LXR linux/net/rds/tcp

   1/*
   2 * Copyright (c) 2006 Oracle.  All rights reserved.
   3 *
   4 * This software is available to you under a choice of one of two
   5 * licenses.  You may choose to be licensed under the terms of the GNU
   6 * General Public License (GPL) Version 2, available from the file
   7 * COPYING in the main directory of this source tree, or the
   8 * OpenIB.org BSD license below:
   9 *
  10 *     Redistribution and use in source and binary forms, with or
  11 *     without modification, are permitted provided that the following
  12 *     conditions are met:
  13 *
  14 *      - Redistributions of source code must retain the above
  15 *        copyright notice, this list of conditions and the following
  16 *        disclaimer.
  17 *
  18 *      - Redistributions in binary form must reproduce the above
  19 *        copyright notice, this list of conditions and the following
  20 *        disclaimer in the documentation and/or other materials
  21 *        provided with the distribution.
  22 *
  23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30 * SOFTWARE.
  31 *
  32 */
  33#include <linux/kernel.h>
  34#include <linux/in.h>
  35#include <net/tcp.h>
  36
  37#include "rds_single_path.h"
  38#include "rds.h"
  39#include "tcp.h"
  40
  41static void rds_tcp_cork(struct socket *sock, int val)
  42{
  43        mm_segment_t oldfs;
  44
  45        oldfs = get_fs();
  46        set_fs(KERNEL_DS);
  47        sock->ops->setsockopt(sock, SOL_TCP, TCP_CORK, (char __user *)&val,
  48                              sizeof(val));
  49        set_fs(oldfs);
  50}
  51
  52void rds_tcp_xmit_path_prepare(struct rds_conn_path *cp)
  53{
  54        struct rds_tcp_connection *tc = cp->cp_transport_data;
  55
  56        rds_tcp_cork(tc->t_sock, 1);
  57}
  58
  59void rds_tcp_xmit_path_complete(struct rds_conn_path *cp)
  60{
  61        struct rds_tcp_connection *tc = cp->cp_transport_data;
  62
  63        rds_tcp_cork(tc->t_sock, 0);
  64}
  65
  66/* the core send_sem serializes this with other xmit and shutdown */
  67static int rds_tcp_sendmsg(struct socket *sock, void *data, unsigned int len)
  68{
  69        struct kvec vec = {
  70                .iov_base = data,
  71                .iov_len = len,
  72        };
  73        struct msghdr msg = {
  74                .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL,
  75        };
  76
  77        return kernel_sendmsg(sock, &msg, &vec, 1, vec.iov_len);
  78}
  79
  80/* the core send_sem serializes this with other xmit and shutdown */
  81int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
  82                 unsigned int hdr_off, unsigned int sg, unsigned int off)
  83{
  84        struct rds_conn_path *cp = rm->m_inc.i_conn_path;
  85        struct rds_tcp_connection *tc = cp->cp_transport_data;
  86        int done = 0;
  87        int ret = 0;
  88        int more;
  89
  90        if (hdr_off == 0) {
  91                /*
  92                 * m_ack_seq is set to the sequence number of the last byte of
  93                 * header and data.  see rds_tcp_is_acked().
