path: root/net/rds/send.c

/*
 * Copyright (c) 2006 Oracle.  All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 */
#include <linux/kernel.h>
#include <linux/moduleparam.h>
#include <linux/gfp.h>
#include <net/sock.h>
#include <linux/in.h>
#include <linux/list.h>
#include <linux/ratelimit.h>
#include <linux/export.h>

#include "rds.h"

/* When transmitting messages in rds_send_xmit, we need to emerge from
 * time to time and briefly release the CPU. Otherwise the softlock watchdog
 * will kick our shin.
 * Also, it seems fairer to not let one busy connection stall all the
 * others.
 *
 * send_batch_count is the number of times we'll loop in send_xmit. Setting
 * it to 0 will restore the old behavior (where we looped until we had
 * drained the queue).
 */
static int send_batch_count = 64;
module_param(send_batch_count, int, 0444);
MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");

static void rds_send_remove_from_sock(struct list_head *messages, int status);

/*
 * Reset the send state.  Callers must ensure that this doesn't race with
 * rds_send_xmit().
 */
void rds_send_reset(struct rds_connection *conn)
{
	struct rds_message *rm, *tmp;
	unsigned long flags;

	if (conn->c_xmit_rm) {
		rm = conn->c_xmit_rm;
		conn->c_xmit_rm = NULL;
		/* Tell the user the RDMA op is no longer mapped by the
		 * transport. This isn't entirely true (it's flushed out
		 * independently) but as the connection is down, there's
		 * no ongoing RDMA to/from that memory */
		rds_message_unmapped(rm);
		rds_message_put(rm);
	}

	conn->c_xmit_sg = 0;
	conn->c_xmit_hdr_off = 0;
	conn->c_xmit_data_off = 0;
	conn->c_xmit_atomic_sent = 0;
	conn->c_xmit_rdma_sent = 0;
	conn->c_xmit_data_sent = 0;

	conn->c_map_queued = 0;

	conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
	conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;

	/* Mark messages as retransmissions, and move them to the send q */
	spin_lock_irqsave(&conn->c_lock, flags);
	list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
		set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
		set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
	}
	list_splice_init(&conn->c_retrans, &conn->c_send_queue);
	spin_unlock_irqrestore(&conn->c_lock, flags);
}

static int acquire_in_xmit(struct rds_connection *conn)
{
	return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
}

static void release_in_xmit(struct rds_connection *conn)
{
	clear_bit(RDS_IN_XMIT, &conn->c_flags);
	smp_mb__after_clear_bit();
	/*
	 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
	 * hot path and finding waiters is very rare.  We don't want to walk
	 * the system-wide hashed waitqueue buckets in the fast path only to
	 * almost never find waiters.
	 */
	if (waitqueue_active(&conn->c_waitq))
		wake_up_all(&conn->c_waitq);
}

/*
 * We're making the conscious trade-off here to only send one message
 * down the connection at a time.
 *   Pro:
 *      - tx queueing is a simple fifo list
 *   	- reassembly is optional and easily done by transports per conn
 *      - no per flow rx lookup at all, straight to the socket
 *   	- less per-frag memory and wire overhead
 *   Con:
 *      - queued acks can be delayed behind large messages
 *   Depends:
 *      - small message latency is higher behind queued large messages
 *      - large message latency isn't starved by intervening small sends
 */
int rds_send_xmit(struct rds_connection *conn)
{
	struct rds_message *rm;
	unsigned long flags;
	unsigned int tmp;
	struct scatterlist *sg;
	int ret = 0;
	LIST_HEAD(to_be_dropped);

restart:

	/*
	 * sendmsg calls here after having queued its message on the send
	 * queue.  We only have one task feeding the connection at a time.  If
	 * another thread is already feeding the queue then we back off.  This
	 * avoids blocking the caller and trading per-connection data between
	 * caches per message.
	 */
	if (!acquire_in_xmit(conn)) {
		rds_stats_inc(s_send_lock_contention);
		ret = -ENOMEM;
		goto