net/rds/tcp_send.c

0001 /*
0002  * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved.
0003  *
0004  * This software is available to you under a choice of one of two
0005  * licenses.  You may choose to be licensed under the terms of the GNU
0006  * General Public License (GPL) Version 2, available from the file
0007  * COPYING in the main directory of this source tree, or the
0008  * OpenIB.org BSD license below:
0009  *
0010  *     Redistribution and use in source and binary forms, with or
0011  *     without modification, are permitted provided that the following
0012  *     conditions are met:
0013  *
0014  *      - Redistributions of source code must retain the above
0015  *        copyright notice, this list of conditions and the following
0016  *        disclaimer.
0017  *
0018  *      - Redistributions in binary form must reproduce the above
0019  *        copyright notice, this list of conditions and the following
0020  *        disclaimer in the documentation and/or other materials
0021  *        provided with the distribution.
0022  *
0023  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0024  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0025  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0026  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0027  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0028  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0029  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0030  * SOFTWARE.
0031  *
0032  */
0033 #include <linux/kernel.h>
0034 #include <linux/in.h>
0035 #include <net/tcp.h>
0036
0037 #include "rds_single_path.h"
0038 #include "rds.h"
0039 #include "tcp.h"
0040
0041 void rds_tcp_xmit_path_prepare(struct rds_conn_path *cp)
0042 {
0043     struct rds_tcp_connection *tc = cp->cp_transport_data;
0044
0045     tcp_sock_set_cork(tc->t_sock->sk, true);
0046 }
0047
0048 void rds_tcp_xmit_path_complete(struct rds_conn_path *cp)
0049 {
0050     struct rds_tcp_connection *tc = cp->cp_transport_data;
0051
0052     tcp_sock_set_cork(tc->t_sock->sk, false);
0053 }
0054
0055 /* the core send_sem serializes this with other xmit and shutdown */
0056 static int rds_tcp_sendmsg(struct socket *sock, void *data, unsigned int len)
0057 {
0058     struct kvec vec = {
0059         .iov_base = data,
0060         .iov_len = len,
0061     };
0062     struct msghdr msg = {
0063         .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL,
0064     };
0065
0066     return kernel_sendmsg(sock, &msg, &vec, 1, vec.iov_len);
0067 }
0068
0069 /* the core send_sem serializes this with other xmit and shutdown */
0070 int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm,
0071          unsigned int hdr_off, unsigned int sg, unsigned int off)
0072 {
0073     struct rds_conn_path *cp = rm->m_inc.i_conn_path;
0074     struct rds_tcp_connection *tc = cp->cp_transport_data;
0075     int done = 0;
0076     int ret = 0;
0077     int more;
0078
0079     if (hdr_off == 0) {
0080         /*
0081          * m_ack_seq is set to the sequence number of the last byte of
0082          * header and data.  see rds_tcp_is_acked().
0083          */
0084         tc->t_last_sent_nxt = rds_tcp_write_seq(tc);
0085         rm->m_ack_seq = tc->t_last_sent_nxt +
0086                 sizeof(struct rds_header) +
0087                 be32_to_cpu(rm->m_inc.i_hdr.h_len) - 1;
0088         smp_mb__before_atomic();
0089         set_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags);
0090         tc->t_last_expected_una = rm->m_ack_seq + 1;
0091
0092         if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
0093             rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
0094
0095         rdsdebug("rm %p tcp nxt %u ack_seq %llu\n",
0096              rm, rds_tcp_write_seq(tc),
0097              (unsigned long long)rm->m_ack_seq);
0098     }
0099
0100     if (hdr_off < sizeof(struct rds_header)) {
0101         /* see rds_tcp_write_space() */
0102         set_bit(SOCK_NOSPACE, &tc->t_sock->sk->sk_socket->flags);
0103
0104         ret = rds_tcp_sendmsg(tc->t_sock,
0105                       (void *)&rm->m_inc.i_hdr + hdr_off,
0106                       sizeof(rm->m_inc.i_hdr) - hdr_off);
0107         if (ret < 0)
0108             goto out;
0109         done += ret;
0110         if (hdr_off + done != sizeof(struct rds_header))
