[deliverable/linux.git] / fs / orangefs / waitqueue.c

/*
 * (C) 2001 Clemson University and The University of Chicago
 * (C) 2011 Omnibond Systems
 *
 * Changes by Acxiom Corporation to implement generic service_operation()
 * function, Copyright Acxiom Corporation, 2005.
 *
 * See COPYING in top-level directory.
 */

/*
 *  In-kernel waitqueue operations.
 */

#include "protocol.h"
#include "orangefs-kernel.h"
#include "orangefs-bufmap.h"

/*
 * What we do in this function is to walk the list of operations that are
 * present in the request queue and mark them as purged.
 * NOTE: This is called from the device close after client-core has
 * guaranteed that no new operations could appear on the list since the
 * client-core is anyway going to exit.
 */
void purge_waiting_ops(void)
{
	struct orangefs_kernel_op_s *op;

	spin_lock(&orangefs_request_list_lock);
	list_for_each_entry(op, &orangefs_request_list, list) {
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "pvfs2-client-core: purging op tag %llu %s\n",
			     llu(op->tag),
			     get_opname_string(op));
		spin_lock(&op->lock);
		set_op_state_purged(op);
		spin_unlock(&op->lock);
	}
	spin_unlock(&orangefs_request_list_lock);
}

static inline void
add_op_to_request_list(struct orangefs_kernel_op_s *op)
{
	spin_lock(&orangefs_request_list_lock);
	spin_lock(&op->lock);
	set_op_state_waiting(op);
	list_add_tail(&op->list, &orangefs_request_list);
	spin_unlock(&orangefs_request_list_lock);
	spin_unlock(&op->lock);
	wake_up_interruptible(&orangefs_request_list_waitq);
}

static inline
void add_priority_op_to_request_list(struct orangefs_kernel_op_s *op)
{
	spin_lock(&orangefs_request_list_lock);
	spin_lock(&op->lock);
	set_op_state_waiting(op);

	list_add(&op->list, &orangefs_request_list);
	spin_unlock(&orangefs_request_list_lock);
	spin_unlock(&op->lock);
	wake_up_interruptible(&orangefs_request_list_waitq);
}

/*
 * submits a ORANGEFS operation and waits for it to complete
 *
 * Note op->downcall.status will contain the status of the operation (in
 * errno format), whether provided by pvfs2-client or a result of failure to
 * service the operation.  If the caller wishes to distinguish, then
 * op->state can be checked to see if it was serviced or not.
 *
 * Returns contents of op->downcall.status for convenience
 */
int service_operation(struct orangefs_kernel_op_s *op,
		      const char *op_name,
		      int flags)
{
	/* flags to modify behavior */
	sigset_t orig_sigset;
	int ret = 0;

	/* irqflags and wait_entry are only used IF the client-core aborts */
	unsigned long irqflags;

	DEFINE_WAIT(wait_entry);

	op->upcall.tgid = current->tgid;
	op->upcall.pid = current->pid;

retry_servicing:
	op->downcall.status = 0;
	gossip_debug(GOSSIP_WAIT_DEBUG,
		     "orangefs: service_operation: %s %p\n",
		     op_name,
		     op);
	gossip_debug(GOSSIP_WAIT_DEBUG,
		     "orangefs: operation posted by process: %s, pid: %i\n",
		     current->comm,
		     current->pid);

	/* mask out signals if this operation is not to be interrupted */
	if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
		orangefs_block_signals(&orig_sigset);

	if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) {
		ret = mutex_lock_interruptible(&request_mutex);
		/*
		 * check to see if we were interrupted while waiting for
		 * semaphore
		 */
		if (ret < 0) {
			if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
				orangefs_set_signals(&orig_sigset);
			op->downcall.status = ret;
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "orangefs: service_operation interrupted.\n");
			return ret;
		}
	}

	gossip_debug(GOSSIP_WAIT_DEBUG,
		     "%s:About to call is_daemon_in_service().\n",
		     __func__);

	if (is_daemon_in_service() < 0) {
		/*
		 * By incrementing the per-operation attempt counter, we
		 * directly go into the timeout logic while waiting for
		 * the matching downcall to be read
		 */
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "%s:client core is NOT in service(%d).\n",
			     __func__,
			     is_daemon_in_service());
		op->attempts++;
	}

	/* queue up the operation */
	if (flags & ORANGEFS_OP_PRIORITY) {
		add_priority_op_to_request_list(op);
	} else {
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "%s:About to call add_op_to_request_list().\n",
			     __func__);
		add_op_to_request_list(op);
	}

	if (!(flags & ORANGEFS_OP_NO_SEMAPHORE))
		mutex_unlock(&request_mutex);

