605fe3f60de38811df9a41719b5ec15d6464917f
[linux-2.6.git] / net / rds / connection.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/list.h>
35 #include <net/inet_hashtables.h>
36
37 #include "rds.h"
38 #include "loop.h"
39 #include "rdma.h"
40
41 #define RDS_CONNECTION_HASH_BITS 12
42 #define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS)
43 #define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1)
44
45 /* converting this to RCU is a chore for another day.. */
46 static DEFINE_SPINLOCK(rds_conn_lock);
47 static unsigned long rds_conn_count;
48 static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES];
49 static struct kmem_cache *rds_conn_slab;
50
51 static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr)
52 {
53         /* Pass NULL, don't need struct net for hash */
54         unsigned long hash = inet_ehashfn(NULL,
55                                           be32_to_cpu(laddr), 0,
56                                           be32_to_cpu(faddr), 0);
57         return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK];
58 }
59
60 #define rds_conn_info_set(var, test, suffix) do {               \
61         if (test)                                               \
62                 var |= RDS_INFO_CONNECTION_FLAG_##suffix;       \
63 } while (0)
64
65 static inline int rds_conn_is_sending(struct rds_connection *conn)
66 {
67         int ret = 0;
68
69         if (!mutex_trylock(&conn->c_send_lock))
70                 ret = 1;
71         else
72                 mutex_unlock(&conn->c_send_lock);
73
74         return ret;
75 }
76
77 static struct rds_connection *rds_conn_lookup(struct hlist_head *head,
78                                               __be32 laddr, __be32 faddr,
79                                               struct rds_transport *trans)
80 {
81         struct rds_connection *conn, *ret = NULL;
82         struct hlist_node *pos;
83
84         hlist_for_each_entry(conn, pos, head, c_hash_node) {
85                 if (conn->c_faddr == faddr && conn->c_laddr == laddr &&
86                                 conn->c_trans == trans) {
87                         ret = conn;
88                         break;
89                 }
90         }
91         rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret,
92                  &laddr, &faddr);
93         return ret;
94 }
95
96 /*
97  * This is called by transports as they're bringing down a connection.
98  * It clears partial message state so that the transport can start sending
99  * and receiving over this connection again in the future.  It is up to
100  * the transport to have serialized this call with its send and recv.
101  */
102 void rds_conn_reset(struct rds_connection *conn)
103 {
104         rdsdebug("connection %pI4 to %pI4 reset\n",
105           &conn->c_laddr, &conn->c_faddr);
106
107         rds_stats_inc(s_conn_reset);
108         rds_send_reset(conn);
109         conn->c_flags = 0;
110
111         /* Do not clear next_rx_seq here, else we cannot distinguish
112          * retransmitted packets from new packets, and will hand all
113          * of them to the application. That is not consistent with the
114          * reliability guarantees of RDS. */
115 }
116
117 /*
118  * There is only every one 'conn' for a given pair of addresses in the
119  * system at a time.  They contain messages to be retransmitted and so
120  * span the lifetime of the actual underlying transport connections.
121  *
122  * For now they are not garbage collected once they're created.  They
123  * are torn down as the module is removed, if ever.
124  */
125 static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr,
126                                        struct rds_transport *trans, gfp_t gfp,
127                                        int is_outgoing)
128 {
129         struct rds_connection *conn, *tmp, *parent = NULL;
130         struct hlist_head *head = rds_conn_bucket(laddr, faddr);
131         unsigned long flags;
132         int ret;
133
134         spin_lock_irqsave(&rds_conn_lock, flags);
135         conn = rds_conn_lookup(head, laddr, faddr, trans);
136         if (conn
137          && conn->c_loopback
138          && conn->c_trans != &rds_loop_transport
139          && !is_outgoing) {
140                 /* This is a looped back IB connection, and we're
141                  * called by the code handling the incoming connect.
