drivers/staging/greybus/operation.c - kernel/msm-4.9 - Gitiles

 /*
  * Greybus operations
  *
  * Copyright 2014 Google Inc.
  *
  * Released under the GPLv2 only.
  */

 #include <linux/kernel.h>
 #include <linux/slab.h>
 #include <linux/module.h>
 #include <linux/workqueue.h>

 #include "greybus.h"

 /*
  * The top bit of the type in an operation message header indicates
  * whether the message is a request (bit clear) or response (bit set)
  */
 #define GB_OPERATION_TYPE_RESPONSE	0x80

 #define OPERATION_TIMEOUT_DEFAULT	1000	/* milliseconds */

 /*
  * XXX This needs to be coordinated with host driver parameters
  * XXX May need to reduce to allow for message header within a page
  */
 #define GB_OPERATION_MESSAGE_SIZE_MAX	4096

 static struct kmem_cache *gb_operation_cache;

 /* Workqueue to handle Greybus operation completions. */
 static struct workqueue_struct *gb_operation_recv_workqueue;

 /*
  * All operation messages (both requests and responses) begin with
  * a common header that encodes the size of the data (header
  * included).  This header also contains a unique identifier, which
  * is used to keep track of in-flight operations.  Finally, the
  * header contains a operation type field, whose interpretation is
  * dependent on what type of device lies on the other end of the
  * connection.  Response messages are distinguished from request
  * messages by setting the high bit (0x80) in the operation type
  * value.
  *
  * The wire format for all numeric fields in the header is little
  * endian.  Any operation-specific data begins immediately after the
  * header, and is 64-bit aligned.
  */
 struct gb_operation_msg_hdr {
 	__le16	size;	/* Size in bytes of header + payload */
 	__le16	id;	/* Operation unique id */
 	__u8	type;	/* E.g GB_I2C_TYPE_* or GB_GPIO_TYPE_* */
 	/* 3 bytes pad, must be zero (ignore when read) */
 } __aligned(sizeof(u64));

 /* XXX Could be per-host device, per-module, or even per-connection */
 static DEFINE_SPINLOCK(gb_operations_lock);

 static void gb_pending_operation_insert(struct gb_operation *operation)
 {
 	struct gb_connection *connection = operation->connection;
 	struct gb_operation_msg_hdr *header;

 	/*
 	 * Assign the operation's id and move it into its
 	 * connection's pending list.
 	 */
 	spin_lock_irq(&gb_operations_lock);
 	operation->id = ++connection->op_cycle;
 	list_move_tail(&operation->links, &connection->pending);
 	spin_unlock_irq(&gb_operations_lock);

 	/* Store the operation id in the request header */
 	header = operation->request.gbuf.transfer_buffer;
 	header->id = cpu_to_le16(operation->id);
 }

 static void gb_pending_operation_remove(struct gb_operation *operation)
 {
 	struct gb_connection *connection = operation->connection;

 	/* Take us off of the list of pending operations */
 	spin_lock_irq(&gb_operations_lock);
 	list_move_tail(&operation->links, &connection->operations);
 	spin_unlock_irq(&gb_operations_lock);
 }

 static struct gb_operation *
 gb_pending_operation_find(struct gb_connection *connection, u16 id)
 {
 	struct gb_operation *operation;
 	bool found = false;

 	spin_lock_irq(&gb_operations_lock);
 	list_for_each_entry(operation, &connection->pending, links)
 		if (operation->id == id) {
 			found = true;
 			break;
 		}
 	spin_unlock_irq(&gb_operations_lock);

 	return found ? operation : NULL;
 }

 static int greybus_submit_gbuf(struct gbuf *gbuf, gfp_t gfp_mask)
 {
 	gbuf->status = -EINPROGRESS;

 	return gbuf->hd->driver->submit_gbuf(gbuf, gfp_mask);
 }

 static void greybus_kill_gbuf(struct gbuf *gbuf)
 {
 	if (gbuf->status != -EINPROGRESS)
 		return;

 	gbuf->hd->driver->kill_gbuf(gbuf);
 }
 /*
  * An operations's response message has arrived.  If no callback was
  * supplied it was submitted for asynchronous completion, so we notify
  * any waiters.  Otherwise we assume calling the completion is enough
  * and nobody else will be waiting.
  */
 static void gb_operation_complete(struct gb_operation *operation)
 {
 	if (operation->callback)
 		operation->callback(operation);
 	else
 		complete_all(&operation->completion);
 }