	/*
	 * If we are asked to service an asynchronous operation from
	 * VFS perspective, we are done.
	 */
	if (flags & ORANGEFS_OP_ASYNC)
		return 0;

	if (flags & ORANGEFS_OP_CANCELLATION) {
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "%s:"
			     "About to call wait_for_cancellation_downcall.\n",
			     __func__);
		ret = wait_for_cancellation_downcall(op);
	} else {
		ret = wait_for_matching_downcall(op);
	}

	if (ret < 0) {
		/* failed to get matching downcall */
		if (ret == -ETIMEDOUT) {
			gossip_err("orangefs: %s -- wait timed out; aborting attempt.\n",
				   op_name);
		}
		op->downcall.status = ret;
	} else {
		/* got matching downcall; make sure status is in errno format */
		op->downcall.status =
		    orangefs_normalize_to_errno(op->downcall.status);
		ret = op->downcall.status;
	}

	if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
		orangefs_set_signals(&orig_sigset);

	BUG_ON(ret != op->downcall.status);
	/* retry if operation has not been serviced and if requested */
	if (!op_state_serviced(op) && op->downcall.status == -EAGAIN) {
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "orangefs: tag %llu (%s)"
			     " -- operation to be retried (%d attempt)\n",
			     llu(op->tag),
			     op_name,
			     op->attempts + 1);

		if (!op->uses_shared_memory)
			/*
			 * this operation doesn't use the shared memory
			 * system
			 */
			goto retry_servicing;

		/* op uses shared memory */
		if (orangefs_get_bufmap_init() == 0) {
			/*
			 * This operation uses the shared memory system AND
			 * the system is not yet ready. This situation occurs
			 * when the client-core is restarted AND there were
			 * operations waiting to be processed or were already
			 * in process.
			 */
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "uses_shared_memory is true.\n");
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "Client core in-service status(%d).\n",
				     is_daemon_in_service());
			gossip_debug(GOSSIP_WAIT_DEBUG, "bufmap_init:%d.\n",
				     orangefs_get_bufmap_init());
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "operation's status is 0x%0x.\n",
				     op->op_state);

			/*
			 * let process sleep for a few seconds so shared
			 * memory system can be initialized.
			 */
			spin_lock_irqsave(&op->lock, irqflags);
			prepare_to_wait(&orangefs_bufmap_init_waitq,
					&wait_entry,
					TASK_INTERRUPTIBLE);
			spin_unlock_irqrestore(&op->lock, irqflags);

			/*
			 * Wait for orangefs_bufmap_initialize() to wake me up
			 * within the allotted time.
			 */
			ret = schedule_timeout(MSECS_TO_JIFFIES
				(1000 * ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS));

			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "Value returned from schedule_timeout:"
				     "%d.\n",
				     ret);
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "Is shared memory available? (%d).\n",
				     orangefs_get_bufmap_init());

			spin_lock_irqsave(&op->lock, irqflags);
			finish_wait(&orangefs_bufmap_init_waitq, &wait_entry);
			spin_unlock_irqrestore(&op->lock, irqflags);

			if (orangefs_get_bufmap_init() == 0) {
				gossip_err("%s:The shared memory system has not started in %d seconds after the client core restarted.  Aborting user's request(%s).\n",
					   __func__,
					   ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS,
					   get_opname_string(op));
				return -EIO;
			}

			/*
			 * Return to the calling function and re-populate a
			 * shared memory buffer.
			 */
			return -EAGAIN;
		}
	}

	gossip_debug(GOSSIP_WAIT_DEBUG,
		     "orangefs: service_operation %s returning: %d for %p.\n",
		     op_name,
		     ret,
		     op);
	return ret;
}

static inline void remove_op_from_request_list(struct orangefs_kernel_op_s *op)
{
	struct list_head *tmp = NULL;
	struct list_head *tmp_safe = NULL;
	struct orangefs_kernel_op_s *tmp_op = NULL;

	spin_lock(&orangefs_request_list_lock);
	list_for_each_safe(tmp, tmp_safe, &orangefs_request_list) {
		tmp_op = list_entry(tmp,
				    struct orangefs_kernel_op_s,
				    list);
		if (tmp_op && (tmp_op == op)) {
			list_del(&tmp_op->list);
			break;
		}
	}
	spin_unlock(&orangefs_request_list_lock);
}

static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
{
	/*
	 * handle interrupted cases depending on what state we were in when
	 * the interruption is detected.  there is a coarse grained lock
	 * across the operation.
	 *
	 * NOTE: be sure not to reverse lock ordering by locking an op lock
	 * while holding the request_list lock.  Here, we first lock the op
	 * and then lock the appropriate list.
	 */
	if (!op) {
		gossip_debug(GOSSIP_WAIT_DEBUG,
			    "%s: op is null, ignoring\n",
			     __func__);
		return;
	}