  94                 */
  95                tc->t_last_sent_nxt = rds_tcp_snd_nxt(tc);
  96                rm->m_ack_seq = tc->t_last_sent_nxt +
  97                                sizeof(struct rds_header) +
  98                                be32_to_cpu(rm->m_inc.i_hdr.h_len) - 1;
  99                smp_mb__before_atomic();
 100                set_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags);
 101                tc->t_last_expected_una = rm->m_ack_seq + 1;
 102
 103                rdsdebug("rm %p tcp nxt %u ack_seq %llu\n",
 104                         rm, rds_tcp_snd_nxt(tc),
 105                         (unsigned long long)rm->m_ack_seq);
 106        }
 107
 108        if (hdr_off < sizeof(struct rds_header)) {
 109                /* see rds_tcp_write_space() */
 110                set_bit(SOCK_NOSPACE, &tc->t_sock->sk->sk_socket->flags);
 111
 112                ret = rds_tcp_sendmsg(tc->t_sock,
 113                                      (void *)&rm->m_inc.i_hdr + hdr_off,
 114                                      sizeof(rm->m_inc.i_hdr) - hdr_off);
 115                if (ret < 0)
 116                        goto out;
 117                done += ret;
 118                if (hdr_off + done != sizeof(struct rds_header))
 119                        goto out;
 120        }
 121
 122        more = rm->data.op_nents > 1 ? (MSG_MORE | MSG_SENDPAGE_NOTLAST) : 0;
 123        while (sg < rm->data.op_nents) {
 124                int flags = MSG_DONTWAIT | MSG_NOSIGNAL | more;
 125
 126                ret = tc->t_sock->ops->sendpage(tc->t_sock,
 127                                                sg_page(&rm->data.op_sg[sg]),
 128                                                rm->data.op_sg[sg].offset + off,
 129                                                rm->data.op_sg[sg].length - off,
 130                                                flags);
 131                rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->data.op_sg[sg]),
 132                         rm->data.op_sg[sg].offset + off, rm->data.op_sg[sg].length - off,
 133                         ret);
 134                if (ret <= 0)
 135                        break;
 136
 137                off += ret;
 138                done += ret;
 139                if (off == rm->data.op_sg[sg].length) {
 140                        off = 0;
 141                        sg++;
 142                }
 143                if (sg == rm->data.op_nents - 1)
 144                        more = 0;
 145        }
 146
 147out:
 148        if (ret <= 0) {
 149                /* write_space will hit after EAGAIN, all else fatal */
 150                if (ret == -EAGAIN) {
 151                        rds_tcp_stats_inc(s_tcp_sndbuf_full);
 152                        ret = 0;
 153                } else {
 154                        /* No need to disconnect/reconnect if path_drop
 155                         * has already been triggered, because, e.g., of
 156                         * an incoming RST.
 157                         */
 158                        if (rds_conn_path_up(cp)) {
 159                                pr_warn("RDS/tcp: send to %pI4 on cp [%d]"
 160                                        "returned %d, "
 161                                        "disconnecting and reconnecting\n",
 162                                        &conn->c_faddr, cp->cp_index, ret);
 163                                rds_conn_path_drop(cp);
 164                        }
 165                }
 166        }
 167        if (done == 0)
 168                done = ret;
 169        return done;
 170}
 171
 172/*
 173 * rm->m_ack_seq is set to the tcp sequence number that corresponds to the
 174 * last byte of the message, including the header.  This means that the
 175 * entire message has been received if rm->m_ack_seq is "before" the next
 176 * unacked byte of the TCP sequence space.  We have to do very careful
 177 * wrapping 32bit comparisons here.
 178 */
 179static int rds_tcp_is_acked(struct rds_message *rm, uint64_t ack)
 180{
 181        if (!test_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags))
 182                return 0;
 183        return (__s32)((u32)rm->m_ack_seq - (u32)ack) < 0;
 184}
 185
 186void rds_tcp_write_space(struct sock *sk)
 187{
 188        void (*write_space)(struct sock *sk);
 189        struct rds_conn_path *cp;
 190        struct rds_tcp_connection *tc;
 191
 192        read_lock_bh(&sk->sk_callback_lock);
 193        cp = sk->sk_user_data;
 194        if (!cp) {
 195                write_space = sk->sk_write_space;
 196                goto out;
 197        }
 198
 199        tc = cp->cp_transport_data;
 200        rdsdebug("write_space for tc %p\n", tc);
 201        write_space = tc->t_orig_write_space;
 202        rds_tcp_stats_inc(s_tcp_write_space_calls);
 203
 204        rdsdebug("tcp una %u\n", rds_tcp_snd_una(tc));
 205        tc->t_last_seen_una = rds_tcp_snd_una(tc);
 206        rds_send_path_drop_acked(cp, rds_tcp_snd_una(tc), rds_tcp_is_acked);
 207
 208        if ((atomic_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf)
 209                queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
 210
 211out:
 212        read_unlock_bh(&sk->sk_callback_lock);
 213
 214        /*
 215         * write_space is only called when data leaves tcp's send queue if
 216         * SOCK_NOSPACE is set.  We set SOCK_NOSPACE every time we put
 217         * data in tcp's send queue because we use write_space to parse the
 218         * sequence numbers and notice that rds messages have been fully
 219         * received.
 220         *
 221         * tcp's write_space clears SOCK_NOSPACE if the send queue has more
 222         * than a certain amount of space. So we need to set it again *after*
 223         * we call tcp's write_space or else we might only get called on the
 224         * first of a series of incoming tcp acks.
 225         */
 226        write_space(sk);
 227
 228        if (sk->sk_socket)
 229                set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 230}
 231