xref: /linux/net/rds/loop.c (revision c809195f5523dd4d09403bbb1c9732d548aa0d1e)
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/slab.h>
35 #include <linux/in.h>
36 #include <net/net_namespace.h>
37 #include <net/netns/generic.h>
38 
39 #include "rds_single_path.h"
40 #include "rds.h"
41 #include "loop.h"
42 
43 static DEFINE_SPINLOCK(loop_conns_lock);
44 static LIST_HEAD(loop_conns);
45 static atomic_t rds_loop_unloading = ATOMIC_INIT(0);
46 
47 static void rds_loop_set_unloading(void)
48 {
49 	atomic_set(&rds_loop_unloading, 1);
50 }
51 
52 static bool rds_loop_is_unloading(struct rds_connection *conn)
53 {
54 	return atomic_read(&rds_loop_unloading) != 0;
55 }
56 
57 /*
58  * This 'loopback' transport is a special case for flows that originate
59  * and terminate on the same machine.
60  *
61  * Connection build-up notices if the destination address is thought of
62  * as a local address by a transport.  At that time it decides to use the
63  * loopback transport instead of the bound transport of the sending socket.
64  *
65  * The loopback transport's sending path just hands the sent rds_message
66  * straight to the receiving path via an embedded rds_incoming.
67  */
68 
69 /*
70  * Usually a message transits both the sender and receiver's conns as it
71  * flows to the receiver.  In the loopback case, though, the receive path
72  * is handed the sending conn so the sense of the addresses is reversed.
73  */
74 static int rds_loop_xmit(struct rds_connection *conn, struct rds_message *rm,
75 			 unsigned int hdr_off, unsigned int sg,
76 			 unsigned int off)
77 {
78 	struct scatterlist *sgp = &rm->data.op_sg[sg];
79 	int ret = sizeof(struct rds_header) +
80 			be32_to_cpu(rm->m_inc.i_hdr.h_len);
81 
82 	/* Do not send cong updates to loopback */
83 	if (rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
84 		rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
85 		ret = min_t(int, ret, sgp->length - conn->c_xmit_data_off);
86 		goto out;
87 	}
88 
89 	BUG_ON(hdr_off || sg || off);
90 
91 	rds_inc_init(&rm->m_inc, conn, conn->c_laddr);
92 	/* For the embedded inc. Matching put is in loop_inc_free() */
93 	rds_message_addref(rm);
94 
95 	rds_recv_incoming(conn, conn->c_laddr, conn->c_faddr, &rm->m_inc,
96 			  GFP_KERNEL);
97 
98 	rds_send_drop_acked(conn, be64_to_cpu(rm->m_inc.i_hdr.h_sequence),
99 			    NULL);
100 
101 	rds_inc_put(&rm->m_inc);
102 out:
103 	return ret;
104 }
105 
106 /*
107  * See rds_loop_xmit(). Since our inc is embedded in the rm, we
108  * make sure the rm lives at least until the inc is done.
109  */
110 static void rds_loop_inc_free(struct rds_incoming *inc)
111 {
112 	struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
113 
114 	rds_message_put(rm);
115 }
116 
117 /* we need to at least give the thread something to succeed */
118 static int rds_loop_recv_path(struct rds_conn_path *cp)
119 {
120 	return 0;
121 }
122 
123 struct rds_loop_connection {
124 	struct list_head loop_node;
125 	struct rds_connection *conn;
126 };
127 
128 /*
129  * Even the loopback transport needs to keep track of its connections,
130  * so it can call rds_conn_destroy() on them on exit. N.B. there are
131  * 1+ loopback addresses (127.*.*.*) so it's not a bug to have
132  * multiple loopback conns allocated, although rather useless.