	/*
	 * one more sanity check, make sure it's in one of the possible states
	 * or don't try to cancel it
	 */
	if (!(op_state_waiting(op) ||
	      op_state_in_progress(op) ||
	      op_state_serviced(op) ||
	      op_state_purged(op))) {
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "%s: op %p not in a valid state (%0x), "
			     "ignoring\n",
			     __func__,
			     op,
			     op->op_state);
		return;
	}

	spin_lock(&op->lock);

	if (op_state_waiting(op)) {
		/*
		 * upcall hasn't been read; remove op from upcall request
		 * list.
		 */
		spin_unlock(&op->lock);
		remove_op_from_request_list(op);
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "Interrupted: Removed op %p from request_list\n",
			     op);
	} else if (op_state_in_progress(op)) {
		/* op must be removed from the in progress htable */
		spin_unlock(&op->lock);
		spin_lock(&htable_ops_in_progress_lock);
		list_del(&op->list);
		spin_unlock(&htable_ops_in_progress_lock);
		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "Interrupted: Removed op %p"
			     " from htable_ops_in_progress\n",
			     op);
	} else if (!op_state_serviced(op)) {
		spin_unlock(&op->lock);
		gossip_err("interrupted operation is in a weird state 0x%x\n",
			   op->op_state);
	} else {
		/*
		 * It is not intended for execution to flow here,
		 * but having this unlock here makes sparse happy.
		 */
		gossip_err("%s: can't get here.\n", __func__);
		spin_unlock(&op->lock);
	}
}

/*
 * sleeps on waitqueue waiting for matching downcall.
 * if client-core finishes servicing, then we are good to go.
 * else if client-core exits, we get woken up here, and retry with a timeout
 *
 * Post when this call returns to the caller, the specified op will no
 * longer be on any list or htable.
 *
 * Returns 0 on success and -errno on failure
 * Errors are:
 * EAGAIN in case we want the caller to requeue and try again..
 * EINTR/EIO/ETIMEDOUT indicating we are done trying to service this
 * operation since client-core seems to be exiting too often
 * or if we were interrupted.
 */
static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op)
{
	int ret = -EINVAL;
	DEFINE_WAIT(wait_entry);

	while (1) {
		spin_lock(&op->lock);
		prepare_to_wait(&op->waitq, &wait_entry, TASK_INTERRUPTIBLE);
		if (op_state_serviced(op)) {
			spin_unlock(&op->lock);
			ret = 0;
			break;
		}
		spin_unlock(&op->lock);

		if (!signal_pending(current)) {
			/*
			 * if this was our first attempt and client-core
			 * has not purged our operation, we are happy to
			 * simply wait
			 */
			spin_lock(&op->lock);
			if (op->attempts == 0 && !op_state_purged(op)) {
				spin_unlock(&op->lock);
				schedule();
			} else {
				spin_unlock(&op->lock);
				/*
				 * subsequent attempts, we retry exactly once
				 * with timeouts
				 */
				if (!schedule_timeout(MSECS_TO_JIFFIES
				      (1000 * op_timeout_secs))) {
					gossip_debug(GOSSIP_WAIT_DEBUG,
						     "*** %s:"
						     " operation timed out (tag"
						     " %llu, %p, att %d)\n",
						     __func__,
						     llu(op->tag),
						     op,
						     op->attempts);
					ret = -ETIMEDOUT;
					orangefs_clean_up_interrupted_operation
					    (op);
					break;
				}
			}
			spin_lock(&op->lock);
			op->attempts++;
			/*
			 * if the operation was purged in the meantime, it
			 * is better to requeue it afresh but ensure that
			 * we have not been purged repeatedly. This could
			 * happen if client-core crashes when an op
			 * is being serviced, so we requeue the op, client
			 * core crashes again so we requeue the op, client
			 * core starts, and so on...
			 */
			if (op_state_purged(op)) {
				ret = (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ?
					 -EAGAIN :
					 -EIO;
				spin_unlock(&op->lock);
				gossip_debug(GOSSIP_WAIT_DEBUG,
					     "*** %s:"
					     " operation purged (tag "
					     "%llu, %p, att %d)\n",
					     __func__,
					     llu(op->tag),
					     op,
					     op->attempts);
				orangefs_clean_up_interrupted_operation(op);
				break;
			}
			spin_unlock(&op->lock);
			continue;
		}

		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "*** %s:"
			     " operation interrupted by a signal (tag "
			     "%llu, op %p)\n",
			     __func__,
			     llu(op->tag),
			     op);
		orangefs_clean_up_interrupted_operation(op);
		ret = -EINTR;
		break;
	}

	spin_lock(&op->lock);
	finish_wait(&op->waitq, &wait_entry);
	spin_unlock(&op->lock);

	return ret;
}

/*
 * similar to wait_for_matching_downcall(), but used in the special case
 * of I/O cancellations.
 *
 * Note we need a special wait function because if this is called we already
 *      know that a signal is pending in current and need to service the
 *      cancellation upcall anyway.  the only way to exit this is to either
 *      timeout or have the cancellation be serviced properly.
 */
static int wait_for_cancellation_downcall(struct orangefs_kernel_op_s *op)
{
	int ret = -EINVAL;
	DEFINE_WAIT(wait_entry);

	while (1) {
		spin_lock(&op->lock);
		prepare_to_wait(&op->waitq, &wait_entry, TASK_INTERRUPTIBLE);
		if (op_state_serviced(op)) {
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "%s:op-state is SERVICED.\n",
				     __func__);
			spin_unlock(&op->lock);
			ret = 0;
			break;
		}
		spin_unlock(&op->lock);

		if (signal_pending(current)) {
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "%s:operation interrupted by a signal (tag"
				     " %llu, op %p)\n",
				     __func__,
				     llu(op->tag),
				     op);
			orangefs_clean_up_interrupted_operation(op);
			ret = -EINTR;
			break;
		}

