drop the extra ref.
508	* this may drop it completely. */
509	dlm_lock_put(lock);
510	dlm_lockres_release_ast(dlm, res);
511	}
512
513	while (!list_empty(&dlm->pending_basts)) {
514	lock = list_entry(dlm->pending_basts.next,
515	struct dlm_lock, bast_list);
516	/* get an extra ref on lock */
517	dlm_lock_get(lock);
518	res = lock->lockres;
519
520	BUG_ON(!lock->bast_pending);
521
522	/* get the highest blocked lock, and reset */
523	spin_lock(&lock->spinlock);
524	BUG_ON(lock->ml.highest_blocked <= LKM_IVMODE);
525	hi = lock->ml.highest_blocked;
526	lock->ml.highest_blocked = LKM_IVMODE;
527	spin_unlock(&lock->spinlock);
528
529	/* remove from list (including ref) */
530	list_del_init(&lock->bast_list);
531	dlm_lock_put(lock);
532	spin_unlock(&dlm->ast_lock);
533
534	mlog(0, "delivering a bast for this lockres "
535	"(blocked = %d\n", hi);
536
537	if (lock->ml.node != dlm->node_num) {
538	ret = dlm_send_proxy_bast(dlm, res, lock, hi);
539	if (ret < 0)
540	mlog_errno(ret);
541	} else
542	dlm_do_local_bast(dlm, res, lock, hi);
543
544	spin_lock(&dlm->ast_lock);
545
546	/* possible that another bast was queued while
547	* we were delivering the last one */
548	if (!list_empty(&lock->bast_list)) {
549	mlog(0, "aha another bast got queued while "
550	"we were finishing the last one. will "
551	"keep the bast_pending flag set.\n");
552	} else
553	lock->bast_pending = 0;
554
555	/* drop the extra ref.
556	* this may drop it completely. */
557	dlm_lock_put(lock);
558	dlm_lockres_release_ast(dlm, res);
559	}
560	wake_up(&dlm->ast_wq);
561	spin_unlock(&dlm->ast_lock);
562	}
563
564
565	#define DLM_THREAD_TIMEOUT_MS (4 * 1000)
566	#define DLM_THREAD_MAX_DIRTY 100
567	#define DLM_THREAD_MAX_ASTS 10
568
569	static int dlm_thread(void *data)
570	{
571	struct dlm_lock_resource *res;
572	struct dlm_ctxt *dlm = data;
573	unsigned long timeout = msecs_to_jiffies(DLM_THREAD_TIMEOUT_MS);
574
575	mlog(0, "dlm thread running for %s...\n", dlm->name);
576
577	while (!kthread_should_stop()) {
578	int n = DLM_THREAD_MAX_DIRTY;
579
580	/* dlm_shutting_down is very point-in-time, but that
581	* doesn't matter as we'll just loop back around if we
582	* get false on the leading edge of a state
583	* transition. */
584	dlm_run_purge_list(dlm, dlm_shutting_down(dlm));
585
586	/* We really don't want to hold dlm->spinlock while
587	* calling dlm_shuffle_lists on each lockres that
588	* needs to have its queues adjusted and AST/BASTs
589	* run. So let's pull each entry off the dirty_list
590	* and drop dlm->spinlock ASAP. Once off the list,
591	* res->spinlock needs to be taken again to protect
592	* the queues while calling dlm_shuffle_lists. */
593	spin_lock(&dlm->spinlock);
594	while (!list_empty(&dlm->dirty_list)) {
595	int delay = 0;
596	res = list_entry(dlm->dirty_list.next,
597	struct dlm_lock_resource, dirty);
598
599	/* peel a lockres off, remove it from the list,
600	* unset the dirty flag and drop the dlm lock */
601	BUG_ON(!res);
602	dlm_lockres_get(res);
603
604	spin_lock(&res->spinlock);
605	res->state &= ~DLM_LOCK_RES_DIRTY;
606	list_del_init(&res->dirty);
607	spin_unlock(&res->spinlock);
608	spin_unlock(&dlm->spinlock);
609
610	/* lockres can be re-dirtied/re-added to the
611	* dirty_list in this gap, but that is ok */
612
613	spin_lock(&res->spinlock);
614	if (res->owner != dlm->node_num) {
615	__dlm_print_one_lock_resource(res);
616	mlog(ML_ERROR, "inprog:%s, mig:%s, reco:%s, dirty:%s\n",
617	res->state & DLM_LOCK_RES_IN_PROGRESS ? "yes" : "no",
618	res->state & DLM_LOCK_RES_MIGRATING ? "yes" : "no",
619	res->state & DLM_LOCK_RES_RECOVERING ? "yes" : "no",
620	res->state & DLM_LOCK_RES_DIRTY ? "yes" : "no");
621	}
622	BUG_ON(res->owner != dlm->node_num);
623
624	/* it is now ok to move lockreses in these states
625	* to the dirty list, assuming that they will only be
626	* dirty for a short while. */
627	if (res->state & (DLM_LOCK_RES_IN_PROGRESS \|
628	DLM_LOCK_RES_MIGRATING \|
629	DLM_LOCK_RES_RECOVERING)) {
630	/* move it to the tail and keep going */
631	spin_unlock(&res->spinlock);
632	mlog(0, "delaying list shuffling for in-"
633	"progress lockres %.*s, state=%d\n",
634	res->lockname.len, res->lockname.name,
635	res->state);
636	delay = 1;
637	goto in_progress;
638	}
639
640	/* at this point the lockres is not migrating/
641	* recovering/in-progress. we have the lockres
642	* spinlock and do NOT have the dlm lock.
643	* safe to reserve/queue asts and run the lists. */
644
8d79d088 KH	645	mlog(0, "calling dlm_shuffle_lists with dlm=%s, "
	646	"res=%.*s\n", dlm->name,
	647	res->lockname.len, res->lockname.name);
6714d8e8 KH	648
	649	/* called while holding lockres lock */
	650	dlm_shuffle_lists(dlm, res);
	651	spin_unlock(&res->spinlock);
	652
	653	dlm_lockres_calc_usage(dlm, res);
	654
	655	in_progress:
	656
	657	spin_lock(&dlm->spinlock);
	658	/* if the lock was in-progress, stick
	659	* it on the back of the list */
	660	if (delay) {
	661	spin_lock(&res->spinlock);
	662	list_add_tail(&res->dirty, &dlm->dirty_list);
	663	res->state \|= DLM_LOCK_RES_DIRTY;
	664	spin_unlock(&res->spinlock);
	665	}
	666	dlm_lockres_put(res);
	667
	668	/* unlikely, but we may need to give time to
	669	* other tasks */
	670	if (!--n) {
	671	mlog(0, "throttling dlm_thread\n");
	672	break;
	673	}
	674	}
	675
	676	spin_unlock(&dlm->spinlock);
	677	dlm_flush_asts(dlm);
	678
	679	/* yield and continue right away if there is more work to do */
	680	if (!n) {
	681	yield();
	682	continue;
	683	}
	684
	685	wait_event_interruptible_timeout(dlm->dlm_thread_wq,
	686	!dlm_dirty_list_empty(dlm) \|\|
	687	kthread_should_stop(),
	688	timeout);
	689	}
	690
	691	mlog(0, "quitting DLM thread\n");
	692	return 0;
	693	}