 /* Wait for a submitted operation to complete */
 int gb_operation_wait(struct gb_operation *operation)
 {
 	int ret;

 	ret = wait_for_completion_interruptible(&operation->completion);
 	/* If interrupted, cancel the in-flight buffer */
 	if (ret < 0)
 		greybus_kill_gbuf(&operation->request.gbuf);
 	return ret;

 }

 static void gb_operation_request_handle(struct gb_operation *operation)
 {
 	struct gb_protocol *protocol = operation->connection->protocol;
 	struct gb_operation_msg_hdr *header;

 	header = operation->request.gbuf.transfer_buffer;

 	/*
 	 * If the protocol has no incoming request handler, report
 	 * an error and mark the request bad.
 	 */
 	if (protocol->request_recv) {
 		protocol->request_recv(header->type, operation);
 		goto out;
 	}

 	gb_connection_err(operation->connection,
 		"unexpected incoming request type 0x%02hhx\n", header->type);
 	operation->result = GB_OP_PROTOCOL_BAD;
 out:
 	gb_operation_complete(operation);
 }

 /*
  * Either this operation contains an incoming request, or its
  * response has arrived.  An incoming request will have a null
  * response buffer pointer (it is the responsibility of the request
  * handler to allocate and fill in the response buffer).
  */
 static void gb_operation_recv_work(struct work_struct *recv_work)
 {
 	struct gb_operation *operation;
 	bool incoming_request;

 	operation = container_of(recv_work, struct gb_operation, recv_work);
 	incoming_request = operation->response.gbuf.transfer_buffer == NULL;
 	if (incoming_request)
 		gb_operation_request_handle(operation);
 	gb_operation_complete(operation);
 }

 /*
  * Timeout call for the operation.
  *
  * If this fires, something went wrong, so mark the result as timed out, and
  * run the completion handler, which (hopefully) should clean up the operation
  * properly.
  */
 static void operation_timeout(struct work_struct *work)
 {
 	struct gb_operation *operation;

 	operation = container_of(work, struct gb_operation, timeout_work.work);
 	pr_debug("%s: timeout!\n", __func__);

 	operation->result = GB_OP_TIMEOUT;
 	gb_operation_complete(operation);
 }

 /*
  * Allocate a buffer to be used for an operation request or response
  * message.  For outgoing messages, both types of message contain a
  * common header, which is filled in here.  Incoming requests or
  * responses also contain the same header, but there's no need to
  * initialize it here (it'll be overwritten by the incoming
  * message).
  */
 static int gb_operation_message_init(struct gb_operation *operation,
 					u8 type, size_t size,
 					bool request, bool outbound)
 {
 	struct gb_connection *connection = operation->connection;
 	struct greybus_host_device *hd = connection->hd;
 	struct gb_message *message;
 	struct gb_operation_msg_hdr *header;
 	struct gbuf *gbuf;
 	gfp_t gfp_flags = request && !outbound ? GFP_ATOMIC : GFP_KERNEL;
 	u16 dest_cport_id;
 	int ret;

 	if (size > GB_OPERATION_MESSAGE_SIZE_MAX)
 		return -E2BIG;
 	size += sizeof(*header);

 	if (request) {
 		message = &operation->request;
 	} else {
 		message = &operation->response;
 		type |= GB_OPERATION_TYPE_RESPONSE;
 	}
 	gbuf = &message->gbuf;

 	if (outbound)
 		dest_cport_id = connection->interface_cport_id;
 	else
 		dest_cport_id = CPORT_ID_BAD;

 	ret = hd->driver->alloc_gbuf_data(gbuf, size, gfp_flags);
 	if (ret)
 		return ret;
 	gbuf->hd = hd;
 	gbuf->dest_cport_id = dest_cport_id;
 	gbuf->status = -EBADR;	/* Initial value--means "never set" */