		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "%s:About to call schedule_timeout.\n",
			     __func__);
		ret =
		    schedule_timeout(MSECS_TO_JIFFIES(1000 * op_timeout_secs));

		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "%s:Value returned from schedule_timeout(%d).\n",
			     __func__,
			     ret);
		if (!ret) {
			gossip_debug(GOSSIP_WAIT_DEBUG,
				     "%s:*** operation timed out: %p\n",
				     __func__,
				     op);
			orangefs_clean_up_interrupted_operation(op);
			ret = -ETIMEDOUT;
			break;
		}

		gossip_debug(GOSSIP_WAIT_DEBUG,
			     "%s:Breaking out of loop, regardless of value returned by schedule_timeout.\n",
			     __func__);
		ret = -ETIMEDOUT;
		break;
	}

	spin_lock(&op->lock);
	finish_wait(&op->waitq, &wait_entry);
	spin_unlock(&op->lock);

	gossip_debug(GOSSIP_WAIT_DEBUG,
		     "%s:returning ret(%d)\n",
		     __func__,
		     ret);

	return ret;
}
Commit	Line	Data
1182fca3 MM	1	/*
	2	* (C) 2001 Clemson University and The University of Chicago
	3	* (C) 2011 Omnibond Systems
	4	*
	5	* Changes by Acxiom Corporation to implement generic service_operation()
	6	* function, Copyright Acxiom Corporation, 2005.
	7	*
	8	* See COPYING in top-level directory.
	9	*/
	10
	11	/*
	12	* In-kernel waitqueue operations.
	13	*/
	14
	15	#include "protocol.h"
575e9461 MM	16	#include "orangefs-kernel.h"
575e9461 MM	17	#include "orangefs-bufmap.h"
1182fca3 MM	18
	19	/*
	20	* What we do in this function is to walk the list of operations that are
	21	* present in the request queue and mark them as purged.
	22	* NOTE: This is called from the device close after client-core has
	23	* guaranteed that no new operations could appear on the list since the
	24	* client-core is anyway going to exit.
	25	*/
	26	void purge_waiting_ops(void)
	27	{
8bb8aefd	28	struct orangefs_kernel_op_s *op;
1182fca3	29
8bb8aefd YL	30	spin_lock(&orangefs_request_list_lock);
8bb8aefd YL	31	list_for_each_entry(op, &orangefs_request_list, list) {
1182fca3 MM	32	gossip_debug(GOSSIP_WAIT_DEBUG,
	33	"pvfs2-client-core: purging op tag %llu %s\n",
	34	llu(op->tag),
	35	get_opname_string(op));
	36	spin_lock(&op->lock);
	37	set_op_state_purged(op);
	38	spin_unlock(&op->lock);
1182fca3	39	}
8bb8aefd	40	spin_unlock(&orangefs_request_list_lock);
1182fca3 MM	41	}
1182fca3 MM	42
fc916da5 AV	43	static inline void
	44	add_op_to_request_list(struct orangefs_kernel_op_s *op)
	45	{
	46	spin_lock(&orangefs_request_list_lock);
	47	spin_lock(&op->lock);
	48	set_op_state_waiting(op);
	49	list_add_tail(&op->list, &orangefs_request_list);
	50	spin_unlock(&orangefs_request_list_lock);
	51	spin_unlock(&op->lock);
	52	wake_up_interruptible(&orangefs_request_list_waitq);
	53	}
	54
	55	static inline
	56	void add_priority_op_to_request_list(struct orangefs_kernel_op_s *op)
	57	{
	58	spin_lock(&orangefs_request_list_lock);
	59	spin_lock(&op->lock);
	60	set_op_state_waiting(op);
	61
	62	list_add(&op->list, &orangefs_request_list);
	63	spin_unlock(&orangefs_request_list_lock);
	64	spin_unlock(&op->lock);
	65	wake_up_interruptible(&orangefs_request_list_waitq);
	66	}
	67
1182fca3	68	/*
8bb8aefd	69	* submits a ORANGEFS operation and waits for it to complete
1182fca3 MM	70	*
	71	* Note op->downcall.status will contain the status of the operation (in
	72	* errno format), whether provided by pvfs2-client or a result of failure to
	73	* service the operation. If the caller wishes to distinguish, then
	74	* op->state can be checked to see if it was serviced or not.
	75	*
	76	* Returns contents of op->downcall.status for convenience
	77	*/
8bb8aefd	78	int service_operation(struct orangefs_kernel_op_s *op,
1182fca3 MM	79	const char *op_name,
	80	int flags)