142                  * We need a second connection object into which we
143                  * can stick the other QP. */
144                 parent = conn;
145                 conn = parent->c_passive;
146         }
147         spin_unlock_irqrestore(&rds_conn_lock, flags);
148         if (conn)
149                 goto out;
150
151         conn = kmem_cache_zalloc(rds_conn_slab, gfp);
152         if (conn == NULL) {
153                 conn = ERR_PTR(-ENOMEM);
154                 goto out;
155         }
156
157         INIT_HLIST_NODE(&conn->c_hash_node);
158         conn->c_laddr = laddr;
159         conn->c_faddr = faddr;
160         spin_lock_init(&conn->c_lock);
161         conn->c_next_tx_seq = 1;
162
163         mutex_init(&conn->c_send_lock);
164         INIT_LIST_HEAD(&conn->c_send_queue);
165         INIT_LIST_HEAD(&conn->c_retrans);
166
167         ret = rds_cong_get_maps(conn);
168         if (ret) {
169                 kmem_cache_free(rds_conn_slab, conn);
170                 conn = ERR_PTR(ret);
171                 goto out;
172         }
173
174         /*
175          * This is where a connection becomes loopback.  If *any* RDS sockets
176          * can bind to the destination address then we'd rather the messages
177          * flow through loopback rather than either transport.
178          */
179         if (rds_trans_get_preferred(faddr)) {
180                 conn->c_loopback = 1;
181                 if (is_outgoing && trans->t_prefer_loopback) {
182                         /* "outgoing" connection - and the transport
183                          * says it wants the connection handled by the
184                          * loopback transport. This is what TCP does.
185                          */
186                         trans = &rds_loop_transport;
187                 }
188         }
189
190         conn->c_trans = trans;
191
192         ret = trans->conn_alloc(conn, gfp);
193         if (ret) {
194                 kmem_cache_free(rds_conn_slab, conn);
195                 conn = ERR_PTR(ret);
196                 goto out;
197         }
198
199         atomic_set(&conn->c_state, RDS_CONN_DOWN);
200         conn->c_reconnect_jiffies = 0;
201         INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker);
202         INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker);
203         INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker);
204         INIT_WORK(&conn->c_down_w, rds_shutdown_worker);
205         mutex_init(&conn->c_cm_lock);
206         conn->c_flags = 0;
207
208         rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n",
209           conn, &laddr, &faddr,
210           trans->t_name ? trans->t_name : "[unknown]",
211           is_outgoing ? "(outgoing)" : "");
212
213         spin_lock_irqsave(&rds_conn_lock, flags);
214         if (parent == NULL) {
215                 tmp = rds_conn_lookup(head, laddr, faddr, trans);
216                 if (tmp == NULL)
217                         hlist_add_head(&conn->c_hash_node, head);
218         } else {
219                 tmp = parent->c_passive;
220                 if (!tmp)
221                         parent->c_passive = conn;
222         }
223
224         if (tmp) {
225                 trans->conn_free(conn->c_transport_data);
226                 kmem_cache_free(rds_conn_slab, conn);
227                 conn = tmp;
228         } else {
229                 rds_cong_add_conn(conn);
230                 rds_conn_count++;
231         }
232
233         spin_unlock_irqrestore(&rds_conn_lock, flags);
234
235 out:
236         return conn;
237 }
238
239 struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr,
240                                        struct rds_transport *trans, gfp_t gfp)
241 {
242         return __rds_conn_create(laddr, faddr, trans, gfp, 0);
243 }
244
245 struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr,
246                                        struct rds_transport *trans, gfp_t gfp)
247 {
248         return __rds_conn_create(laddr, faddr, trans, gfp, 1);
249 }
250
251 void rds_conn_destroy(struct rds_connection *conn)
252 {
253         struct rds_message *rm, *rtmp;
254
255         rdsdebug("freeing conn %p for %pI4 -> "
256                  "%pI4\n", conn, &conn->c_laddr,
257                  &conn->c_faddr);
258
259         hlist_del_init(&conn->c_hash_node);
260
261         /* wait for the rds thread to shut it down */
262         atomic_set(&conn->c_state, RDS_CONN_ERROR);
263         cancel_delayed_work(&conn->c_conn_w);
264         queue_work(rds_wq, &conn->c_down_w);
265         flush_workqueue(rds_wq);
266
267         /* tear down queued messages */
268         list_for_each_entry_safe(rm, rtmp,
269                                  &conn->c_send_queue,
270                                  m_conn_item) {
271                 list_del_init(&rm->m_conn_item);
272                 BUG_ON(!list_empty(&rm->m_sock_item));
273                 rds_message_put(rm);
274         }
275         if (conn->c_xmit_rm)
276                 rds_message_put(conn->c_xmit_rm);
277
278         conn->c_trans->conn_free(conn->c_transport_data);
279
280         /*
281          * The congestion maps aren't freed up here.  They're
282          * freed by rds_cong_exit() after all the connections
283          * have been freed.