 	/* Fill in the header structure */
 	header = (struct gb_operation_msg_hdr *)gbuf->transfer_buffer;
 	header->size = cpu_to_le16(size);
 	header->id = 0;		/* Filled in when submitted */
 	header->type = type;

 	message->payload = header + 1;
 	message->operation = operation;

 	return 0;
 }

 static void gb_operation_message_exit(struct gb_message *message)
 {
 	message->operation = NULL;
 	message->payload = NULL;
 	message->gbuf.hd->driver->free_gbuf_data(&message->gbuf);
 }

 /*
  * Create a Greybus operation to be sent over the given connection.
  * The request buffer will big enough for a payload of the given
  * size.  Outgoing requests must specify the size of the response
  * buffer size, which must be sufficient to hold all expected
  * response data.
  *
  * Incoming requests will supply a response size of 0, and in that
  * case no response buffer is allocated.  (A response always
  * includes a status byte, so 0 is not a valid size.)  Whatever
  * handles the operation request is responsible for allocating the
  * response buffer.
  *
  * Returns a pointer to the new operation or a null pointer if an
  * error occurs.
  */
 struct gb_operation *gb_operation_create(struct gb_connection *connection,
 					u8 type, size_t request_size,
 					size_t response_size)
 {
 	struct gb_operation *operation;
 	gfp_t gfp_flags = response_size ? GFP_KERNEL : GFP_ATOMIC;
 	bool outgoing = response_size != 0;
 	int ret;

 	operation = kmem_cache_zalloc(gb_operation_cache, gfp_flags);
 	if (!operation)
 		return NULL;
 	operation->connection = connection;

 	ret = gb_operation_message_init(operation, type, request_size,
 						true, outgoing);
 	if (ret)
 		goto err_cache;

 	if (outgoing) {
 		ret = gb_operation_message_init(operation, type, response_size,
 						false, false);
 		if (ret)
 			goto err_request;
 	}

 	INIT_WORK(&operation->recv_work, gb_operation_recv_work);
 	operation->callback = NULL;	/* set at submit time */
 	init_completion(&operation->completion);
 	INIT_DELAYED_WORK(&operation->timeout_work, operation_timeout);
 	kref_init(&operation->kref);

 	spin_lock_irq(&gb_operations_lock);
 	list_add_tail(&operation->links, &connection->operations);
 	spin_unlock_irq(&gb_operations_lock);

 	return operation;

 err_request:
 	gb_operation_message_exit(&operation->request);
 err_cache:
 	kmem_cache_free(gb_operation_cache, operation);

 	return NULL;
 }

 /*
  * Destroy a previously created operation.
  */
 static void _gb_operation_destroy(struct kref *kref)
 {
 	struct gb_operation *operation;

 	operation = container_of(kref, struct gb_operation, kref);

 	/* XXX Make sure it's not in flight */
 	spin_lock_irq(&gb_operations_lock);
 	list_del(&operation->links);
 	spin_unlock_irq(&gb_operations_lock);

 	gb_operation_message_exit(&operation->response);
 	gb_operation_message_exit(&operation->request);

 	kmem_cache_free(gb_operation_cache, operation);
 }

 void gb_operation_put(struct gb_operation *operation)
 {
 	if (!WARN_ON(!operation))
 		kref_put(&operation->kref, _gb_operation_destroy);
 }

 /*
  * Send an operation request message.  The caller has filled in
  * any payload so the request message is ready to go.  If non-null,
  * the callback function supplied will be called when the response
  * message has arrived indicating the operation is complete.  A null
  * callback function is used for a synchronous request; return from
  * this function won't occur until the operation is complete (or an
  * interrupt occurs).
  */
 int gb_operation_request_send(struct gb_operation *operation,
 				gb_operation_callback callback)
 {
 	unsigned long timeout;
 	int ret;

 	if (operation->connection->state != GB_CONNECTION_STATE_ENABLED)
 		return -ENOTCONN;