	81	{
	82	/* flags to modify behavior */
	83	sigset_t orig_sigset;
	84	int ret = 0;
	85
	86	/* irqflags and wait_entry are only used IF the client-core aborts */
	87	unsigned long irqflags;
	88
ce6c414e	89	DEFINE_WAIT(wait_entry);
1182fca3 MM	90
	91	op->upcall.tgid = current->tgid;
	92	op->upcall.pid = current->pid;
	93
	94	retry_servicing:
	95	op->downcall.status = 0;
	96	gossip_debug(GOSSIP_WAIT_DEBUG,
8bb8aefd	97	"orangefs: service_operation: %s %p\n",
1182fca3 MM	98	op_name,
	99	op);
	100	gossip_debug(GOSSIP_WAIT_DEBUG,
8bb8aefd	101	"orangefs: operation posted by process: %s, pid: %i\n",
1182fca3 MM	102	current->comm,
	103	current->pid);
	104
	105	/* mask out signals if this operation is not to be interrupted */
8bb8aefd	106	if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
c146c0b8	107	orangefs_block_signals(&orig_sigset);
1182fca3	108
8bb8aefd	109	if (!(flags & ORANGEFS_OP_NO_SEMAPHORE)) {
1182fca3 MM	110	ret = mutex_lock_interruptible(&request_mutex);
	111	/*
	112	* check to see if we were interrupted while waiting for
	113	* semaphore
	114	*/
	115	if (ret < 0) {
8bb8aefd	116	if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
c146c0b8	117	orangefs_set_signals(&orig_sigset);
1182fca3 MM	118	op->downcall.status = ret;
1182fca3 MM	119	gossip_debug(GOSSIP_WAIT_DEBUG,
8bb8aefd	120	"orangefs: service_operation interrupted.\n");
1182fca3 MM	121	return ret;
	122	}
	123	}
	124
	125	gossip_debug(GOSSIP_WAIT_DEBUG,
	126	"%s:About to call is_daemon_in_service().\n",
	127	__func__);
	128
	129	if (is_daemon_in_service() < 0) {
	130	/*
	131	* By incrementing the per-operation attempt counter, we
	132	* directly go into the timeout logic while waiting for
	133	* the matching downcall to be read
	134	*/
	135	gossip_debug(GOSSIP_WAIT_DEBUG,
	136	"%s:client core is NOT in service(%d).\n",
	137	__func__,
	138	is_daemon_in_service());
	139	op->attempts++;
	140	}
	141
	142	/* queue up the operation */
8bb8aefd	143	if (flags & ORANGEFS_OP_PRIORITY) {
1182fca3 MM	144	add_priority_op_to_request_list(op);
	145	} else {
	146	gossip_debug(GOSSIP_WAIT_DEBUG,
	147	"%s:About to call add_op_to_request_list().\n",
	148	__func__);
	149	add_op_to_request_list(op);
	150	}
	151
8bb8aefd	152	if (!(flags & ORANGEFS_OP_NO_SEMAPHORE))
1182fca3 MM	153	mutex_unlock(&request_mutex);
	154
	155	/*
	156	* If we are asked to service an asynchronous operation from
	157	* VFS perspective, we are done.
	158	*/
8bb8aefd	159	if (flags & ORANGEFS_OP_ASYNC)
1182fca3 MM	160	return 0;
1182fca3 MM	161
8bb8aefd	162	if (flags & ORANGEFS_OP_CANCELLATION) {
1182fca3 MM	163	gossip_debug(GOSSIP_WAIT_DEBUG,
	164	"%s:"
	165	"About to call wait_for_cancellation_downcall.\n",
	166	__func__);
	167	ret = wait_for_cancellation_downcall(op);
	168	} else {
	169	ret = wait_for_matching_downcall(op);
	170	}
	171
	172	if (ret < 0) {
	173	/* failed to get matching downcall */
	174	if (ret == -ETIMEDOUT) {
8bb8aefd	175	gossip_err("orangefs: %s -- wait timed out; aborting attempt.\n",
1182fca3 MM	176	op_name);
	177	}
	178	op->downcall.status = ret;
	179	} else {
	180	/* got matching downcall; make sure status is in errno format */
	181	op->downcall.status =
8bb8aefd	182	orangefs_normalize_to_errno(op->downcall.status);
1182fca3 MM	183	ret = op->downcall.status;
	184	}
	185
8bb8aefd	186	if (!(flags & ORANGEFS_OP_INTERRUPTIBLE))
c146c0b8	187	orangefs_set_signals(&orig_sigset);
1182fca3 MM	188
	189	BUG_ON(ret != op->downcall.status);
	190	/* retry if operation has not been serviced and if requested */
	191	if (!op_state_serviced(op) && op->downcall.status == -EAGAIN) {
	192	gossip_debug(GOSSIP_WAIT_DEBUG,
8bb8aefd	193	"orangefs: tag %llu (%s)"
