net/rds/threads.c

   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/random.h>
  35 #include <linux/export.h>
  36
  37 #include "rds.h"
  38
  39 /*
  40  * All of connection management is simplified by serializing it through
  41  * work queues that execute in a connection managing thread.
  42  *
  43  * TCP wants to send acks through sendpage() in response to data_ready(),
  44  * but it needs a process context to do so.
  45  *
  46  * The receive paths need to allocate but can't drop packets (!) so we have
  47  * a thread around to block allocating if the receive fast path sees an
  48  * allocation failure.
  49  */
  50
  51 /* Grand Unified Theory of connection life cycle:
  52  * At any point in time, the connection can be in one of these states:
  53  * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
  54  *
  55  * The following transitions are possible:
  56  *  ANY           -> ERROR
  57  *  UP            -> DISCONNECTING
  58  *  ERROR         -> DISCONNECTING
  59  *  DISCONNECTING -> DOWN
  60  *  DOWN          -> CONNECTING
  61  *  CONNECTING    -> UP
  62  *
  63  * Transition to state DISCONNECTING/DOWN:
  64  *  -   Inside the shutdown worker; synchronizes with xmit path
  65  *      through RDS_IN_XMIT, and with connection management callbacks
  66  *      via c_cm_lock.
  67  *
  68  *      For receive callbacks, we rely on the underlying transport
  69  *      (TCP, IB/RDMA) to provide the necessary synchronisation.
  70  */
  71 struct workqueue_struct *rds_wq;
  72 EXPORT_SYMBOL_GPL(rds_wq);
  73
  74 void rds_connect_complete(struct rds_connection *conn)
  75 {
  76         if (!rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_UP)) {
  77                 printk(KERN_WARNING "%s: Cannot transition to state UP, "
  78                                 "current state is %d\n",
  79                                 __func__,
  80                                 atomic_read(&conn->c_state));
  81                 atomic_set(&conn->c_state, RDS_CONN_ERROR);
  82                 queue_work(rds_wq, &conn->c_down_w);
  83                 return;
  84         }
  85
  86         rdsdebug("conn %p for %pI4 to %pI4 complete\n",
  87           conn, &conn->c_laddr, &conn->c_faddr);
  88
  89         conn->c_reconnect_jiffies = 0;
  90         set_bit(0, &conn->c_map_queued);
  91         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
  92         queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
  93 }
  94 EXPORT_SYMBOL_GPL(rds_connect_complete);
  95
  96 /*
  97  * This random exponential backoff is relied on to eventually resolve racing
  98  * connects.
  99  *
 100  * If connect attempts race then both parties drop both connections and come
 101  * here to wait for a random amount of time before trying again.  Eventually
 102  * the backoff range will be so much greater than the time it takes to
 103  * establish a connection that one of the pair will establish the connection
 104  * before the other's random delay fires.
 105  *
 106  * Connection attempts that arrive while a connection is already established
 107  * are also considered to be racing connects.  This lets a connection from
 108  * a rebooted machine replace an existing stale connection before the transport
 109  * notices that the connection has failed.
 110  *
 111  * We should *always* start with a random backoff; otherwise a broken connection
 112  * will always take several iterations to be re-established.
 113  */
 114 void rds_queue_reconnect(struct rds_connection *conn)
 115 {
 116         unsigned long rand;
 117
 118         rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
 119           conn, &conn->c_laddr, &conn->c_faddr,
 120           conn->c_reconnect_jiffies);
 121
 122         set_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
 123         if (conn->c_reconnect_jiffies == 0) {
 124                 conn->c_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
 125                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
 126                 return;
 127         }
 128
 129         get_random_bytes(&rand, sizeof(rand));
 130         rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
 131                  rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
 132                  conn, &conn->c_laddr, &conn->c_faddr);
 133         queue_delayed_work(rds_wq, &conn->c_conn_w,
 134                            rand % conn->c_reconnect_jiffies);
 135
 136         conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
 137                                         rds_sysctl_reconnect_max_jiffies);
 138 }
 139
 140 void rds_connect_worker(struct work_struct *work)
 141 {
 142         struct rds_connection *conn = container_of(work, struct rds_connection, c_conn_w.work);
 143         int ret;
 144
 145         clear_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
 146         if (rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 147                 ret = conn->c_trans->conn_connect(conn);
 148                 rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
 149                         conn, &conn->c_laddr, &conn->c_faddr, ret);
 150
 151                 if (ret) {
 152                         if (rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_DOWN))
 153                                 rds_queue_reconnect(conn);
 154                         else
 155                                 rds_conn_error(conn, "RDS: connect failed\n");
 156                 }
 157         }
 158 }
 159
 160 void rds_send_worker(struct work_struct *work)
 161 {
 162         struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
 163         int ret;
 164
 165         if (rds_conn_state(conn) == RDS_CONN_UP) {
 166                 ret = rds_send_xmit(conn);
 167                 rdsdebug("conn %p ret %d\n", conn, ret);
 168                 switch (ret) {
 169                 case -EAGAIN:
 170                         rds_stats_inc(s_send_immediate_retry);
 171                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
 172                         break;
 173                 case -ENOMEM:
 174                         rds_stats_inc(s_send_delayed_retry);
 175                         queue_delayed_work(rds_wq, &conn->c_send_w, 2);
 176                 default:
 177                         break;
 178                 }
 179         }
 180 }
 181
 182 void rds_recv_worker(struct work_struct *work)
 183 {
 184         struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
 185         int ret;
 186
 187         if (rds_conn_state(conn) == RDS_CONN_UP) {
 188                 ret = conn->c_trans->recv(conn);
 189                 rdsdebug("conn %p ret %d\n", conn, ret);
 190                 switch (ret) {
 191                 case -EAGAIN:
 192                         rds_stats_inc(s_recv_immediate_retry);
 193                         queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
 194                         break;
 195                 case -ENOMEM:
 196                         rds_stats_inc(s_recv_delayed_retry);
 197                         queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
 198                 default:
 199                         break;
 200                 }
 201         }
 202 }
 203
 204 void rds_shutdown_worker(struct work_struct *work)
 205 {
 206         struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
 207
 208         rds_conn_shutdown(conn);
 209 }
 210
 211 void rds_threads_exit(void)
 212 {
 213         destroy_workqueue(rds_wq);
 214 }
 215
 216 int rds_threads_init(void)
 217 {
 218         rds_wq = create_singlethread_workqueue("krdsd");
 219         if (!rds_wq)
 220                 return -ENOMEM;
 221
 222         return 0;
 223 }