 	/*
 	 * XXX
 	 * I think the order of operations is going to be
 	 * significant, and if so, we may need a mutex to surround
 	 * setting the operation id and submitting the gbuf.
 	 */
 	operation->callback = callback;
 	gb_pending_operation_insert(operation);
 	ret = greybus_submit_gbuf(&operation->request.gbuf, GFP_KERNEL);
 	if (ret)
 		return ret;

 	/* We impose a time limit for requests to complete.  */
 	timeout = msecs_to_jiffies(OPERATION_TIMEOUT_DEFAULT);
 	schedule_delayed_work(&operation->timeout_work, timeout);
 	if (!callback)
 		ret = gb_operation_wait(operation);

 	return ret;
 }

 /*
  * Send a response for an incoming operation request.
  */
 int gb_operation_response_send(struct gb_operation *operation)
 {
 	gb_operation_destroy(operation);

 	return 0;
 }

 /*
  * Handle data arriving on a connection.  As soon as we return, the
  * incoming data buffer will be reused, so we need to copy the data
  * into one of our own operation message buffers.
  *
  * If the incoming data is an operation response message, look up
  * the operation and copy the incoming data into its response
  * buffer.  Otherwise allocate a new operation and copy the incoming
  * data into its request buffer.
  *
  * This is called in interrupt context, so just copy the incoming
  * data into the buffer and do remaining handling via a work queue.
  *
  */
 void gb_connection_operation_recv(struct gb_connection *connection,
 				void *data, size_t size)
 {
 	struct gb_operation_msg_hdr *header;
 	struct gb_operation *operation;
 	struct gbuf *gbuf;
 	u16 msg_size;

 	if (connection->state != GB_CONNECTION_STATE_ENABLED)
 		return;

 	if (size < sizeof(*header)) {
 		gb_connection_err(connection, "message too small");
 		return;
 	}

 	header = data;
 	msg_size = le16_to_cpu(header->size);
 	if (header->type & GB_OPERATION_TYPE_RESPONSE) {
 		u16 id = le16_to_cpu(header->id);

 		operation = gb_pending_operation_find(connection, id);
 		if (!operation) {
 			gb_connection_err(connection, "operation not found");
 			return;
 		}
 		cancel_delayed_work(&operation->timeout_work);
 		gb_pending_operation_remove(operation);
 		gbuf = &operation->response.gbuf;
 		if (size > gbuf->transfer_buffer_length) {
 			operation->result = GB_OP_OVERFLOW;
 			gb_connection_err(connection, "recv buffer too small");
 			return;
 		}
 		operation->result = GB_OP_SUCCESS;
 	} else {
 		WARN_ON(msg_size != size);
 		operation = gb_operation_create(connection, header->type,
 							msg_size, 0);
 		if (!operation) {
 			gb_connection_err(connection, "can't create operation");
 			return;
 		}
 		gbuf = &operation->request.gbuf;
 	}

 	memcpy(gbuf->transfer_buffer, data, msg_size);

 	/* The rest will be handled in work queue context */
 	queue_work(gb_operation_recv_workqueue, &operation->recv_work);
 }

 /*
  * Cancel an operation.
  */
 void gb_operation_cancel(struct gb_operation *operation)
 {
 	operation->canceled = true;
 	greybus_kill_gbuf(&operation->request.gbuf);
 	if (operation->response.gbuf.transfer_buffer)
 		greybus_kill_gbuf(&operation->response.gbuf);
 }

 int gb_operation_init(void)
 {
 	gb_operation_cache = kmem_cache_create("gb_operation_cache",
 				sizeof(struct gb_operation), 0, 0, NULL);
 	if (!gb_operation_cache)
 		return -ENOMEM;

 	gb_operation_recv_workqueue = alloc_workqueue("greybus_recv", 0, 1);
 	if (!gb_operation_recv_workqueue) {
 		kmem_cache_destroy(gb_operation_cache);
 		gb_operation_cache = NULL;
 		return -ENOMEM;
 	}

 	return 0;
 }

 void gb_operation_exit(void)
 {
 	destroy_workqueue(gb_operation_recv_workqueue);
 	gb_operation_recv_workqueue = NULL;
 	kmem_cache_destroy(gb_operation_cache);
 	gb_operation_cache = NULL;
 }
	/*
	* Greybus operations
	*
	* Copyright 2014 Google Inc.
	*
	* Released under the GPLv2 only.
	*/