1182fca3 MM	194	" -- operation to be retried (%d attempt)\n",
	195	llu(op->tag),
	196	op_name,
	197	op->attempts + 1);
	198
	199	if (!op->uses_shared_memory)
	200	/*
	201	* this operation doesn't use the shared memory
	202	* system
	203	*/
	204	goto retry_servicing;
	205
	206	/* op uses shared memory */
7d221485	207	if (orangefs_get_bufmap_init() == 0) {
1182fca3 MM	208	/*
	209	* This operation uses the shared memory system AND
	210	* the system is not yet ready. This situation occurs
	211	* when the client-core is restarted AND there were
	212	* operations waiting to be processed or were already
	213	* in process.
	214	*/
	215	gossip_debug(GOSSIP_WAIT_DEBUG,
	216	"uses_shared_memory is true.\n");
	217	gossip_debug(GOSSIP_WAIT_DEBUG,
	218	"Client core in-service status(%d).\n",
	219	is_daemon_in_service());
	220	gossip_debug(GOSSIP_WAIT_DEBUG, "bufmap_init:%d.\n",
7d221485	221	orangefs_get_bufmap_init());
1182fca3 MM	222	gossip_debug(GOSSIP_WAIT_DEBUG,
	223	"operation's status is 0x%0x.\n",
	224	op->op_state);
	225
	226	/*
	227	* let process sleep for a few seconds so shared
	228	* memory system can be initialized.
	229	*/
	230	spin_lock_irqsave(&op->lock, irqflags);
ce6c414e MM	231	prepare_to_wait(&orangefs_bufmap_init_waitq,
	232	&wait_entry,
	233	TASK_INTERRUPTIBLE);
1182fca3 MM	234	spin_unlock_irqrestore(&op->lock, irqflags);
1182fca3 MM	235
1182fca3	236	/*
8bb8aefd	237	* Wait for orangefs_bufmap_initialize() to wake me up
1182fca3 MM	238	* within the allotted time.
	239	*/
	240	ret = schedule_timeout(MSECS_TO_JIFFIES
8bb8aefd	241	(1000 * ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS));
1182fca3 MM	242
	243	gossip_debug(GOSSIP_WAIT_DEBUG,
	244	"Value returned from schedule_timeout:"
	245	"%d.\n",
	246	ret);
	247	gossip_debug(GOSSIP_WAIT_DEBUG,
	248	"Is shared memory available? (%d).\n",
7d221485	249	orangefs_get_bufmap_init());
1182fca3 MM	250
1182fca3 MM	251	spin_lock_irqsave(&op->lock, irqflags);
ce6c414e	252	finish_wait(&orangefs_bufmap_init_waitq, &wait_entry);
1182fca3 MM	253	spin_unlock_irqrestore(&op->lock, irqflags);
1182fca3 MM	254
7d221485	255	if (orangefs_get_bufmap_init() == 0) {
1182fca3 MM	256	gossip_err("%s:The shared memory system has not started in %d seconds after the client core restarted. Aborting user's request(%s).\n",
1182fca3 MM	257	__func__,
8bb8aefd	258	ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS,
1182fca3 MM	259	get_opname_string(op));
	260	return -EIO;
	261	}
	262
	263	/*
	264	* Return to the calling function and re-populate a
	265	* shared memory buffer.
	266	*/
	267	return -EAGAIN;
	268	}
	269	}
	270
	271	gossip_debug(GOSSIP_WAIT_DEBUG,
8bb8aefd	272	"orangefs: service_operation %s returning: %d for %p.\n",
1182fca3 MM	273	op_name,
	274	ret,
	275	op);
	276	return ret;
	277	}
	278
fc916da5 AV	279	static inline void remove_op_from_request_list(struct orangefs_kernel_op_s *op)
	280	{
	281	struct list_head *tmp = NULL;
	282	struct list_head *tmp_safe = NULL;
	283	struct orangefs_kernel_op_s *tmp_op = NULL;
	284
	285	spin_lock(&orangefs_request_list_lock);
	286	list_for_each_safe(tmp, tmp_safe, &orangefs_request_list) {
	287	tmp_op = list_entry(tmp,
	288	struct orangefs_kernel_op_s,
	289	list);
	290	if (tmp_op && (tmp_op == op)) {
	291	list_del(&tmp_op->list);
	292	break;
	293	}
	294	}
	295	spin_unlock(&orangefs_request_list_lock);
	296	}
	297
e07db0a2	298	static void orangefs_clean_up_interrupted_operation(struct orangefs_kernel_op_s *op)
1182fca3 MM	299	{
	300	/*
	301	* handle interrupted cases depending on what state we were in when
	302	* the interruption is detected. there is a coarse grained lock
	303	* across the operation.
	304	*
	305	* NOTE: be sure not to reverse lock ordering by locking an op lock