133  */
134 static int rds_loop_conn_alloc(struct rds_connection *conn, gfp_t gfp)
135 {
136 	struct rds_loop_connection *lc;
137 	unsigned long flags;
138 
139 	lc = kzalloc(sizeof(struct rds_loop_connection), gfp);
140 	if (!lc)
141 		return -ENOMEM;
142 
143 	INIT_LIST_HEAD(&lc->loop_node);
144 	lc->conn = conn;
145 	conn->c_transport_data = lc;
146 
147 	spin_lock_irqsave(&loop_conns_lock, flags);
148 	list_add_tail(&lc->loop_node, &loop_conns);
149 	spin_unlock_irqrestore(&loop_conns_lock, flags);
150 
151 	return 0;
152 }
153 
154 static void rds_loop_conn_free(void *arg)
155 {
156 	struct rds_loop_connection *lc = arg;
157 	unsigned long flags;
158 
159 	rdsdebug("lc %p\n", lc);
160 	spin_lock_irqsave(&loop_conns_lock, flags);
161 	list_del(&lc->loop_node);
162 	spin_unlock_irqrestore(&loop_conns_lock, flags);
163 	kfree(lc);
164 }
165 
166 static int rds_loop_conn_path_connect(struct rds_conn_path *cp)
167 {
168 	rds_connect_complete(cp->cp_conn);
169 	return 0;
170 }
171 
172 static void rds_loop_conn_path_shutdown(struct rds_conn_path *cp)
173 {
174 }
175 
176 void rds_loop_exit(void)
177 {
178 	struct rds_loop_connection *lc, *_lc;
179 	LIST_HEAD(tmp_list);
180 
181 	rds_loop_set_unloading();
182 	synchronize_rcu();
183 	/* avoid calling conn_destroy with irqs off */
184 	spin_lock_irq(&loop_conns_lock);
185 	list_splice(&loop_conns, &tmp_list);
186 	INIT_LIST_HEAD(&loop_conns);
187 	spin_unlock_irq(&loop_conns_lock);
188 
189 	list_for_each_entry_safe(lc, _lc, &tmp_list, loop_node) {
190 		WARN_ON(lc->conn->c_passive);
191 		rds_conn_destroy(lc->conn);
192 	}
193 }
194 
195 static void rds_loop_kill_conns(struct net *net)
196 {
197 	struct rds_loop_connection *lc, *_lc;
198 	LIST_HEAD(tmp_list);
199 
200 	spin_lock_irq(&loop_conns_lock);
201 	list_for_each_entry_safe(lc, _lc, &loop_conns, loop_node)  {
202 		struct net *c_net = read_pnet(&lc->conn->c_net);
203 
204 		if (net != c_net)
205 			continue;
206 		list_move_tail(&lc->loop_node, &tmp_list);
207 	}
208 	spin_unlock_irq(&loop_conns_lock);
209 
210 	list_for_each_entry_safe(lc, _lc, &tmp_list, loop_node) {
211 		WARN_ON(lc->conn->c_passive);
212 		rds_conn_destroy(lc->conn);
213 	}
214 }
215 
216 static void __net_exit rds_loop_exit_net(struct net *net)
217 {
218 	rds_loop_kill_conns(net);
219 }
220 
221 static struct pernet_operations rds_loop_net_ops = {
222 	.exit = rds_loop_exit_net,
223 };
224 
225 int rds_loop_net_init(void)
226 {
227 	return register_pernet_device(&rds_loop_net_ops);
228 }
229 
230 void rds_loop_net_exit(void)
231 {
232 	unregister_pernet_device(&rds_loop_net_ops);
233 }
234 
235 /*
236  * This is missing .xmit_* because loop doesn't go through generic
237  * rds_send_xmit() and doesn't call rds_recv_incoming().  .listen_stop and
238  * .laddr_check are missing because transport.c doesn't iterate over
239  * rds_loop_transport.
240  */
241 struct rds_transport rds_loop_transport = {
242 	.xmit			= rds_loop_xmit,
243 	.recv_path		= rds_loop_recv_path,
244 	.conn_alloc		= rds_loop_conn_alloc,
245 	.conn_free		= rds_loop_conn_free,
246 	.conn_path_connect	= rds_loop_conn_path_connect,
247 	.conn_path_shutdown	= rds_loop_conn_path_shutdown,
248 	.inc_copy_to_user	= rds_message_inc_copy_to_user,
249 	.inc_free		= rds_loop_inc_free,
250 	.t_name			= "loopback",
251 	.t_type			= RDS_TRANS_LOOP,
252 	.t_unloading		= rds_loop_is_unloading,
253 };
254