	#include <linux/kernel.h>
	#include <linux/slab.h>
	#include <linux/module.h>
	#include <linux/workqueue.h>

	#include "greybus.h"

	/*
	* The top bit of the type in an operation message header indicates
	* whether the message is a request (bit clear) or response (bit set)
	*/
	#define GB_OPERATION_TYPE_RESPONSE 0x80

	#define OPERATION_TIMEOUT_DEFAULT 1000 /* milliseconds */

	/*
	* XXX This needs to be coordinated with host driver parameters
	* XXX May need to reduce to allow for message header within a page
	*/
	#define GB_OPERATION_MESSAGE_SIZE_MAX 4096

	static struct kmem_cache *gb_operation_cache;

	/* Workqueue to handle Greybus operation completions. */
	static struct workqueue_struct *gb_operation_recv_workqueue;

	/*
	* All operation messages (both requests and responses) begin with
	* a common header that encodes the size of the data (header
	* included). This header also contains a unique identifier, which
	* is used to keep track of in-flight operations. Finally, the
	* header contains a operation type field, whose interpretation is
	* dependent on what type of device lies on the other end of the
	* connection. Response messages are distinguished from request
	* messages by setting the high bit (0x80) in the operation type
	* value.
	*
	* The wire format for all numeric fields in the header is little
	* endian. Any operation-specific data begins immediately after the
	* header, and is 64-bit aligned.
	*/
	struct gb_operation_msg_hdr {
	__le16 size; /* Size in bytes of header + payload */
	__le16 id; /* Operation unique id */
	__u8 type; /* E.g GB_I2C_TYPE_* or GB_GPIO_TYPE_* */
	/* 3 bytes pad, must be zero (ignore when read) */
	} __aligned(sizeof(u64));

	/* XXX Could be per-host device, per-module, or even per-connection */
	static DEFINE_SPINLOCK(gb_operations_lock);

	static void gb_pending_operation_insert(struct gb_operation *operation)
	{
	struct gb_connection *connection = operation->connection;
	struct gb_operation_msg_hdr *header;

	/*
	* Assign the operation's id and move it into its
	* connection's pending list.
	*/
	spin_lock_irq(&gb_operations_lock);
	operation->id = ++connection->op_cycle;
	list_move_tail(&operation->links, &connection->pending);
	spin_unlock_irq(&gb_operations_lock);

	/* Store the operation id in the request header */
	header = operation->request.gbuf.transfer_buffer;
	header->id = cpu_to_le16(operation->id);
	}

	static void gb_pending_operation_remove(struct gb_operation *operation)
	{
	struct gb_connection *connection = operation->connection;

	/* Take us off of the list of pending operations */
	spin_lock_irq(&gb_operations_lock);
	list_move_tail(&operation->links, &connection->operations);
	spin_unlock_irq(&gb_operations_lock);
	}

	static struct gb_operation *
	gb_pending_operation_find(struct gb_connection *connection, u16 id)
	{
	struct gb_operation *operation;
	bool found = false;

	spin_lock_irq(&gb_operations_lock);
	list_for_each_entry(operation, &connection->pending, links)
	if (operation->id == id) {
	found = true;
	break;
	}
	spin_unlock_irq(&gb_operations_lock);

	return found ? operation : NULL;
	}

	static int greybus_submit_gbuf(struct gbuf *gbuf, gfp_t gfp_mask)
	{
	gbuf->status = -EINPROGRESS;