	306	* while holding the request_list lock. Here, we first lock the op
	307	* and then lock the appropriate list.
	308	*/
	309	if (!op) {
	310	gossip_debug(GOSSIP_WAIT_DEBUG,
	311	"%s: op is null, ignoring\n",
	312	__func__);
	313	return;
	314	}
	315
	316	/*
	317	* one more sanity check, make sure it's in one of the possible states
	318	* or don't try to cancel it
	319	*/
	320	if (!(op_state_waiting(op) \|\|
	321	op_state_in_progress(op) \|\|
	322	op_state_serviced(op) \|\|
	323	op_state_purged(op))) {
	324	gossip_debug(GOSSIP_WAIT_DEBUG,
	325	"%s: op %p not in a valid state (%0x), "
	326	"ignoring\n",
	327	__func__,
	328	op,
	329	op->op_state);
	330	return;
	331	}
	332
	333	spin_lock(&op->lock);
	334
	335	if (op_state_waiting(op)) {
	336	/*
	337	* upcall hasn't been read; remove op from upcall request
	338	* list.
	339	*/
	340	spin_unlock(&op->lock);
	341	remove_op_from_request_list(op);
	342	gossip_debug(GOSSIP_WAIT_DEBUG,
	343	"Interrupted: Removed op %p from request_list\n",
	344	op);
	345	} else if (op_state_in_progress(op)) {
	346	/* op must be removed from the in progress htable */
	347	spin_unlock(&op->lock);
	348	spin_lock(&htable_ops_in_progress_lock);
	349	list_del(&op->list);
	350	spin_unlock(&htable_ops_in_progress_lock);
	351	gossip_debug(GOSSIP_WAIT_DEBUG,
	352	"Interrupted: Removed op %p"
	353	" from htable_ops_in_progress\n",
	354	op);
	355	} else if (!op_state_serviced(op)) {
	356	spin_unlock(&op->lock);
	357	gossip_err("interrupted operation is in a weird state 0x%x\n",
	358	op->op_state);
84d02150 MM	359	} else {
	360	/*
	361	* It is not intended for execution to flow here,
	362	* but having this unlock here makes sparse happy.
	363	*/
	364	gossip_err("%s: can't get here.\n", __func__);
	365	spin_unlock(&op->lock);
1182fca3 MM	366	}
	367	}
	368
	369	/*
	370	* sleeps on waitqueue waiting for matching downcall.
	371	* if client-core finishes servicing, then we are good to go.
	372	* else if client-core exits, we get woken up here, and retry with a timeout
	373	*
	374	* Post when this call returns to the caller, the specified op will no
	375	* longer be on any list or htable.
	376	*
	377	* Returns 0 on success and -errno on failure
	378	* Errors are:
	379	* EAGAIN in case we want the caller to requeue and try again..
	380	* EINTR/EIO/ETIMEDOUT indicating we are done trying to service this
	381	* operation since client-core seems to be exiting too often
	382	* or if we were interrupted.
	383	*/
b7ae37b0	384	static int wait_for_matching_downcall(struct orangefs_kernel_op_s *op)
1182fca3 MM	385	{
1182fca3 MM	386	int ret = -EINVAL;
ce6c414e	387	DEFINE_WAIT(wait_entry);
1182fca3 MM	388
1182fca3 MM	389	while (1) {
1182fca3	390	spin_lock(&op->lock);
ce6c414e	391	prepare_to_wait(&op->waitq, &wait_entry, TASK_INTERRUPTIBLE);
1182fca3 MM	392	if (op_state_serviced(op)) {
	393	spin_unlock(&op->lock);
	394	ret = 0;
	395	break;
	396	}
	397	spin_unlock(&op->lock);
	398
	399	if (!signal_pending(current)) {
	400	/*
	401	* if this was our first attempt and client-core
	402	* has not purged our operation, we are happy to
	403	* simply wait
	404	*/
	405	spin_lock(&op->lock);
	406	if (op->attempts == 0 && !op_state_purged(op)) {
	407	spin_unlock(&op->lock);
	408	schedule();
	409	} else {
	410	spin_unlock(&op->lock);
	411	/*
	412	* subsequent attempts, we retry exactly once
	413	* with timeouts
	414	*/
	415	if (!schedule_timeout(MSECS_TO_JIFFIES
	416	(1000 * op_timeout_secs))) {
	417	gossip_debug(GOSSIP_WAIT_DEBUG,
	418	"*** %s:"
	419	" operation timed out (tag"
	420	" %llu, %p, att %d)\n",
	421	__func__,
	422	llu(op->tag),
	423	op,
	424	op->attempts);
	425	ret = -ETIMEDOUT;