284          */
285         rds_cong_remove_conn(conn);
286
287         BUG_ON(!list_empty(&conn->c_retrans));
288         kmem_cache_free(rds_conn_slab, conn);
289
290         rds_conn_count--;
291 }
292
293 static void rds_conn_message_info(struct socket *sock, unsigned int len,
294                                   struct rds_info_iterator *iter,
295                                   struct rds_info_lengths *lens,
296                                   int want_send)
297 {
298         struct hlist_head *head;
299         struct hlist_node *pos;
300         struct list_head *list;
301         struct rds_connection *conn;
302         struct rds_message *rm;
303         unsigned long flags;
304         unsigned int total = 0;
305         size_t i;
306
307         len /= sizeof(struct rds_info_message);
308
309         spin_lock_irqsave(&rds_conn_lock, flags);
310
311         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
312              i++, head++) {
313                 hlist_for_each_entry(conn, pos, head, c_hash_node) {
314                         if (want_send)
315                                 list = &conn->c_send_queue;
316                         else
317                                 list = &conn->c_retrans;
318
319                         spin_lock(&conn->c_lock);
320
321                         /* XXX too lazy to maintain counts.. */
322                         list_for_each_entry(rm, list, m_conn_item) {
323                                 total++;
324                                 if (total <= len)
325                                         rds_inc_info_copy(&rm->m_inc, iter,
326                                                           conn->c_laddr,
327                                                           conn->c_faddr, 0);
328                         }
329
330                         spin_unlock(&conn->c_lock);
331                 }
332         }
333
334         spin_unlock_irqrestore(&rds_conn_lock, flags);
335
336         lens->nr = total;
337         lens->each = sizeof(struct rds_info_message);
338 }
339
340 static void rds_conn_message_info_send(struct socket *sock, unsigned int len,
341                                        struct rds_info_iterator *iter,
342                                        struct rds_info_lengths *lens)
343 {
344         rds_conn_message_info(sock, len, iter, lens, 1);
345 }
346
347 static void rds_conn_message_info_retrans(struct socket *sock,
348                                           unsigned int len,
349                                           struct rds_info_iterator *iter,
350                                           struct rds_info_lengths *lens)
351 {
352         rds_conn_message_info(sock, len, iter, lens, 0);
353 }
354
355 void rds_for_each_conn_info(struct socket *sock, unsigned int len,
356                           struct rds_info_iterator *iter,
357                           struct rds_info_lengths *lens,
358                           int (*visitor)(struct rds_connection *, void *),
359                           size_t item_len)
360 {
361         uint64_t buffer[(item_len + 7) / 8];
362         struct hlist_head *head;
363         struct hlist_node *pos;
364         struct hlist_node *tmp;
365         struct rds_connection *conn;
366         unsigned long flags;
367         size_t i;
368
369         spin_lock_irqsave(&rds_conn_lock, flags);
370
371         lens->nr = 0;
372         lens->each = item_len;
373
374         for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash);
375              i++, head++) {
376                 hlist_for_each_entry_safe(conn, pos, tmp, head, c_hash_node) {
377
378                         /* XXX no c_lock usage.. */
379                         if (!visitor(conn, buffer))