	return gbuf->hd->driver->submit_gbuf(gbuf, gfp_mask);
	}

	static void greybus_kill_gbuf(struct gbuf *gbuf)
	{
	if (gbuf->status != -EINPROGRESS)
	return;

	gbuf->hd->driver->kill_gbuf(gbuf);
	}
	/*
	* An operations's response message has arrived. If no callback was
	* supplied it was submitted for asynchronous completion, so we notify
	* any waiters. Otherwise we assume calling the completion is enough
	* and nobody else will be waiting.
	*/
	static void gb_operation_complete(struct gb_operation *operation)
	{
	if (operation->callback)
	operation->callback(operation);
	else
	complete_all(&operation->completion);
	}

	/* Wait for a submitted operation to complete */
	int gb_operation_wait(struct gb_operation *operation)
	{
	int ret;

	ret = wait_for_completion_interruptible(&operation->completion);
	/* If interrupted, cancel the in-flight buffer */
	if (ret < 0)
	greybus_kill_gbuf(&operation->request.gbuf);
	return ret;

	}

	static void gb_operation_request_handle(struct gb_operation *operation)
	{
	struct gb_protocol *protocol = operation->connection->protocol;
	struct gb_operation_msg_hdr *header;

	header = operation->request.gbuf.transfer_buffer;

	/*
	* If the protocol has no incoming request handler, report
	* an error and mark the request bad.
	*/
	if (protocol->request_recv) {
	protocol->request_recv(header->type, operation);
	goto out;
	}

	gb_connection_err(operation->connection,
	"unexpected incoming request type 0x%02hhx\n", header->type);
	operation->result = GB_OP_PROTOCOL_BAD;
	out:
	gb_operation_complete(operation);
	}

	/*
	* Either this operation contains an incoming request, or its
	* response has arrived. An incoming request will have a null
	* response buffer pointer (it is the responsibility of the request
	* handler to allocate and fill in the response buffer).
	*/
	static void gb_operation_recv_work(struct work_struct *recv_work)
	{
	struct gb_operation *operation;
	bool incoming_request;

	operation = container_of(recv_work, struct gb_operation, recv_work);
	incoming_request = operation->response.gbuf.transfer_buffer == NULL;
	if (incoming_request)
	gb_operation_request_handle(operation);
	gb_operation_complete(operation);
	}

	/*
	* Timeout call for the operation.
	*
	* If this fires, something went wrong, so mark the result as timed out, and
	* run the completion handler, which (hopefully) should clean up the operation
	* properly.
	*/
	static void operation_timeout(struct work_struct *work)
	{
	struct gb_operation *operation;

	operation = container_of(work, struct gb_operation, timeout_work.work);
	pr_debug("%s: timeout!\n", __func__);

	operation->result = GB_OP_TIMEOUT;
	gb_operation_complete(operation);
	}

	/*
	* Allocate a buffer to be used for an operation request or response
	* message. For outgoing messages, both types of message contain a
	* common header, which is filled in here. Incoming requests or
	* responses also contain the same header, but there's no need to
	* initialize it here (it'll be overwritten by the incoming
	* message).
	*/
	static int gb_operation_message_init(struct gb_operation *operation,
	u8 type, size_t size,
	bool request, bool outbound)
	{
	struct gb_connection *connection = operation->connection;
	struct greybus_host_device *hd = connection->hd;
	struct gb_message *message;
	struct gb_operation_msg_hdr *header;
	struct gbuf *gbuf;
	gfp_t gfp_flags = request && !outbound ? GFP_ATOMIC : GFP_KERNEL;
	u16 dest_cport_id;
	int ret;

	if (size > GB_OPERATION_MESSAGE_SIZE_MAX)
	return -E2BIG;
	size += sizeof(*header);

	if (request) {
	message = &operation->request;
	} else {
	message = &operation->response;
	type \|= GB_OPERATION_TYPE_RESPONSE;
	}
	gbuf = &message->gbuf;

	if (outbound)
	dest_cport_id = connection->interface_cport_id;
	else
	dest_cport_id = CPORT_ID_BAD;

	ret = hd->driver->alloc_gbuf_data(gbuf, size, gfp_flags);
	if (ret)
	return ret;
	gbuf->hd = hd;
	gbuf->dest_cport_id = dest_cport_id;
	gbuf->status = -EBADR; /* Initial value--means "never set" */