8bb8aefd	426	orangefs_clean_up_interrupted_operation
1182fca3 MM	427	(op);
	428	break;
	429	}
	430	}
	431	spin_lock(&op->lock);
	432	op->attempts++;
	433	/*
	434	* if the operation was purged in the meantime, it
	435	* is better to requeue it afresh but ensure that
	436	* we have not been purged repeatedly. This could
	437	* happen if client-core crashes when an op
	438	* is being serviced, so we requeue the op, client
	439	* core crashes again so we requeue the op, client
	440	* core starts, and so on...
	441	*/
	442	if (op_state_purged(op)) {
8bb8aefd	443	ret = (op->attempts < ORANGEFS_PURGE_RETRY_COUNT) ?
1182fca3 MM	444	-EAGAIN :
	445	-EIO;
	446	spin_unlock(&op->lock);
	447	gossip_debug(GOSSIP_WAIT_DEBUG,
	448	"*** %s:"
	449	" operation purged (tag "
	450	"%llu, %p, att %d)\n",
	451	__func__,
	452	llu(op->tag),
	453	op,
	454	op->attempts);
8bb8aefd	455	orangefs_clean_up_interrupted_operation(op);
1182fca3 MM	456	break;
	457	}
	458	spin_unlock(&op->lock);
	459	continue;
	460	}
	461
	462	gossip_debug(GOSSIP_WAIT_DEBUG,
	463	"*** %s:"
	464	" operation interrupted by a signal (tag "
	465	"%llu, op %p)\n",
	466	__func__,
	467	llu(op->tag),
	468	op);
8bb8aefd	469	orangefs_clean_up_interrupted_operation(op);
1182fca3 MM	470	ret = -EINTR;
	471	break;
	472	}
	473
1182fca3	474	spin_lock(&op->lock);
ce6c414e	475	finish_wait(&op->waitq, &wait_entry);
1182fca3 MM	476	spin_unlock(&op->lock);
	477
	478	return ret;
	479	}
	480
	481	/*
	482	* similar to wait_for_matching_downcall(), but used in the special case
	483	* of I/O cancellations.
	484	*
	485	* Note we need a special wait function because if this is called we already
	486	* know that a signal is pending in current and need to service the
	487	* cancellation upcall anyway. the only way to exit this is to either
	488	* timeout or have the cancellation be serviced properly.
	489	*/
b7ae37b0	490	static int wait_for_cancellation_downcall(struct orangefs_kernel_op_s *op)
1182fca3 MM	491	{
1182fca3 MM	492	int ret = -EINVAL;
ce6c414e	493	DEFINE_WAIT(wait_entry);
1182fca3 MM	494
1182fca3 MM	495	while (1) {
1182fca3	496	spin_lock(&op->lock);
ce6c414e	497	prepare_to_wait(&op->waitq, &wait_entry, TASK_INTERRUPTIBLE);
1182fca3 MM	498	if (op_state_serviced(op)) {
	499	gossip_debug(GOSSIP_WAIT_DEBUG,
	500	"%s:op-state is SERVICED.\n",
	501	__func__);
	502	spin_unlock(&op->lock);
	503	ret = 0;
	504	break;
	505	}
	506	spin_unlock(&op->lock);
	507
	508	if (signal_pending(current)) {
	509	gossip_debug(GOSSIP_WAIT_DEBUG,
	510	"%s:operation interrupted by a signal (tag"
	511	" %llu, op %p)\n",
	512	__func__,
	513	llu(op->tag),
	514	op);
8bb8aefd	515	orangefs_clean_up_interrupted_operation(op);
1182fca3 MM	516	ret = -EINTR;
	517	break;
	518	}
	519
	520	gossip_debug(GOSSIP_WAIT_DEBUG,
	521	"%s:About to call schedule_timeout.\n",
	522	__func__);
	523	ret =
	524	schedule_timeout(MSECS_TO_JIFFIES(1000 * op_timeout_secs));
	525
	526	gossip_debug(GOSSIP_WAIT_DEBUG,
	527	"%s:Value returned from schedule_timeout(%d).\n",
	528	__func__,
	529	ret);
	530	if (!ret) {
	531	gossip_debug(GOSSIP_WAIT_DEBUG,
	532	"%s:*** operation timed out: %p\n",
	533	__func__,
	534	op);
8bb8aefd	535	orangefs_clean_up_interrupted_operation(op);
1182fca3 MM	536	ret = -ETIMEDOUT;
	537	break;
	538	}
	539
	540	gossip_debug(GOSSIP_WAIT_DEBUG,
	541	"%s:Breaking out of loop, regardless of value returned by schedule_timeout.\n",
	542	__func__);
	543	ret = -ETIMEDOUT;
	544	break;
	545	}
	546
1182fca3	547	spin_lock(&op->lock);
ce6c414e	548	finish_wait(&op->waitq, &wait_entry);
1182fca3 MM	549	spin_unlock(&op->lock);
	550
	551	gossip_debug(GOSSIP_WAIT_DEBUG,
	552	"%s:returning ret(%d)\n",
	553	__func__,
	554	ret);
	555
	556	return ret;
	557	}