380                                 continue;
381
382                         /* We copy as much as we can fit in the buffer,
383                          * but we count all items so that the caller
384                          * can resize the buffer. */
385                         if (len >= item_len) {
386                                 rds_info_copy(iter, buffer, item_len);
387                                 len -= item_len;
388                         }
389                         lens->nr++;
390                 }
391         }
392
393         spin_unlock_irqrestore(&rds_conn_lock, flags);
394 }
395
396 static int rds_conn_info_visitor(struct rds_connection *conn,
397                                   void *buffer)
398 {
399         struct rds_info_connection *cinfo = buffer;
400
401         cinfo->next_tx_seq = conn->c_next_tx_seq;
402         cinfo->next_rx_seq = conn->c_next_rx_seq;
403         cinfo->laddr = conn->c_laddr;
404         cinfo->faddr = conn->c_faddr;
405         strncpy(cinfo->transport, conn->c_trans->t_name,
406                 sizeof(cinfo->transport));
407         cinfo->flags = 0;
408
409         rds_conn_info_set(cinfo->flags,
410                           rds_conn_is_sending(conn), SENDING);
411         /* XXX Future: return the state rather than these funky bits */
412         rds_conn_info_set(cinfo->flags,
413                           atomic_read(&conn->c_state) == RDS_CONN_CONNECTING,
414                           CONNECTING);
415         rds_conn_info_set(cinfo->flags,
416                           atomic_read(&conn->c_state) == RDS_CONN_UP,
417                           CONNECTED);
418         return 1;
419 }
420
421 static void rds_conn_info(struct socket *sock, unsigned int len,
422                           struct rds_info_iterator *iter,
423                           struct rds_info_lengths *lens)
424 {
425         rds_for_each_conn_info(sock, len, iter, lens,
426                                 rds_conn_info_visitor,
427                                 sizeof(struct rds_info_connection));
428 }
429
430 int __init rds_conn_init(void)
431 {
432         rds_conn_slab = kmem_cache_create("rds_connection",
433                                           sizeof(struct rds_connection),
434                                           0, 0, NULL);
435         if (rds_conn_slab == NULL)
436                 return -ENOMEM;
437
438         rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info);
439         rds_info_register_func(RDS_INFO_SEND_MESSAGES,
440                                rds_conn_message_info_send);
441         rds_info_register_func(RDS_INFO_RETRANS_MESSAGES,
442                                rds_conn_message_info_retrans);
443
444         return 0;
445 }
446
447 void rds_conn_exit(void)
448 {
449         rds_loop_exit();
450
451         WARN_ON(!hlist_empty(rds_conn_hash));
452
453         kmem_cache_destroy(rds_conn_slab);
454
455         rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info);
456         rds_info_deregister_func(RDS_INFO_SEND_MESSAGES,
457                                  rds_conn_message_info_send);
458         rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES,
459                                  rds_conn_message_info_retrans);
460 }
461
462 /*
463  * Force a disconnect
464  */
465 void rds_conn_drop(struct rds_connection *conn)
466 {
467         atomic_set(&conn->c_state, RDS_CONN_ERROR);
468         queue_work(rds_wq, &conn->c_down_w);
469 }
470
471 /*
472  * An error occurred on the connection
473  */
474 void
475 __rds_conn_error(struct rds_connection *conn, const char *fmt, ...)
476 {
477         va_list ap;
478
479         va_start(ap, fmt);
480         vprintk(fmt, ap);
481         va_end(ap);
482
483         rds_conn_drop(conn);
484 }