	/* Fill in the header structure */
	header = (struct gb_operation_msg_hdr *)gbuf->transfer_buffer;
	header->size = cpu_to_le16(size);
	header->id = 0; /* Filled in when submitted */
	header->type = type;

	message->payload = header + 1;
	message->operation = operation;

	return 0;
	}

	static void gb_operation_message_exit(struct gb_message *message)
	{
	message->operation = NULL;
	message->payload = NULL;
	message->gbuf.hd->driver->free_gbuf_data(&message->gbuf);
	}

	/*
	* Create a Greybus operation to be sent over the given connection.
	* The request buffer will big enough for a payload of the given
	* size. Outgoing requests must specify the size of the response
	* buffer size, which must be sufficient to hold all expected
	* response data.
	*
	* Incoming requests will supply a response size of 0, and in that
	* case no response buffer is allocated. (A response always
	* includes a status byte, so 0 is not a valid size.) Whatever
	* handles the operation request is responsible for allocating the
	* response buffer.
	*
	* Returns a pointer to the new operation or a null pointer if an
	* error occurs.
	*/
	struct gb_operation gb_operation_create(struct gb_connection connection,
	u8 type, size_t request_size,
	size_t response_size)
	{
	struct gb_operation *operation;
	gfp_t gfp_flags = response_size ? GFP_KERNEL : GFP_ATOMIC;
	bool outgoing = response_size != 0;
	int ret;

	operation = kmem_cache_zalloc(gb_operation_cache, gfp_flags);
	if (!operation)
	return NULL;
	operation->connection = connection;

	ret = gb_operation_message_init(operation, type, request_size,
	true, outgoing);
	if (ret)
	goto err_cache;

	if (outgoing) {
	ret = gb_operation_message_init(operation, type, response_size,
	false, false);
	if (ret)
	goto err_request;
	}

	INIT_WORK(&operation->recv_work, gb_operation_recv_work);
	operation->callback = NULL; /* set at submit time */
	init_completion(&operation->completion);
	INIT_DELAYED_WORK(&operation->timeout_work, operation_timeout);
	kref_init(&operation->kref);

	spin_lock_irq(&gb_operations_lock);
	list_add_tail(&operation->links, &connection->operations);
	spin_unlock_irq(&gb_operations_lock);

	return operation;

	err_request:
	gb_operation_message_exit(&operation->request);
	err_cache:
	kmem_cache_free(gb_operation_cache, operation);

	return NULL;
	}

	/*
	* Destroy a previously created operation.
	*/
	static void _gb_operation_destroy(struct kref *kref)
	{
	struct gb_operation *operation;

	operation = container_of(kref, struct gb_operation, kref);

	/* XXX Make sure it's not in flight */
	spin_lock_irq(&gb_operations_lock);
	list_del(&operation->links);
	spin_unlock_irq(&gb_operations_lock);

	gb_operation_message_exit(&operation->response);
	gb_operation_message_exit(&operation->request);

	kmem_cache_free(gb_operation_cache, operation);
	}

	void gb_operation_put(struct gb_operation *operation)
	{
	if (!WARN_ON(!operation))
	kref_put(&operation->kref, _gb_operation_destroy);
	}

	/*
	* Send an operation request message. The caller has filled in
	* any payload so the request message is ready to go. If non-null,
	* the callback function supplied will be called when the response
	* message has arrived indicating the operation is complete. A null
	* callback function is used for a synchronous request; return from
	* this function won't occur until the operation is complete (or an
	* interrupt occurs).
	*/
	int gb_operation_request_send(struct gb_operation *operation,
	gb_operation_callback callback)
	{
	unsigned long timeout;
	int ret;

	if (operation->connection->state != GB_CONNECTION_STATE_ENABLED)
	return -ENOTCONN;