0111             goto out;
0112     }
0113
0114     more = rm->data.op_nents > 1 ? (MSG_MORE | MSG_SENDPAGE_NOTLAST) : 0;
0115     while (sg < rm->data.op_nents) {
0116         int flags = MSG_DONTWAIT | MSG_NOSIGNAL | more;
0117
0118         ret = tc->t_sock->ops->sendpage(tc->t_sock,
0119                         sg_page(&rm->data.op_sg[sg]),
0120                         rm->data.op_sg[sg].offset + off,
0121                         rm->data.op_sg[sg].length - off,
0122                         flags);
0123         rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->data.op_sg[sg]),
0124              rm->data.op_sg[sg].offset + off, rm->data.op_sg[sg].length - off,
0125              ret);
0126         if (ret <= 0)
0127             break;
0128
0129         off += ret;
0130         done += ret;
0131         if (off == rm->data.op_sg[sg].length) {
0132             off = 0;
0133             sg++;
0134         }
0135         if (sg == rm->data.op_nents - 1)
0136             more = 0;
0137     }
0138
0139 out:
0140     if (ret <= 0) {
0141         /* write_space will hit after EAGAIN, all else fatal */
0142         if (ret == -EAGAIN) {
0143             rds_tcp_stats_inc(s_tcp_sndbuf_full);
0144             ret = 0;
0145         } else {
0146             /* No need to disconnect/reconnect if path_drop
0147              * has already been triggered, because, e.g., of
0148              * an incoming RST.
0149              */
0150             if (rds_conn_path_up(cp)) {
0151                 pr_warn("RDS/tcp: send to %pI6c on cp [%d]"
0152                     "returned %d, "
0153                     "disconnecting and reconnecting\n",
0154                     &conn->c_faddr, cp->cp_index, ret);
0155                 rds_conn_path_drop(cp, false);
0156             }
0157         }
0158     }
0159     if (done == 0)
0160         done = ret;
0161     return done;
0162 }
0163
0164 /*
0165  * rm->m_ack_seq is set to the tcp sequence number that corresponds to the
0166  * last byte of the message, including the header.  This means that the
0167  * entire message has been received if rm->m_ack_seq is "before" the next
0168  * unacked byte of the TCP sequence space.  We have to do very careful
0169  * wrapping 32bit comparisons here.
0170  */
0171 static int rds_tcp_is_acked(struct rds_message *rm, uint64_t ack)
0172 {
0173     if (!test_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags))
0174         return 0;
0175     return (__s32)((u32)rm->m_ack_seq - (u32)ack) < 0;
0176 }
0177
0178 void rds_tcp_write_space(struct sock *sk)
0179 {
0180     void (*write_space)(struct sock *sk);
0181     struct rds_conn_path *cp;
0182     struct rds_tcp_connection *tc;
0183
0184     read_lock_bh(&sk->sk_callback_lock);
0185     cp = sk->sk_user_data;
0186     if (!cp) {
0187         write_space = sk->sk_write_space;
0188         goto out;
0189     }
0190
0191     tc = cp->cp_transport_data;
0192     rdsdebug("write_space for tc %p\n", tc);
0193     write_space = tc->t_orig_write_space;
0194     rds_tcp_stats_inc(s_tcp_write_space_calls);
0195
0196     rdsdebug("tcp una %u\n", rds_tcp_snd_una(tc));
0197     tc->t_last_seen_una = rds_tcp_snd_una(tc);
0198     rds_send_path_drop_acked(cp, rds_tcp_snd_una(tc), rds_tcp_is_acked);
0199
0200     rcu_read_lock();
0201     if ((refcount_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf &&
0202         !rds_destroy_pending(cp->cp_conn))
0203         queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
0204     rcu_read_unlock();
0205
0206 out:
0207     read_unlock_bh(&sk->sk_callback_lock);
0208
0209     /*
0210      * write_space is only called when data leaves tcp's send queue if
0211      * SOCK_NOSPACE is set.  We set SOCK_NOSPACE every time we put
0212      * data in tcp's send queue because we use write_space to parse the
0213      * sequence numbers and notice that rds messages have been fully
0214      * received.
0215      *
0216      * tcp's write_space clears SOCK_NOSPACE if the send queue has more
0217      * than a certain amount of space. So we need to set it again *after*
0218      * we call tcp's write_space or else we might only get called on the
0219      * first of a series of incoming tcp acks.
0220      */
0221     write_space(sk);
0222
0223     if (sk->sk_socket)
0224         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
0225 }