	/*
	* XXX
	* I think the order of operations is going to be
	* significant, and if so, we may need a mutex to surround
	* setting the operation id and submitting the gbuf.
	*/
	operation->callback = callback;
	gb_pending_operation_insert(operation);
	ret = greybus_submit_gbuf(&operation->request.gbuf, GFP_KERNEL);
	if (ret)
	return ret;

	/* We impose a time limit for requests to complete. */
	timeout = msecs_to_jiffies(OPERATION_TIMEOUT_DEFAULT);
	schedule_delayed_work(&operation->timeout_work, timeout);
	if (!callback)
	ret = gb_operation_wait(operation);

	return ret;
	}

	/*
	* Send a response for an incoming operation request.
	*/
	int gb_operation_response_send(struct gb_operation *operation)
	{
	gb_operation_destroy(operation);

	return 0;
	}

	/*
	* Handle data arriving on a connection. As soon as we return, the
	* incoming data buffer will be reused, so we need to copy the data
	* into one of our own operation message buffers.
	*
	* If the incoming data is an operation response message, look up
	* the operation and copy the incoming data into its response
	* buffer. Otherwise allocate a new operation and copy the incoming
	* data into its request buffer.
	*
	* This is called in interrupt context, so just copy the incoming
	* data into the buffer and do remaining handling via a work queue.
	*
	*/
	void gb_connection_operation_recv(struct gb_connection *connection,
	void *data, size_t size)
	{
	struct gb_operation_msg_hdr *header;
	struct gb_operation *operation;
	struct gbuf *gbuf;
	u16 msg_size;

	if (connection->state != GB_CONNECTION_STATE_ENABLED)
	return;

	if (size < sizeof(*header)) {
	gb_connection_err(connection, "message too small");
	return;
	}

	header = data;
	msg_size = le16_to_cpu(header->size);
	if (header->type & GB_OPERATION_TYPE_RESPONSE) {
	u16 id = le16_to_cpu(header->id);

	operation = gb_pending_operation_find(connection, id);
	if (!operation) {
	gb_connection_err(connection, "operation not found");
	return;
	}
	cancel_delayed_work(&operation->timeout_work);
	gb_pending_operation_remove(operation);
	gbuf = &operation->response.gbuf;
	if (size > gbuf->transfer_buffer_length) {
	operation->result = GB_OP_OVERFLOW;
	gb_connection_err(connection, "recv buffer too small");
	return;
	}
	operation->result = GB_OP_SUCCESS;
	} else {
	WARN_ON(msg_size != size);
	operation = gb_operation_create(connection, header->type,
	msg_size, 0);
	if (!operation) {
	gb_connection_err(connection, "can't create operation");
	return;
	}
	gbuf = &operation->request.gbuf;
	}

	memcpy(gbuf->transfer_buffer, data, msg_size);

	/* The rest will be handled in work queue context */
	queue_work(gb_operation_recv_workqueue, &operation->recv_work);
	}

	/*
	* Cancel an operation.
	*/
	void gb_operation_cancel(struct gb_operation *operation)
	{
	operation->canceled = true;
	greybus_kill_gbuf(&operation->request.gbuf);
	if (operation->response.gbuf.transfer_buffer)
	greybus_kill_gbuf(&operation->response.gbuf);
	}

	int gb_operation_init(void)
	{
	gb_operation_cache = kmem_cache_create("gb_operation_cache",
	sizeof(struct gb_operation), 0, 0, NULL);
	if (!gb_operation_cache)
	return -ENOMEM;

	gb_operation_recv_workqueue = alloc_workqueue("greybus_recv", 0, 1);
	if (!gb_operation_recv_workqueue) {
	kmem_cache_destroy(gb_operation_cache);
	gb_operation_cache = NULL;
	return -ENOMEM;
	}

	return 0;
	}

	void gb_operation_exit(void)
	{
	destroy_workqueue(gb_operation_recv_workqueue);
	gb_operation_recv_workqueue = NULL;
	kmem_cache_destroy(gb_operation_cache);
	gb_operation_cache = NULL;
	}